Ground penetrating radar data generation method and device based on road material constraint adversarial neural network, equipment and medium

Abstract

The invention discloses a ground penetrating radar data generation method and device based on a road material constraint adversarial neural network, equipment and a medium, and relates to the technical field of road detection, and the method comprises the steps: obtaining ground penetrating radar B-SCAN original data and road material parameters of the same road detection area, carrying out the standardization of the original data, carrying out the channel splicing, obtaining a multi-modal feature, and carrying out the recognition of the B-SCAN original data and the road material parameters; combining noise and geological constraints to construct joint input features; single-channel data is generated step by step through a three-stage hierarchical generator, a generated sample and a real sample are constructed, and confrontation training is carried out through a two-dimensional discriminator until convergence to obtain a target generator; and the B-SCAN data of the target ground penetrating radar can be output by inputting the data to be detected. According to the method, multi-modal geological constraint and layered generation are realized, data are more real and reasonable, the problems of insufficient samples and generation distortion are effectively solved, richer ground penetrating radar B-SCAN data are provided for a road damage area, and the intelligent detection level of road internal damage can be improved.

Description

Technical Field

[0001] This invention relates to the field of road detection technology, and in particular to a method, apparatus, equipment and medium for generating ground-penetrating radar data based on a road material constraint adversarial neural network. Background Technology

[0002] Currently, the industry primarily uses traditional generative adversarial neural network (GAN) models for generating and enhancing ground-penetrating radar (GPR) B-SCAN data. The main approach is as follows: First, the raw GPR data undergoes basic preprocessing such as simple amplitude normalization. Then, the preprocessed radar data is fed as a single input into a single-level generator. The generator uses random noise and simple feature mapping to generate simulated GPR data. This generated data is then input along with real data into a discriminator for authentication. Through iterative training of the generator and discriminator, the generator gradually approximates the real data distribution, ultimately achieving the generation of GPR data. Some improved solutions simply append a small amount of geological parameters to the data input, but fail to construct targeted constraint mechanisms for road material parameters.

[0003] Traditional generative adversarial neural network (GAN) models, in the generation of ground-penetrating radar (GPR) B-SCAN data, fail to consider the hierarchical characteristics of GPR data. The single-level generator structure used cannot accurately model the different levels of data features, easily leading to problems such as global geological structure distortion, blurred underground target outlines, and missing signal details. Furthermore, traditional methods lack effective constraint mechanisms for road material parameters. They neither standardize the discrete and continuous parameters of road materials differently nor incorporate the constraint effect of road materials on radar signals during the generation process. This results in poor adaptability of the generated data to the actual road geological background, failing to reproduce the differences in radar echo signal responses corresponding to different road materials. In addition, existing discriminators only judge the overall authenticity of the generated data from a single dimension, failing to simultaneously evaluate the geological logic and physical authenticity of the signals. This makes it difficult to provide effective dual supervision of the generator, resulting in low stability of model training and low reliability of generated data.

[0004] Therefore, how to combine road material parameters to construct a targeted constraint mechanism, realize the hierarchical and accurate generation of ground-penetrating radar B-SCAN data, and improve the accuracy and applicability of road detection has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, device, and medium for generating ground-penetrating radar data based on road material constraint adversarial neural networks, aiming to solve the technical problem of how to improve the accuracy of the generated ground-penetrating radar data.

[0006] To achieve the above objectives, this application proposes a ground-penetrating radar data generation method based on a road material-constrained adversarial neural network, comprising: Acquire raw ground-penetrating radar B-SCAN data and road material parameters for the same road detection area; The raw data from the ground-penetrating radar B-SCAN and the road material parameters are respectively subjected to standardization preprocessing to obtain standardized radar data and standardized road material parameters. The standardized radar data and standardized road material parameters are spliced ​​together to obtain multimodal features; Based on the multimodal features, random noise vectors and geological constraint vectors are fused to construct joint input features; The combined input features are input into a three-level hierarchical generator to generate ground-penetrating radar B-SCAN single-channel data step by step. The ground-penetrating radar B-SCAN single-channel data is spliced ​​with the road material parameters of the corresponding batch to form a generated sample, and the ground-penetrating radar B-SCAN data of the real road detection area is spliced ​​with the real road material parameters of the corresponding batch as the real sample. The generated samples and the real samples are input into a two-dimensional discriminator for adversarial discrimination. The three-level hierarchical generator and the two-dimensional discriminator are trained alternately and iteratively until the model meets the preset convergence condition. The current three-level hierarchical generator is then output as the target three-level hierarchical generator. The raw ground-penetrating radar B-SCAN data of the road detection area to be detected and the road material parameters are input into the target three-level layer generator to obtain the target ground-penetrating radar B-SCAN data.

[0007] In one embodiment, the step of performing standardization preprocessing on the raw ground-penetrating radar B-SCAN data and the road material parameters to obtain standardized radar data and standardized road material parameters includes: The discrete parameters in the road material parameters are converted into numerical features using one-hot encoding to obtain the encoded discrete parameters; The continuous parameters in the road material parameters are mapped to a preset numerical range through minimum-maximum normalization to obtain normalized continuous parameters; By fusing the encoded discrete parameters with the normalized continuous parameters, standardized road material parameters are obtained. Noise cancellation is performed on the raw data of the ground-penetrating radar B-SCAN to obtain denoised radar data. The denoised radar data is normalized in amplitude and mapped to a preset signal range, and phase calibration is performed to obtain standardized radar data.

[0008] In one embodiment, the step of performing channel stitching of the standardized radar data and standardized road material parameters to obtain multimodal features includes: The sensitivity coefficients between ground-penetrating radar signals and various road material parameters were calculated using Pearson correlation analysis. Based on the sensitivity coefficient, generate the weight coefficients corresponding to each road material parameter channel; The standardized road material parameters are weighted and enhanced based on the weighting coefficients to obtain the enhanced road material parameters. The enhanced road material parameters and the standardized radar data are aligned in the channel dimension to obtain the aligned features. Perform channel splicing on the aligned features to obtain initial multimodal features; The initial multimodal features are validated and corrected in terms of feature dimensions to obtain the multimodal features.

[0009] In one embodiment, the step of constructing joint input features based on the fusion of the multimodal feature random noise vector and the geological constraint vector includes: Random noise vectors that follow a normal distribution are randomly sampled according to a preset batch size, and geological constraint vectors corresponding to the batch are extracted from the road material parameter set. The random noise vector and the geological constraint vector are weighted and fused to obtain a noise-constraint fused vector; The noise-constraint fusion vector and the multimodal features are fused along their feature dimensions to obtain a fused feature vector; The fused feature vector is normalized to obtain normalized fused features; The normalized fusion features are mapped to the input dimension of the three-level hierarchical generator to obtain joint input features.

[0010] In one embodiment, the step of inputting the joint input features into a three-level hierarchical generator to generate ground-penetrating radar B-SCAN single-channel data step by step includes: The joint input features are input into the geological framework layer of the three-level hierarchical generator to generate a preset low-resolution macroscopic geological framework feature map. The macroscopic geological framework feature map is input into the target contour layer of the three-level layer generator to generate a target contour feature map with a preset medium resolution. The target contour feature map is input into the signal detail layer of the three-level hierarchical generator to generate a preset high-resolution signal detail feature map. The signal detail feature map is subjected to feature smoothing and dimensionality compression to obtain a single-channel feature matrix; The single-channel feature matrix is ​​converted into the ground-penetrating radar B-SCAN data format to obtain ground-penetrating radar B-SCAN single-channel data.

[0011] In one embodiment, the step of inputting the joint input features into the geological framework layer of a three-level hierarchical generator to generate a preset low-resolution macroscopic geological framework feature map includes: The joint input features are input into the geological parameter encoding module of the geological framework layer and converted into a global constraint feature map. The global constraint feature map is convolved using a large kernel convolution with a preset kernel size to obtain a convolutional feature map; The convolutional feature map is upsampled by multi-layer transposed convolution, batch normalization and activation function to obtain the upsampled feature map; The upsampled feature map is mapped to a preset low resolution to obtain an initial geological framework feature map. The road material constraint adaptability of the initial geological framework feature map is verified to obtain a macroscopic geological framework feature map. The step of inputting the macroscopic geological framework feature map into the target contour layer of the three-level layer generator to generate a target contour feature map with a preset medium resolution includes: The macroscopic geological framework feature map is upsampled by transposed convolution to obtain a medium-resolution geological feature map with a preset medium resolution. Based on the standardized road material parameters, a target candidate region extraction operation is performed on the medium-resolution geological feature map, and a target candidate mask is output. The features within the target candidate mask are processed by reinforcement learning through the boundary constraint convolution module of the target contour layer to obtain an enhanced feature map. A geological compatibility verification operation for road material constraints is performed on the enhanced feature map to obtain a verified medium-resolution feature map. The verified medium-resolution feature map is subjected to feature fusion and smoothing to generate a target contour feature map with a preset medium resolution. The step of inputting the target contour feature map into the signal detail layer of a three-level hierarchical generator to generate a preset high-resolution signal detail feature map includes: The target contour feature map is upsampled to a preset high resolution through transposed convolution to obtain a high-resolution contour feature map; The radar echo intensity is simulated based on the finite-difference time-domain model of the high-resolution contour feature map to generate a signal amplitude feature map. An adaptive clutter signal is generated based on the noise statistical distribution of real ground-penetrating radar data to obtain adaptive clutter characteristics. The adaptive clutter features are superimposed on the signal amplitude feature map to obtain a clutter-containing signal feature map. Phase and amplitude corrections are performed on the clutter-containing signal feature map to obtain a preset high-resolution signal detail feature map.

[0012] In one embodiment, the step of inputting the generated samples and the real samples into a two-dimensional discriminator for adversarial discrimination, and iteratively training the three-level hierarchical generator and the two-dimensional discriminator alternately until the model meets a preset convergence condition, and outputting the current three-level hierarchical generator as the target three-level hierarchical generator includes: The generated samples and the real samples are input into a two-dimensional discriminator to extract two-dimensional features of geological structures and radar signals, respectively, to obtain the feature set of the generated samples and the feature set of the real samples. The generated sample feature set and the real sample feature set are compared and evaluated by a two-dimensional discriminator, and the authenticity discrimination score and the real sample discrimination score are output. The adversarial loss function values ​​of the three-level hierarchical generator and the two-dimensional discriminator are calculated respectively. The network parameters of the three-level hierarchical generator and the two-dimensional discriminator are updated according to the adversarial loss function value by backpropagation algorithm to obtain an optimized generator. When the fluctuation range of the adversarial loss function value within a consecutive preset iteration batch is less than a preset threshold, and the authenticity discrimination score of the generated sample exceeds the preset proportion of the authenticity discrimination score of the real sample, the current optimized generator is output as the target three-level hierarchical generator.

[0013] Furthermore, to achieve the above objectives, this application also proposes a ground-penetrating radar data generation device based on a road material-constrained adversarial neural network, wherein the ground-penetrating radar data generation device based on the road material-constrained adversarial neural network includes: The data acquisition module is used to acquire raw ground-penetrating radar B-SCAN data and road material parameters for the same road detection area; The standardization preprocessing module is used to perform standardization preprocessing on the raw data of the ground penetrating radar B-SCAN and the road material parameters respectively to obtain standardized radar data and standardized road material parameters. A multimodal stitching module is used to stitch together the standardized radar data and standardized road material parameters to obtain multimodal features; The joint input construction module is used to construct joint input features based on the fusion of random noise vectors and geological constraint vectors of the multimodal features; The layered generation module is used to input the joint input features into the three-level layered generator to generate ground-penetrating radar B-SCAN single-channel data step by step. The sample construction module is used to stitch together the ground-penetrating radar B-SCAN single-channel data with the road material parameters of the corresponding batch to form a sample, and to stitch together the ground-penetrating radar B-SCAN data of the real road detection area with the real road material parameters of the corresponding batch as the real sample. The adversarial training module is used to input the generated samples and the real samples into the two-dimensional discriminator for adversarial discrimination, and to alternately iterate the training of the three-level hierarchical generator and the two-dimensional discriminator until the model meets the preset convergence condition, and outputs the current three-level hierarchical generator as the target three-level hierarchical generator. The data generation module is used to input the raw ground-penetrating radar B-SCAN data of the road detection area to be detected and the road material parameters into the target three-level layer generator to obtain the target ground-penetrating radar B-SCAN data.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the ground-penetrating radar data generation method based on road material constraint adversarial neural network as described above.

[0015] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the ground-penetrating radar data generation method based on road material constraint adversarial neural network as described above.

[0016] This application acquires raw ground-penetrating radar (GPR) B-SCAN data and road material parameters from the same road detection area, standardizes them separately, and splices channels to obtain multimodal features. Noise and geological constraints are then fused to construct joint input features. A three-level hierarchical generator progressively generates single-channel data, constructing generated samples and real samples. A two-dimensional discriminator is used for adversarial training until convergence to obtain the target generator. Inputting the data to be detected outputs target GPR B-SCAN data. This approach achieves multimodal geological constraints and hierarchical generation, resulting in more realistic and reasonable data. It effectively solves the problems of insufficient samples and generation distortion, providing richer GPR B-SCAN data for road damage areas and contributing to improving the intelligent detection level of internal road damage. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a flowchart illustrating the first embodiment of the ground-penetrating radar data generation method based on road material constraint adversarial neural network of this application; Figure 2 This is a flowchart illustrating the second embodiment of the ground-penetrating radar data generation method based on road material constraint adversarial neural network of this application; Figure 3 This is a schematic diagram of the module structure of the ground-penetrating radar data generation device based on road material constraint adversarial neural network of this application; Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the ground-penetrating radar data generation method based on road material constraint adversarial neural network in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Currently, the industry primarily uses traditional generative adversarial neural network (GAN) models for generating and enhancing ground-penetrating radar (GPR) B-SCAN data. The main approach is as follows: First, the raw GPR data undergoes basic preprocessing such as simple amplitude normalization. Then, the preprocessed radar data is fed as a single input into a single-level generator. The generator uses random noise and simple feature mapping to generate simulated GPR data. This generated data is then input along with real data into a discriminator for authentication. Through iterative training of the generator and discriminator, the generator gradually approximates the real data distribution, ultimately achieving the generation of GPR data. Some improved solutions simply append a small amount of geological parameters to the data input, but fail to construct targeted constraint mechanisms for road material parameters.

[0023] Traditional generative adversarial neural network (GAN) models, in the generation of ground-penetrating radar (GPR) B-SCAN data, fail to consider the hierarchical characteristics of GPR data. The single-level generator structure used cannot accurately model the different levels of data features, easily leading to problems such as global geological structure distortion, blurred underground target outlines, and missing signal details. Furthermore, traditional methods lack effective constraint mechanisms for road material parameters. They neither standardize the discrete and continuous parameters of road materials differently nor incorporate the constraint effect of road materials on radar signals during the generation process. This results in poor adaptability of the generated data to the actual road geological background, failing to reproduce the differences in radar echo signal responses corresponding to different road materials. In addition, existing discriminators only judge the overall authenticity of the generated data from a single dimension, failing to simultaneously evaluate the geological logic and physical authenticity of the signals. This makes it difficult to provide effective dual supervision of the generator, resulting in low stability of model training and low reliability of generated data.

[0024] Therefore, how to combine road material parameters to construct a targeted constraint mechanism, realize the hierarchical and accurate generation of ground-penetrating radar B-SCAN data, and improve the accuracy and applicability of road detection has become an urgent problem to be solved.

[0025] Based on the above, this application also provides a method for generating ground-penetrating radar data based on a road material-constrained adversarial neural network, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the ground-penetrating radar data generation method based on road material constraint adversarial neural network of this application.

[0026] In this embodiment, the ground-penetrating radar data generation method based on road material constraint adversarial neural network includes steps S10~S80: Step S10: Obtain raw ground-penetrating radar B-SCAN data and road material parameters for the same road detection area.

[0027] Specifically, ground-penetrating radar (GPR) equipment is used to scan and collect data along the survey line in the same road detection area to obtain the raw B-SCAN data of the area. At the same time, core samples of each structural layer are obtained through on-site core drilling, or the road surface material is tested and analyzed in the laboratory to obtain the road material parameters of the corresponding area, including discrete parameters (such as material types such as asphalt concrete, cement-stabilized crushed stone, and graded crushed stone) and continuous parameters (such as density, moisture content, dielectric constant, and layer thickness). The B-SCAN data and material parameters are matched one-to-one according to spatial location. The purpose of this is to establish an accurate mapping relationship between radar signals and road physical properties to ensure the consistency between input conditions and target output during subsequent model training.

[0028] Step S20: Perform standardization preprocessing on the raw data of ground penetrating radar B-SCAN and road material parameters respectively to obtain standardized radar data and standardized road material parameters.

[0029] It should be noted that step S20 includes: converting discrete parameters in the road material parameters into numerical features using one-hot encoding to obtain encoded discrete parameters; mapping continuous parameters in the road material parameters to a preset numerical range using minimum-maximum normalization to obtain normalized continuous parameters; fusing the encoded discrete parameters and the normalized continuous parameters to obtain standardized road material parameters; performing noise cancellation on the raw ground-penetrating radar B-SCAN data to obtain denoised radar data; and performing amplitude normalization mapping on the denoised radar data to a preset signal range and completing phase calibration to obtain standardized radar data.

[0030] It's important to understand that one-hot encoding is a common data preprocessing method that converts discrete categorical data into numerical features that can be processed by machine learning models. It transforms discrete categories into standardized numerical vectors, allowing discrete parameters to be concatenated and fused with continuous parameters to form multimodal input features. Minimax normalization is a normalization method that maps continuous numerical data to a fixed interval, eliminating numerical differences and dimensional influences between data. Preset numerical intervals are fixed numerical ranges set for the normalization of continuous road material parameters, i.e., the unified [0,1] interval in ground-penetrating radar data generation, providing a unified numerical representation standard for various continuous parameters. Noise cancellation is a signal denoising process performed on the raw ground-penetrating radar B-SCAN data, removing irregular interference signals from the raw radar data and retaining the core features of effective echo signals reflecting underground geological characteristics. Amplitude normalization is a processing operation that maps the amplitude of ground-penetrating radar signals to a fixed interval, uniformly mapping the amplitude values ​​of radar echoes to a fixed interval and eliminating amplitude differences under different detection conditions. The preset signal range is a fixed numerical range set for the amplitude normalization of ground-penetrating radar B-SCAN data, specifically the unified [-1,1] interval in ground-penetrating radar data generation, providing a unified characterization standard for radar signal amplitude. Phase calibration is a correction process performed on the phase characteristics of ground-penetrating radar B-SCAN data, which can correct the phase shift generated by the radar signal during propagation and acquisition, restoring the true physical characteristics of the radar echo phase. The encoded discrete parameters are standardized numerical features obtained by converting the discrete parameters of road materials through one-hot encoding, retaining the classification characteristics of discrete parameters, and can directly participate in the numerical calculation and feature fusion of the model. The normalized continuous parameters are standardized numerical features obtained by normalizing the continuous parameters of road materials through min-max normalization, eliminating dimensional and numerical differences, and can work in conjunction with the encoded discrete parameters to characterize the geological features of road materials.

[0031] Specifically, firstly, discrete parameters (such as material type) in the road material parameters are one-hot encoded, converting each category into a binary vector of length equal to the number of categories. The vector is set to 1 for the corresponding category position and 0 for the rest, resulting in encoded discrete parameters. This converts classification information that cannot be directly computed into a numerical representation while avoiding the introduction of spurious order relationships. Secondly, continuous parameters (such as density and moisture content) in the road material parameters are normalized using min-max normalization, linearly mapping the values ​​to the [0,1] interval, resulting in normalized continuous parameters. This eliminates dimensional differences between different physical quantities, ensuring that all parameters participate in subsequent calculations on the same numerical scale. Then, the encoded discrete parameters and the normalized continuous parameters are concatenated and fused along the feature dimension to obtain standardized road material parameters. This constructs a unified multi-channel feature vector, facilitating channel cascading with radar data. Next, noise cancellation is performed on the raw ground-penetrating radar (B-SCAN) data. A bandpass filter is used to remove antenna coupling waves and low-frequency interference, and wavelet soft thresholding is applied to suppress random high-frequency noise, resulting in denoised radar data. This is done to improve the signal-to-noise ratio and avoid noise interference with the accuracy of subsequent normalization mapping. Finally, amplitude normalization is performed on the denoised radar data, linearly mapping the signal amplitude to the [-1,1] interval. Time zero-point correction is performed based on the first arrival time of the direct wave or the two-way travel time of a known reflection interface, completing phase calibration to align the in-phase axes and obtain standardized radar data. This is done to unify the signal dynamic range and eliminate time reference differences, ensuring phase consistency between multi-channel data and providing well-aligned input data for subsequent multi-modal fusion.

[0032] Step S30: The standardized radar data and standardized road material parameters are spliced ​​together to obtain multimodal features.

[0033] It should be noted that step S30 includes: calculating the sensitivity coefficient between the ground-penetrating radar signal and each road material parameter through Pearson correlation analysis; generating the weight coefficient corresponding to each road material parameter channel based on the sensitivity coefficient; performing weighted enhancement processing on the standardized road material parameters based on the weight coefficient to obtain the enhanced road material parameters; aligning the enhanced road material parameters and standardized radar data in the channel dimension to obtain the aligned features; performing channel stitching operation on the aligned features to obtain the initial multimodal features; and performing feature dimension verification and correction on the initial multimodal features to obtain the multimodal features.

[0034] It's important to understand that Pearson correlation analysis is a statistical indicator that measures the degree of linear correlation between two continuous variables. Its core purpose is to quantify the strength of linear associations between variables, indicating simultaneous increases or decreases, or one variable increasing while the other decreases. It can accurately determine the response of ground-penetrating radar (GPR) signals to various road material parameters. The sensitivity coefficient, calculated using Pearson correlation analysis, characterizes the sensitivity of GPR signals to a single road material parameter. A larger absolute value indicates a more significant impact of that parameter on the radar signal characteristics. The weighting coefficient, generated based on the sensitivity coefficient, quantifies the importance of each road material parameter channel in feature fusion, transforming the sensitivity coefficient into a weighting basis that can be directly applied to parameter features. Weighted enhancement processing adjusts the features of each channel of standardized road material parameters according to the weighting coefficients, increasing the feature representation intensity of parameter channels with high sensitivity coefficients and retaining basic features for channels with low sensitivity coefficients. Channel splicing merges features from different sources after dimensional alignment along the channel dimension, fusing road material parameter features and radar signal features into a single feature matrix, achieving initial fusion of multi-source features. The initial multimodal features are the fused features obtained after channel stitching, integrating core features of road material parameters and ground-penetrating radar signals. They have not yet undergone dimensionality verification and represent the raw form of the multimodal features. Feature dimensionality verification and correction involves checking and adjusting the dimensions of the initial multimodal features to verify whether the feature dimensions match the input requirements of the subsequent model, and performing adaptation corrections on mismatched dimensions. The final multimodal features are the fused features obtained after dimensionality verification and correction of the initial multimodal features. They integrate multi-source features from road materials and radar signals, and their dimensions are fully adapted to the training requirements of the subsequent model.

[0035] Specifically, firstly, the Pearson correlation coefficient between the ground-penetrating radar signal and various road material parameters is calculated. This involves extracting the amplitude sequence of each sampling point in the standardized radar data and calculating the covariance to standard deviation ratio for each dimension of the continuous parameters (density, moisture content, dielectric constant) and the discrete parameters after one-heat encoding in the standardized road material parameters. This yields a sensitivity coefficient ranging from -1 to 1. The purpose of this is to quantify the influence of different material parameters on radar echo intensity and identify key constraint parameters. Secondly, weighting coefficients are generated based on the absolute values ​​of the sensitivity coefficients. Specifically, the absolute values ​​of the sensitivity coefficients of each parameter are softmax normalized so that all weights are between 0 and 1 and sum to 1. Parameters with larger absolute values ​​of sensitivity coefficients are assigned higher weights. This aims to adaptively highlight the contribution of key geological information and suppress interference from low-correlation parameters. Then, the standardized road material parameters are weighted channel-by-channel based on the weighting coefficients. The value of each parameter channel is multiplied by its corresponding weight to obtain enhanced road material parameters. This aims to strengthen the characterization of material features that significantly affect the radar signal and weaken redundant information. Next, the enhanced road material parameters are aligned with the standardized radar data in terms of feature dimensions. By copying or interpolating, the material parameters are expanded to the same spatial dimension (number of sampling channels × number of time sampling points) as the radar data, ensuring consistency in depth and width between the two types of data. This alignment eliminates dimensional mismatches and ensures the feasibility of channel stitching. Subsequently, the aligned features are stitched along the channel dimension, stacking the radar data channels and the weighted multi-channel material parameters along the depth direction to form initial multimodal features. This constructs a joint representation of "signal-material" coupling, providing constraints for the generator. Finally, the initial multimodal features are dimension-validated to check if the number of channels and spatial dimensions meet the generator's input requirements. If dimensional deviations exist, they are corrected through pruning, padding, or linear transformation to obtain the final multimodal features. This ensures the standardization of the input data and avoids network forward propagation failures due to dimensional errors.

[0036] Step S40: Construct joint input features based on the fusion of multimodal features, random noise vectors, and geological constraint vectors.

[0037] It should be noted that step S40 includes: randomly sampling random noise vectors that follow a normal distribution according to a preset batch size, and extracting the corresponding batch of geological constraint vectors from the road material parameter set; performing weighted fusion processing on the random noise vectors and geological constraint vectors to obtain a noise-constraint fusion vector; fusing the noise-constraint fusion vector and multimodal features by feature dimension to obtain a fusion feature vector; normalizing the fusion feature vector to obtain a normalized fusion feature; and mapping the normalized fusion feature to the input dimension of the three-level hierarchical generator to obtain joint input features.

[0038] It's important to understand that the preset batch size is the number of input data samples set for model training. It's a fixed value adapted to model training efficiency and computational resources, ensuring a consistent sample size for each iteration and remaining within the limits of computing power. The normally distributed random noise vector is a random numerical vector following a Gaussian distribution. Its numerical distribution conforms to natural randomness; introducing this vector improves the diversity of generated data and avoids the problem of monotonous patterns in the generated results. The geological constraint vector is a feature vector extracted from the road material parameter set, used to constrain the direction of radar data generation. It condenses the core geological features of road materials and limits the geological background range of the generated data. Weighted fusion processing is an operation that combines the random noise vector and the geological constraint vector after numerical weighting. This balances the diversity of random noise and the directional nature of geological constraints, giving the fused vector both randomness and geological constraints. The noise-constraint fusion vector is a composite vector obtained by weighted fusion of the random noise vector and the geological constraint vector. It retains the generation diversity brought by random noise while incorporating the geological constraint characteristics of road materials. Feature dimension fusion is an operation that integrates feature vectors of different dimensions in the feature space. It can merge noise-constraint fusion vectors and multimodal features into a complete feature vector, achieving deep fusion of multiple types of features. The fused feature vector is a feature set obtained after dimension fusion of the noise-constraint fusion vector and multimodal features, integrating four core features: random noise, geological constraints, road materials, and radar signals. Normalization is an operation that standardizes the values ​​of the fused feature vector within a certain range. The normalized fused feature is a standardized feature obtained after normalizing the fused feature vector, with uniform numerical distribution and consistent magnitude, meeting the numerical feature requirements of the model input. The input dimension of the three-level hierarchical generator is a pre-defined feature dimension adapted to the three-level hierarchical generator network structure, that is, the standard feature vector dimension that the generator can receive and process. The joint input feature is the final input feature obtained by mapping the normalized fused feature to the generator input dimension. It integrates multi-source features and its dimensions are fully adapted to the three-level hierarchical generator, serving as the core input data of the generator.

[0039] Specifically, firstly, random noise vectors are generated by randomly sampling from the standard normal distribution N(0,1) according to a preset batch size (e.g., 64 or 128). Simultaneously, the same number of samples are randomly drawn from the pre-constructed road material parameter set as geological constraint vectors, ensuring a one-to-one correspondence between the noise vectors and geological constraint vectors in the batch dimension. This aims to introduce both random diversity sources and structured conditional constraints into the generator, guaranteeing the diversity and rationality of the generated data. Secondly, the random noise vectors and geological constraint vectors are weighted and fused. Specifically, the geological constraint vectors are mapped to the same dimension as the noise vectors through a fully connected layer, and then linearly interpolated according to a learnable fusion weight α. α is initially set to 0.5 and dynamically adjusted during training. This aims to balance the diversity of generation with the strength of conditional constraints, avoiding noise overwhelming geological information or excessive constraints restricting the generation space. Then, the noise-constraint fusion vector is fused with the multimodal features in terms of feature dimensions. Specifically, the fusion vector is expanded to the same spatial size as the multimodal features through a fully connected layer, and then concatenated with the multimodal features in the channel dimension, or added element-wise through a broadcast mechanism in the spatial dimension to obtain the fused feature vector. The purpose of this is to inject global conditional information into the local feature map, realizing global control of the generation process by the conditions. Next, the fused feature vector is batch normalized. The mean and variance of each channel in the current batch are calculated and standardized, and then transformed by learnable scaling and translation parameters to obtain normalized fused features. The purpose of this is to stabilize the data distribution, accelerate network convergence, and prevent gradient vanishing. Finally, the normalized fused features are mapped to the input dimension required by the first layer of the three-level hierarchical generator (e.g., 64×64×512) through convolution or fully connected layers to obtain joint input features. The purpose of this is to match the entry specifications of the generator network and ensure that the features can smoothly enter the geological framework layer for subsequent generation.

[0040] In step S50, the joint input features are input into the three-level hierarchical generator to generate ground-penetrating radar B-SCAN single-channel data step by step.

[0041] It should be noted that, firstly, the joint input features are fed into the geological framework layer of the three-level hierarchical generator to extract global geological features and generate low-resolution feature maps, thereby determining the overall distribution of underground structures and ensuring that the generated data conforms to the geological constraints of road materials. Secondly, the features output from the geological framework layer are fed into the target contour layer to further refine the location and boundary information of underground targets, improving the clarity of target areas in the generated data. Then, the features obtained from the target contour layer are fed into the signal detail layer to supplement fine signal features such as radar echo amplitude and phase, restoring the physical characteristics of real ground-penetrating radar data. Finally, the high-resolution features output from the signal detail layer are subjected to dimensionality compression and format conversion to obtain ground-penetrating radar B-SCAN single-channel data, ensuring that the generated results meet the input requirements for subsequent sample construction and adversarial training.

[0042] Step S60: The ground-penetrating radar B-SCAN single-channel data is spliced ​​with the road material parameters of the corresponding batch to form a generated sample, and the ground-penetrating radar B-SCAN data of the real road detection area is spliced ​​with the real road material parameters of the corresponding batch as the real sample.

[0043] It should be noted that, firstly, the ground-penetrating radar B-SCAN single-channel data output by the three-level hierarchical generator is concatenated with the road material parameters of the corresponding index in the current training batch along the channel dimension. Specifically, the single-channel data is used as the first channel, and the standardized road material parameters are used as subsequent channels in sequence to form generated samples. The purpose of this is to provide the discriminator with complete "signal-material" pairing information, enabling the discriminator to evaluate the suitability of the generated signal with the given material conditions. Secondly, real ground-penetrating radar B-SCAN data and their corresponding real road material parameters of the same index are extracted from the training dataset. Real samples are constructed using the same channel concatenation method as the generated samples, ensuring that the real samples and generated samples are strictly consistent in data format and channel order. The purpose of this is to ensure the standardization of the discriminator's input, avoid discrimination bias caused by format differences, and provide a reliable reference benchmark for adversarial training.

[0044] Step S70: Input the generated samples and real samples into the two-dimensional discriminator for adversarial discrimination, and iteratively train the three-level hierarchical generator and the two-dimensional discriminator alternately until the model meets the preset convergence condition, and output the current three-level hierarchical generator as the target three-level hierarchical generator.

[0045] It should be noted that step S70 includes: inputting generated samples and real samples into a two-dimensional discriminator, extracting two-dimensional features of geological structures and radar signals respectively, to obtain a feature set of generated samples and a feature set of real samples; comparing and evaluating the feature set of generated samples and the feature set of real samples through the two-dimensional discriminator, outputting a realism discrimination score and a real sample discrimination score, and calculating the adversarial loss function values ​​of the three-level hierarchical generator and the two-dimensional discriminator respectively; updating the network parameters of the three-level hierarchical generator and the two-dimensional discriminator according to the adversarial loss function values ​​through the backpropagation algorithm to obtain an optimized generator; when the fluctuation range of the adversarial loss function value within a consecutive preset iteration batch is less than a preset threshold, and the realism discrimination score of the generated samples exceeds a preset proportion of the real sample discrimination score, outputting the current optimized generator as the target three-level hierarchical generator.

[0046] It's important to understand that the dual-dimensional discriminator is a model capable of evaluating both geological structure and radar signal characteristics. It can extract features from both dimensions and perform comparative analysis, rather than simply making a single-dimensional judgment on the overall authenticity of the data. Geological structure features are characteristics reflecting the spatial structure of the geological environment, such as road layer distribution, target depth, and media interfaces, and are the core basis for determining the geological rationality of the generated data. Radar signal features are characteristics reflecting the physical characteristics of the ground-penetrating radar echo, such as amplitude, phase, and clutter distribution, and are the core basis for determining the signal authenticity of the generated data. The generated sample feature set is a collection of geological structure and radar signal features extracted by the dual-dimensional discriminator from the generated samples, fully covering the core feature dimensions of the generated samples. The real sample feature set is a collection of geological structure and radar signal features extracted by the dual-dimensional discriminator from the real samples, serving as the benchmark feature set for comparing and evaluating the generated sample feature set. The authenticity discrimination score is the score output by the dual-dimensional discriminator after comparing the generated sample feature set with the real sample feature set, quantitatively representing the feature similarity between the generated and real samples. The real sample discrimination score is the evaluation score of the two-dimensional discriminator on the authenticity of the real sample feature set itself, serving as a benchmark score for determining whether the generated sample meets the standard. The adversarial loss function value is a numerical metric that measures the generation effect of the generator and the discrimination effect of the discriminator. The generator loss value reflects the degree of difference between the generated sample and the real sample, while the discriminator loss value reflects the accuracy of the discrimination result. The backpropagation algorithm is a core algorithm in deep learning used to update network parameters. It propagates the loss function value backward from the output layer to the input layer, adjusting the network parameters through gradient descent to reduce the loss value. Network parameters are the numerical parameters that constitute the three-level hierarchical generator and the two-dimensional discriminator network structure, namely the weights and biases of each convolutional and fully connected layer, directly determining the model's feature extraction and data generation capabilities. The optimized generator is the three-level hierarchical generator whose parameters have been updated using the backpropagation algorithm. Its ability to generate data is optimized compared to before the update, and it more closely resembles the features of real samples. The continuous preset iteration batch is the set number of consecutive iterations for determining model convergence. This means that the model is considered stable only if convergence is achieved across multiple iterations, avoiding the randomness of a single iteration. The fluctuation range of the adversarial loss function value is the range of change in the adversarial loss function value within the continuous preset iteration batch, reflecting the stability of the model training process. The smaller the fluctuation, the closer the model is to convergence. The preset threshold is a critical value set for the fluctuation range of the loss function value; when the fluctuation range is below this value, the model loss value is considered to be stable. The preset ratio is a target ratio set for the generated sample authenticity discrimination score; when the generated sample score reaches this ratio of the real sample score, the authenticity of the generated data is considered to meet the requirements. The target three-level hierarchical generator is the final optimized generator output after the convergence condition is met, possessing the ability to stably generate ground-penetrating radar data that meets geological constraints and signal authenticity requirements.

[0047] Specifically, firstly, the dual-dimensional discriminator constructs two independent feature extraction branches. The geological structure branch extracts the joint distribution features of the material parameter channel and the signal channel, that is, quantifies the spatial logical consistency between the stratigraphic interface strike, target burial depth and material type. Specifically, it is measured by calculating the feature similarity distance between the generated sample and the real sample in the geological structure distribution. The radar signal branch extracts statistical features such as signal amplitude distribution, phase gradient, and spectral characteristics, that is, quantifies the numerical distribution deviation of the echo signal in the time domain and frequency domain. The feature sets of the generated sample and the real sample are extracted respectively. This is done in order to comprehensively evaluate the sample quality from two independent dimensions of geological rationality and signal authenticity, and avoid the evaluation blind spots caused by a single discrimination standard. Secondly, the two feature sets are input into the fully connected discriminant layer, and the sigmoid function outputs a realism score ranging from [0,1], where the realism score is D(fake) and the true sample score is D(real). Based on the binary cross-entropy loss function, the generator loss (focusing on increasing D(fake) to deceive the discriminator) and the discriminator loss (focusing on widening the gap between D(real) and D(fake)) are calculated respectively. The purpose of this is to establish an adversarial game between the generator and the discriminator, driving both to alternately improve their capabilities. Then, the gradient of the loss function with respect to the parameters of each network layer is calculated using the backpropagation algorithm. The Adam optimizer is used to update the weight parameters of the three-level hierarchical generator and the two-dimensional discriminator along the gradient inverse direction, resulting in an optimized generator. The purpose of this is to minimize the loss function through gradient descent, making the generated data distribution gradually approximate the real data distribution. Finally, the fluctuation range of the adversarial loss function value within 100 consecutive iterations is monitored, and the standard deviation is calculated to determine whether it is less than 5%. At the same time, the authenticity discrimination score of the generated sample is compared to whether it reaches more than 90% of the authenticity discrimination score of the real sample. When both conditions are met, the model is judged to have converged, and the current optimized generator is output as the target three-level hierarchical generator. The purpose of this is to establish a quantitative convergence judgment criterion, prevent undertraining or overfitting, and ensure that the output generator has a stable and reliable data generation capability.

[0048] Step S80: Input the raw ground-penetrating radar B-SCAN data of the road detection area to be detected and the road material parameters into the target three-level layer generator to obtain the target ground-penetrating radar B-SCAN data.

[0049] It should be noted that the constraints based on road materials are implemented through a three-layer mechanism: feature embedding, generation verification, and discriminative feedback. Constraint factors include road material type, physical parameters, interlayer distribution, target depth, and media interface. Road material parameters are encoded into geological constraint vectors, which are integrated with multimodal features and noise-constraint fusion vectors to limit the geological background range of the generated data from the input level. In each layer of the three-level hierarchical generator, the geological framework adaptability is verified based on material parameters, target candidate regions are extracted, and radar echo intensity is simulated to constrain geological rationality and signal physical characteristics during the generation process. Finally, a two-dimensional discriminator evaluates the data using quantitative indicators such as geological structure similarity distance, logical consistency score, and signal numerical distribution deviation, combined with preset thresholds and proportions. The reverse constraint generator is then optimized to ensure that the generated data conforms to the characteristics of the road materials.

[0050] Specifically, firstly, raw ground-penetrating radar (B-SCAN) data of the target road detection area is collected. Road material parameters for the area, including pavement structure layer type, thickness, density, moisture content, and dielectric constant, are obtained through field testing or historical databases. This ensures the input data conforms to the format specifications used in the training phase. This provides the target generator with constraints within its cognitive range, guaranteeing the reliability of the generated results. Secondly, the raw B-SCAN data and road material parameters of the target area undergo the same standardized preprocessing procedure as in the training phase. This includes one-hot encoding and normalization of material parameters, noise reduction of radar data, amplitude normalization, and phase calibration to obtain standardized input data. This eliminates distribution differences between the target data and the training data, enabling the generator to correctly process new inputs. Then, the standardized road material parameters are used as a geological constraint vector and weighted with randomly sampled noise vectors according to the fusion weights determined in the training phase. This is then combined with the standardized radar data for multimodal feature fusion and dimensional mapping to construct joint input features. This reproduces the conditional input patterns from the training phase and activates the generator's layered generation capability. Finally, the joint input features are input into the target three-level layered generator, which sequentially generates the macroscopic structure through the geological framework layer, depicts the morphology of the disease through the target contour layer, and supplements the physical echo features through the signal detail layer. After upsampling and refining at each level, the preset high-resolution single-channel data is output. After format conversion, the target ground-penetrating radar B-SCAN data with the same specifications as the real detection data is obtained. The purpose of this is to use the trained converged generator to expand high-fidelity samples for data-scarce scenarios, and support the subsequent intelligent identification and diagnostic analysis of road underground diseases.

[0051] This embodiment acquires raw ground-penetrating radar (GPR) B-SCAN data and road material parameters from the same road detection area, standardizes them separately, and splices channels to obtain multimodal features. Noise and geological constraints are then fused to construct joint input features. A three-level hierarchical generator progressively generates single-channel data, constructing generated samples and real samples. A two-dimensional discriminator is used for adversarial training until convergence to obtain the target generator. Inputting the data to be detected outputs the target GPR B-SCAN data. This achieves multimodal geological constraints and hierarchical generation, resulting in more realistic and reasonable data. It effectively solves the problems of insufficient samples and generation distortion, providing richer GPR B-SCAN data for road damage areas and contributing to improving the intelligent detection level of internal road damage.

[0052] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The ground-penetrating radar data generation method based on road material constraint adversarial neural network, step S50, further includes steps S201 to S205: Step S201: Input the joint input features into the geological framework layer of the three-level layered generator to generate a preset low-resolution macroscopic geological framework feature map.

[0053] It should be noted that step S201 includes: inputting the joint input features into the geological parameter encoding module of the geological framework layer and converting it into a global constraint feature map; performing convolution processing on the global constraint feature map using a large kernel convolution with a preset kernel size to obtain a convolutional feature map; upsampling the convolutional feature map through multi-layer transposed convolution, batch normalization, and activation functions to obtain an upsampled feature map; mapping the upsampled feature map to a preset low resolution to obtain an initial geological framework feature map; and performing road material constraint adaptation verification on the initial geological framework feature map to obtain a macroscopic geological framework feature map.

[0054] Understandably, the geological framework layer is the first processing layer in the three-level hierarchical generator, primarily used to extract and construct the overall underground structural distribution of the road. This layer is responsible for forming the macroscopic geological background, providing basic constraints for subsequent contour and detail generation. The geological parameter encoding module is a feature transformation unit within the geological framework layer, used to encode and map road material-related features. This module transforms discretized parameter information into spatial feature forms that can be used for convolution operations. The large-kernel convolution with a preset kernel size is a convolution operation using a larger kernel size, used to extract large-scale spatial features. This convolution can capture the overall geological structure, avoiding local details from interfering with global morphological generation. Multi-layer transposed convolution is a convolution operation used to upsample the feature map, gradually increasing the size of the feature map. The multi-layer structure makes the upsampling process smoother, avoiding abrupt changes and distortions in features. Batch normalization is an operation to normalize the feature map data, keeping the data distribution stable. This operation can improve the stability of the training process, accelerate model convergence, and reduce gradient fluctuations. The activation function is a function that introduces non-linear expressive power, enabling the model to learn complex geological structural relationships. This function enhances the network's ability to fit real geological morphology. The preset low resolution is a fixed size for the output feature map of the geological framework layer, used to characterize macroscopic geological structures. The road material constraint fit verification is an operation that compares the feature map with road material parameters for consistency. This verification is used to determine whether the generated geological framework conforms to the actual characteristics of the road medium.

[0055] Specifically, firstly, the joint input features are input into the geological parameter encoding module of the geological framework layer. A fully connected layer maps the geological constraint vectors into low-dimensional dense features, which are then reconstructed into a two-dimensional feature map as the global constraint feature map. This aims to convert road material parameters into a spatial feature representation that the generator can process, establishing an initial association between geological conditions and image generation. Secondly, a 7×7 large-kernel convolution is used to convolve the global constraint feature map. This large receptive field captures the global spatial relationship between the distribution of geological bodies and the orientation of stratigraphic interfaces, resulting in a convolutional feature map. This enhances the network's ability to perceive macroscopic geological structures and avoids the local fragmentation problem caused by small convolutional kernels. Then, three layers of transposed convolution (4×4 kernel size, stride 2, padding 1) are alternately combined with batch normalization and the LeakyReLU activation function to progressively upsample the convolutional feature map. The output feature map size is doubled at each layer, resulting in an upsampled feature map. This aims to stabilize the data distribution and introduce non-linear expressive capabilities while expanding the spatial size of the feature map. Next, the upsampled feature map is processed by 1×1 convolution to adjust the number of channels and interpolated to a preset low resolution of 64×64 pixels, resulting in an initial geological framework feature map. This is done to match the target output size and compress redundant channels, forming a preliminary representation of the macroscopic geological framework. Finally, the initial geological framework feature map is validated for road material constraints. Specifically, the deviation between the stratigraphic interface position in the feature map and the layer thickness information in the input material parameters is calculated. If the deviation exceeds a preset threshold, a correction term is introduced through residual connections to adjust the intensity distribution of the feature map to conform to the road structure's stratigraphic logic, resulting in a macroscopic geological framework feature map. This ensures that the generated global geological structure is physically and logically consistent with the input road material parameters, avoiding stratigraphic misalignment or thickness inconsistencies.

[0056] Step S202: Input the macroscopic geological framework feature map into the target contour layer of the three-level layer generator to generate a target contour feature map with a preset medium resolution.

[0057] It should be noted that step S202 includes: upsampling the macroscopic geological framework feature map through transposed convolution to obtain a preset medium-resolution geological feature map; performing target candidate region extraction on the medium-resolution geological feature map based on standardized road material parameters to output a target candidate mask; performing reinforcement learning processing on the features within the target candidate mask through the boundary constraint convolution module of the target contour layer to obtain an enhanced feature map; performing a geological adaptability verification operation of road material constraints on the enhanced feature map to obtain a verified medium-resolution feature map; and performing feature fusion and smoothing processing on the verified medium-resolution feature map to generate a preset medium-resolution target contour feature map.

[0058] Understandably, the target candidate region extraction operation, based on standardized road material parameters, filters out areas from the feature map where underground targets may exist. This operation, combined with road material characteristics, locks down the range of areas that meet the conditions for target existence, reducing interference from invalid features. The target candidate mask is a binary mask image output by the target candidate region extraction operation, used to mark the target candidate region and background region in the feature map. This mask can accurately define the range where the target may exist, providing clear regional guidance for subsequent boundary strengthening. The target contour layer is an intermediate processing layer of the three-level layered generator, inheriting the macroscopic features of the geological framework layer and focusing on depicting the contour of underground targets. This layer is responsible for refining the target boundary, improving the recognizability of the target region, and connecting the macroscopic geological background with microscopic signal details. The boundary constraint convolution module is a feature strengthening unit inside the target contour layer, specifically used to optimize the boundary features of the target candidate region. Reinforcement learning processing is an operation that performs targeted optimization on the features within the target candidate mask, enhancing the correlation between the target boundary and internal features through the boundary constraint convolution module. The geological adaptability verification operation is an operation that verifies the rationality of the strengthened feature map by combining standardized road material parameters. This verification process determines whether the shape and location of the target outline conform to the geological characteristics of the road material, ensuring that the generated target does not deviate from the actual geological background. Feature fusion and smoothing are operations that integrate and optimize the verified feature map, fusing the feature correlations between the target area and the background area, while eliminating noise and abrupt details in the feature map. This processing makes the overall transition of the feature map more natural and the target outline more regular.

[0059] Specifically, the macroscopic geological framework feature map is upsampled using transposed convolution (kernel size 4×4, stride 2), expanding the feature map size from 64×64 to 128×128. Batch normalization and the LeakyReLU activation function are then used to stabilize the distribution and introduce nonlinearity, resulting in a pre-defined medium-resolution geological feature map. This aims to improve spatial resolution while preserving global geological structure information, providing a more refined feature carrier for subsequent target detail characterization. Secondly, based on the density anomaly range and dielectric constant mutation threshold in standardized road material parameters, target candidate region extraction is performed on the medium-resolution geological feature map. Specifically, a sliding window is used to calculate the similarity between local features and typical disease signal templates. Regions with similarity exceeding a pre-defined threshold are marked as 1, and the rest are marked as 0, outputting a binary target candidate mask. This aims to accurately pinpoint the possible distribution range of underground targets (such as cavities and fissures), providing a focused area for subsequent boundary learning and avoiding wasting computational resources in background areas. Then, the features within the target candidate mask are processed using reinforcement learning through a boundary-constrained convolution module. Specifically, 3×3 convolution is used to extract edge gradient features within the mask region, and deformable convolution is used at the mask boundary to adaptively adjust the sampling position to fit the irregular target contour. The internal features and boundary features are then added element-wise to obtain an enhanced feature map. This is done to enhance the model's perception of the target shape and boundary, making the generated target contour clearer and more accurate. Next, a geological adaptability verification operation for road material constraints is performed on the enhanced feature map. Specifically, based on the target type (void / loose / crack) in the input material parameters, a pre-set morphological rule library is called to check whether the burial depth range, aspect ratio, and relative position to the upper and lower layer interfaces of the generated target conform to the development law of this type of disease in the corresponding road structure. If there is a deviation, the feature response intensity is adjusted through a spatial attention mechanism to obtain a verified medium-resolution feature map. This is done to ensure that the geometric shape and spatial position of the target are consistent with the road material conditions and geomechanical laws, avoiding the generation of false targets that do not conform to engineering reality. Finally, the verified medium-resolution feature map is subjected to feature fusion and smoothing. Specifically, the verified features are combined with the same-sized skip connection features from the geological framework layer through channel splicing, then 1×1 convolution is used for dimensionality reduction and fusion, and finally Gaussian filtering is used to smooth noise abrupt changes to generate a 128×128 pixel target contour feature map. The purpose of this is to integrate multi-scale information and eliminate high-frequency artifacts to form a target representation with clear boundaries and accurate location.

[0060] Step S203: Input the target contour feature map into the signal detail layer of the three-level layer generator to generate a preset high-resolution signal detail feature map.

[0061] It should be noted that step S203 includes: upsampling the target contour feature map to a preset high resolution through transposed convolution to obtain a high-resolution contour feature map; simulating the radar echo intensity of the high-resolution contour feature map based on the finite-difference time-domain model to generate a signal amplitude feature map; generating an adaptive clutter signal based on the noise statistical distribution of real ground-penetrating radar data to obtain adaptive clutter features; superimposing the adaptive clutter features onto the signal amplitude feature map to obtain a clutter-containing signal feature map; and performing phase and amplitude correction on the clutter-containing signal feature map to obtain a preset high-resolution signal detail feature map.

[0062] Understandably, the finite-difference time-domain (FDTD) model is a numerical model used to simulate electromagnetic wave propagation and reflection. This model calculates radar echo intensity based on geological and media parameters, ensuring the generated signal conforms to the laws of physical propagation. Radar echo intensity simulation is the process of calculating the magnitude of the radar reflected signal based on the characteristics of the underground medium. This simulation can reproduce the signal response corresponding to different road materials and target structures, ensuring the authenticity of the generated data. The noise statistical distribution of real ground-penetrating radar data is the noise distribution pattern extracted from actual acquired data. This distribution is used to describe the characteristics of interference signals in the real environment, providing a basis for clutter generation. Adaptive clutter features are the characteristic representation of adaptive clutter signals. This feature matches the size of the high-resolution feature map and can be directly superimposed on the main signal. Phase and amplitude correction are numerical adjustments to the signal feature map to correct signal deviations. This correction ensures that the phase relationship and amplitude range of the signal conform to the characteristics of real radar data.

[0063] Specifically, firstly, the 128×128 pixel target contour feature map is upsampled to a preset high resolution of 512×512 pixels through transposed convolution (kernel size 4×4, stride 2), and then processed by batch normalization and LeakyReLU activation function to obtain a high-resolution contour feature map. The purpose of this is to expand the mesoscopic target information to a fine spatial scale that matches the real radar data, providing a high-dimensional feature carrier for the generation of microscopic signal details. Secondly, radar echo intensity is simulated on the high-resolution contour feature map based on a simplified finite-difference time-domain model. Specifically, the reflection coefficient of electromagnetic waves at each interface is calculated based on the input road material parameters (dielectric constant, conductivity), the time position of the reflected signal is determined according to the two-way travel time formula, and then the interface reflection response is superimposed on the feature map through convolution operation to generate a signal amplitude feature map that conforms to the physical propagation law. The purpose of this is to introduce physical constraints on electromagnetic wave propagation, so that the generated signal amplitude distribution has the real medium response characteristics and avoids physical distortion caused by pure data-driven methods. Then, an adaptive clutter signal is generated based on the noise statistical distribution of real ground-penetrating radar data. Specifically, the mean and standard deviation of the background region of the real data in the training set are calculated to establish a Gaussian distribution model. A random clutter matrix is ​​generated by sampling according to the noise intensity level corresponding to the material parameters of the current batch. Then, wavelet transform is used to adjust the spectral characteristics of the clutter to match the frequency band distribution of the real data, resulting in adaptive clutter features. The purpose of this is to simulate random interference in the real detection environment and enhance the realism of the generated data. Next, the adaptive clutter features are superimposed element-wise with the signal amplitude feature map, and a material-related attenuation coefficient is introduced to control the clutter intensity, resulting in a clutter-containing signal feature map. The purpose of this is to restore the background noise level in the real detection while maintaining the main structure of the signal, making the generated data closer to the signal-to-noise ratio characteristics of the measured data. Finally, phase and amplitude corrections are performed on the clutter signal feature map. Specifically, the global phase zero point is adjusted according to the direct wave calibration results, the amplitude distribution is aligned to the statistical range of the real data through histogram matching, and then the block effect is suppressed by edge-preserving filtering to obtain a 512×512 pixel signal detail feature map. The purpose of this is to eliminate phase drift and amplitude offset in the generation process and ensure the consistency of the output data with the real radar data in terms of visual features and numerical distribution.

[0064] Step S204: Perform feature smoothing and dimensionality compression on the signal detail feature map to obtain a single-channel feature matrix.

[0065] Specifically, firstly, the 512×512 pixel signal detail feature map is smoothed. A bilateral filter is used to suppress high-frequency noise and artifacts while maintaining edge sharpness. Nonlinear smoothing is applied to the feature map through joint constraints of spatial domain Gaussian weights and gray-level similarity weights, resulting in a smoothed high-resolution feature map. This aims to eliminate high-frequency artifacts generated during the generation process while protecting the sharpness of target boundaries and geological interfaces, thus improving the visual quality and physical reliability of the data. Secondly, the smoothed high-resolution feature map undergoes dimensionality compression. A 1×1 convolution maps the multi-channel features to a single-channel output, and a Sigmoid activation function constrains the numerical range to the [0,1] interval. Finally, a linear transformation maps the output to the standard amplitude range of the ground-penetrating radar signal, resulting in a single-channel feature matrix. This converts the generator's multi-channel intermediate representation into a two-dimensional matrix consistent with the actual ground-penetrating radar B-SCAN data format, ensuring that the number of channels and numerical range of the output data meet the requirements of subsequent applications.

[0066] Step S205: Convert the single-channel feature matrix into ground-penetrating radar B-SCAN data format to obtain ground-penetrating radar B-SCAN single-channel data.

[0067] Specifically, firstly, the single-channel feature matrix is ​​encapsulated according to the standard format of ground-penetrating radar B-SCAN data. The row dimension of the matrix is ​​mapped to the time sampling axis (depth direction), and the column dimension is mapped to the spatial sampling axis (survey line direction). Corresponding time and spatial coordinate axes are generated based on the original acquisition parameters (sampling window, sampling frequency, channel spacing), resulting in a two-dimensional data array with physical coordinate information. This is done to give the generated data an interpretable physical dimension, enabling it to have a spatiotemporal correspondence consistent with real data. Secondly, the two-dimensional data array is converted to a common ground-penetrating radar data format (such as SEGY, DT1, or a custom binary format), and metadata such as acquisition date, equipment model, antenna frequency, and survey line location are written into the file header. The amplitude data is then encoded and stored according to the format specification, resulting in ground-penetrating radar B-SCAN single-channel data. This is done to enable the generated data to be directly imported into commercial radar processing software or deep learning training frameworks, achieving seamless compatibility and interchangeability with real detection data.

[0068] This embodiment feeds the joint input features into a three-level hierarchical generator, which sequentially generates a low-resolution macroscopic geological framework feature map through the geological framework layer, a medium-resolution target contour feature map through the target contour layer, and a high-resolution signal detail feature map through the signal detail layer. The signal detail feature map is smoothed and its dimensions are compressed to obtain a single-channel feature matrix, which is then converted into the ground-penetrating radar B-SCAN data format to obtain single-channel data. This achieves hierarchical and accurate generation of radar data, taking into account both geological rationality and signal authenticity, improving the clarity and reliability of the generated data, and adapting to the actual needs of road detection.

[0069] Based on the first embodiment of this application, this application also provides a ground-penetrating radar data generation device based on a road material constraint adversarial neural network. Please refer to... Figure 3 The device includes: The data acquisition module 10 is used to acquire raw data from the ground-penetrating radar B-SCAN and road material parameters for the same road detection area.

[0070] The standardization preprocessing module 20 is used to perform standardization preprocessing on the raw data of ground penetrating radar B-SCAN and road material parameters respectively to obtain standardized radar data and standardized road material parameters.

[0071] The multimodal stitching module 30 is used to stitch together standardized radar data and standardized road material parameters to obtain multimodal features.

[0072] The joint input construction module 40 is used to construct joint input features based on the fusion of random noise vectors and geological constraint vectors using multimodal features.

[0073] The layered generation module 50 is used to input the joint input features into the three-level layered generator to generate ground penetrating radar B-SCAN single-channel data step by step.

[0074] The sample construction module 60 is used to stitch together ground-penetrating radar B-SCAN single-channel data with the corresponding batch of road material parameters to form a generated sample, and to stitch together ground-penetrating radar B-SCAN data of the real road detection area with the corresponding batch of real road material parameters as the real sample.

[0075] The adversarial training module 70 is used to input the generated samples and the real samples into the two-dimensional discriminator for adversarial discrimination, and to alternately iterate the training of the three-level hierarchical generator and the two-dimensional discriminator until the model meets the preset convergence condition, and outputs the current three-level hierarchical generator as the target three-level hierarchical generator.

[0076] The data generation module 80 is used to input the raw ground-penetrating radar B-SCAN data of the road detection area to be detected and the road material parameters into the target three-level layer generator to obtain the target ground-penetrating radar B-SCAN data.

[0077] The ground-penetrating radar (GPR) data generation device based on road material constraint adversarial neural networks provided in this application, employing the GPR data generation method based on road material constraint adversarial neural networks described in the above embodiments, can solve the technical problem of how to improve the accuracy of the generated GPR data. Compared with the prior art, the beneficial effects of the GPR data generation device based on road material constraint adversarial neural networks provided in this application are the same as those of the GPR data generation method based on road material constraint adversarial neural networks provided in the above embodiments, and other technical features in the GPR data generation device based on road material constraint adversarial neural networks are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0078] This application provides a ground-penetrating radar data generation device based on road material constraint adversarial neural network. The ground-penetrating radar data generation device based on road material constraint adversarial neural network includes: at least one processor; and a memory communicatively connected to at least one processor; wherein the memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to execute the ground-penetrating radar data generation method based on road material constraint adversarial neural network in the above embodiment 1.

[0079] The following is for reference. Figure 4 This document illustrates a schematic diagram of a ground-penetrating radar (GPR) data generation device suitable for implementing embodiments of this application based on a road material constraint adversarial neural network. The GPR data generation device in this application can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The ground-penetrating radar data generation device based on road material constraint adversarial neural network shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0080] like Figure 4As shown, the ground-penetrating radar data generation device based on road material constraint adversarial neural network may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in ROM 1002 (Read-Only Memory) or the program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the ground-penetrating radar data generation device based on road material constraint adversarial neural network. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the ground-penetrating radar data generation device based on road material constraint adversarial neural networks to communicate wirelessly or wiredly with other devices to exchange data. Although various ground-penetrating radar data generation devices based on road material constraint adversarial neural networks are shown in the figures, it should be understood that it is not required to implement or possess all of those shown. More or fewer may be implemented alternatively.

[0081] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0082] The ground-penetrating radar (GPR) data generation device based on road material constraint adversarial neural networks provided in this application, employing the GPR data generation method based on road material constraint adversarial neural networks described in the above embodiments, can solve the technical problem of how to improve the accuracy of the generated GPR data. Compared with the prior art, the beneficial effects of the GPR data generation device based on road material constraint adversarial neural networks provided in this application are the same as those of the GPR data generation method based on road material constraint adversarial neural networks provided in the above embodiments, and other technical features in this GPR data generation device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0083] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0084] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0085] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the ground-penetrating radar data generation method based on road material constraint adversarial neural network in the above embodiments.

[0086] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0087] The aforementioned computer-readable medium may be included in a ground-penetrating radar data generation device based on a road material constraint adversarial neural network; or it may exist independently and not be assembled into a ground-penetrating radar data generation device based on a road material constraint adversarial neural network.

[0088] The aforementioned computer-readable medium carries one or more programs that, when executed by a ground-penetrating radar data generation device based on a road material constraint adversarial neural network, enable the ground-penetrating radar data generation device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0089] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0090] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0091] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described ground-penetrating radar data generation method based on road material constraint adversarial neural networks, thereby solving the technical problem of how to improve the accuracy of the generated ground-penetrating radar data. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the ground-penetrating radar data generation method based on road material constraint adversarial neural networks provided in the above embodiments, and will not be repeated here.

[0092] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for generating ground-penetrating radar data based on a road material constraint adversarial neural network.

[0093] The computer program product provided in this application solves the technical problem of how to improve the accuracy of generated ground-penetrating radar data. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the ground-penetrating radar data generation method based on road material constraint adversarial neural network provided in the above embodiments, and will not be repeated here.

[0094] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for generating ground-penetrating radar data based on road material-constrained adversarial neural networks, characterized in that, The method includes: Acquire raw ground-penetrating radar B-SCAN data and road material parameters for the same road detection area; The raw data from the ground-penetrating radar B-SCAN and the road material parameters are respectively subjected to standardization preprocessing to obtain standardized radar data and standardized road material parameters. The standardized radar data and standardized road material parameters are spliced ​​together to obtain multimodal features; Based on the multimodal features, random noise vectors and geological constraint vectors are fused to construct joint input features; The combined input features are input into a three-level hierarchical generator to generate ground-penetrating radar B-SCAN single-channel data step by step. The ground-penetrating radar B-SCAN single-channel data is spliced ​​with the road material parameters of the corresponding batch to form a generated sample, and the ground-penetrating radar B-SCAN data of the real road detection area is spliced ​​with the real road material parameters of the corresponding batch as the real sample. The generated samples and the real samples are input into a two-dimensional discriminator for adversarial discrimination. The three-level hierarchical generator and the two-dimensional discriminator are trained alternately and iteratively until the model meets the preset convergence condition. The current three-level hierarchical generator is then output as the target three-level hierarchical generator. The raw ground-penetrating radar B-SCAN data of the road detection area to be detected and the road material parameters are input into the target three-level layer generator to obtain the target ground-penetrating radar B-SCAN data.

2. The ground-penetrating radar data generation method based on road material-constrained adversarial neural networks as described in claim 1, characterized in that, The step of performing standardization preprocessing on the raw ground-penetrating radar B-SCAN data and the road material parameters to obtain standardized radar data and standardized road material parameters includes: The discrete parameters in the road material parameters are converted into numerical features using one-hot encoding to obtain the encoded discrete parameters; The continuous parameters in the road material parameters are mapped to a preset numerical range through minimum-maximum normalization to obtain normalized continuous parameters; By fusing the encoded discrete parameters with the normalized continuous parameters, standardized road material parameters are obtained. Noise cancellation is performed on the raw data of the ground-penetrating radar B-SCAN to obtain denoised radar data. The denoised radar data is normalized in amplitude and mapped to a preset signal range, and phase calibration is performed to obtain standardized radar data.

3. The ground-penetrating radar data generation method based on road material-constrained adversarial neural networks as described in claim 1, characterized in that, The step of stitching together the standardized radar data and standardized road material parameters to obtain multimodal features includes: The sensitivity coefficients between ground-penetrating radar signals and various road material parameters were calculated using Pearson correlation analysis. Based on the sensitivity coefficient, generate the weight coefficients corresponding to each road material parameter channel; The standardized road material parameters are weighted and enhanced based on the weighting coefficients to obtain the enhanced road material parameters. The enhanced road material parameters and the standardized radar data are aligned in the channel dimension to obtain the aligned features. Perform channel splicing on the aligned features to obtain initial multimodal features; The initial multimodal features are validated and corrected in terms of feature dimensions to obtain the multimodal features.

4. The ground-penetrating radar data generation method based on road material-constrained adversarial neural networks as described in claim 1, characterized in that, The step of constructing joint input features based on the fusion of the multimodal feature random noise vector and the geological constraint vector includes: Random noise vectors that follow a normal distribution are randomly sampled according to a preset batch size, and geological constraint vectors corresponding to the batch are extracted from the road material parameter set. The random noise vector and the geological constraint vector are weighted and fused to obtain a noise-constraint fused vector; The noise-constraint fusion vector and the multimodal features are fused along their feature dimensions to obtain a fused feature vector; The fused feature vector is normalized to obtain normalized fused features; The normalized fusion features are mapped to the input dimension of the three-level hierarchical generator to obtain joint input features.

5. The ground-penetrating radar data generation method based on road material constraint adversarial neural network as described in claim 1, characterized in that, The step of inputting the joint input features into a three-level hierarchical generator to generate ground-penetrating radar B-SCAN single-channel data step by step includes: The joint input features are input into the geological framework layer of the three-level hierarchical generator to generate a preset low-resolution macroscopic geological framework feature map. The macroscopic geological framework feature map is input into the target contour layer of the three-level layer generator to generate a target contour feature map with a preset medium resolution. The target contour feature map is input into the signal detail layer of the three-level hierarchical generator to generate a preset high-resolution signal detail feature map. The signal detail feature map is subjected to feature smoothing and dimensionality compression to obtain a single-channel feature matrix; The single-channel feature matrix is ​​converted into the ground-penetrating radar B-SCAN data format to obtain ground-penetrating radar B-SCAN single-channel data.

6. The ground-penetrating radar data generation method based on road material constraint adversarial neural network as described in claim 5, characterized in that, The step of inputting the joint input features into the geological framework layer of the three-level hierarchical generator to generate a preset low-resolution macroscopic geological framework feature map includes: The joint input features are input into the geological parameter encoding module of the geological framework layer and converted into a global constraint feature map. The global constraint feature map is convolved using a large kernel convolution with a preset kernel size to obtain a convolutional feature map; The convolutional feature map is upsampled by multi-layer transposed convolution, batch normalization and activation function to obtain the upsampled feature map; The upsampled feature map is mapped to a preset low resolution to obtain an initial geological framework feature map. The road material constraint adaptability of the initial geological framework feature map is verified to obtain a macroscopic geological framework feature map. The step of inputting the macroscopic geological framework feature map into the target contour layer of the three-level layer generator to generate a target contour feature map with a preset medium resolution includes: The macroscopic geological framework feature map is upsampled by transposed convolution to obtain a medium-resolution geological feature map with a preset medium resolution. Based on the standardized road material parameters, a target candidate region extraction operation is performed on the medium-resolution geological feature map, and a target candidate mask is output. The features within the target candidate mask are processed by reinforcement learning through the boundary constraint convolution module of the target contour layer to obtain an enhanced feature map. A geological compatibility verification operation for road material constraints is performed on the enhanced feature map to obtain a verified medium-resolution feature map. The verified medium-resolution feature map is subjected to feature fusion and smoothing to generate a target contour feature map with a preset medium resolution. The step of inputting the target contour feature map into the signal detail layer of a three-level hierarchical generator to generate a preset high-resolution signal detail feature map includes: The target contour feature map is upsampled to a preset high resolution through transposed convolution to obtain a high-resolution contour feature map; The radar echo intensity is simulated based on the finite-difference time-domain model of the high-resolution contour feature map to generate a signal amplitude feature map. An adaptive clutter signal is generated based on the noise statistical distribution of real ground-penetrating radar data to obtain adaptive clutter characteristics. The adaptive clutter features are superimposed on the signal amplitude feature map to obtain a clutter-containing signal feature map. Phase and amplitude corrections are performed on the clutter-containing signal feature map to obtain a preset high-resolution signal detail feature map.

7. The ground-penetrating radar data generation method based on road material-constrained adversarial neural networks as described in claim 1, characterized in that, The steps of inputting the generated samples and the real samples into a two-dimensional discriminator for adversarial discrimination, iteratively training the three-level hierarchical generator and the two-dimensional discriminator alternately until the model meets the preset convergence condition, and outputting the current three-level hierarchical generator as the target three-level hierarchical generator include: The generated samples and the real samples are input into a two-dimensional discriminator to extract two-dimensional features of geological structures and radar signals, respectively, to obtain the feature set of the generated samples and the feature set of the real samples. The generated sample feature set and the real sample feature set are compared and evaluated by a two-dimensional discriminator, and the authenticity discrimination score and the real sample discrimination score are output. The adversarial loss function values ​​of the three-level hierarchical generator and the two-dimensional discriminator are calculated respectively. The network parameters of the three-level hierarchical generator and the two-dimensional discriminator are updated according to the adversarial loss function value by backpropagation algorithm to obtain an optimized generator. When the fluctuation range of the adversarial loss function value within a consecutive preset iteration batch is less than a preset threshold, and the authenticity discrimination score of the generated sample exceeds the preset proportion of the authenticity discrimination score of the real sample, the current optimized generator is output as the target three-level hierarchical generator.

8. A ground-penetrating radar data generation device based on road material constraint adversarial neural network, characterized in that, The device includes: The data acquisition module is used to acquire raw ground-penetrating radar B-SCAN data and road material parameters for the same road detection area; The standardization preprocessing module is used to perform standardization preprocessing on the raw data of the ground penetrating radar B-SCAN and the road material parameters respectively to obtain standardized radar data and standardized road material parameters. A multimodal stitching module is used to stitch together the standardized radar data and standardized road material parameters to obtain multimodal features; The joint input construction module is used to construct joint input features based on the fusion of random noise vectors and geological constraint vectors of the multimodal features; The layered generation module is used to input the joint input features into the three-level layered generator to generate ground-penetrating radar B-SCAN single-channel data step by step. The sample construction module is used to stitch together the ground-penetrating radar B-SCAN single-channel data with the road material parameters of the corresponding batch to form a sample, and to stitch together the ground-penetrating radar B-SCAN data of the real road detection area with the real road material parameters of the corresponding batch as the real sample. The adversarial training module is used to input the generated samples and the real samples into the two-dimensional discriminator for adversarial discrimination, and to alternately iterate the training of the three-level hierarchical generator and the two-dimensional discriminator until the model meets the preset convergence condition, and outputs the current three-level hierarchical generator as the target three-level hierarchical generator. The data generation module is used to input the raw ground-penetrating radar B-SCAN data of the road detection area to be detected and the road material parameters into the target three-level layer generator to obtain the target ground-penetrating radar B-SCAN data.

9. A ground-penetrating radar data generation device based on road material constraint adversarial neural network, characterized in that, The device includes: a memory, a processor, and a ground-penetrating radar data generation program based on a road material constraint adversarial neural network stored in the memory and running on the processor, wherein the ground-penetrating radar data generation program based on a road material constraint adversarial neural network is configured to implement the steps of the ground-penetrating radar data generation method based on a road material constraint adversarial neural network as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a ground-penetrating radar data generation program based on a road material constraint adversarial neural network. When the ground-penetrating radar data generation program based on a road material constraint adversarial neural network is executed by a processor, it implements the steps of the ground-penetrating radar data generation method based on a road material constraint adversarial neural network as described in any one of claims 1-7.

Images