An asphalt pavement internal crack three-dimensional inversion method, system, device and storage medium

Abstract

The present application relates to the field of road nondestructive testing and ground penetrating radar intelligent inversion, in particular to a three-dimensional inversion method, system and device for internal cracks of asphalt pavement and a storage medium, 24-channel 800MHz three-dimensional ground penetrating radar detection data is acquired, decoupling reconstruction is performed to obtain a three-dimensional field strength data body, and a geometric prior is constructed based on crack area positioning results; a GPR Trans-InvNet network is constructed, a CAE-Block encoder is used to extract multi-dimensional features, a lightweight Transformer bridging module is used to realize global context modeling, and a decoder with an Attention Gate is used to complete high-resolution reconstruction; the network is trained through a composite loss of electromagnetic interpretability, geometric similarity and numerical consistency, and a three-dimensional distribution of crack equivalent dielectric constant is output. The present application converts unconstrained inversion into limited inversion, and takes into account accuracy, efficiency and physical interpretability, and is suitable for engineering scenarios such as road preventive detection, disease survey and emergency detection.

Description

Technical Field

[0001] This invention relates to the field of road non-destructive testing and ground-penetrating radar intelligent inversion technology, specifically to a three-dimensional inversion method, system, equipment, and storage medium for internal cracks in asphalt pavement. Background Technology

[0002] Under the long-term coupled effects of vehicle loads, temperature stress, and water damage, asphalt pavements are prone to developing hidden internal cracks. These cracks are initially invisible to the naked eye before they penetrate to the road surface. If not detected and treated promptly, they will continue to extend upwards, forming surface cracks, which can then lead to structural defects such as network cracking, pumping, base layer softening, and even pavement subsidence. This severely shortens the road's service life and significantly increases the total life-cycle maintenance cost. Ground-penetrating radar (GPR), as a highly efficient, non-contact, and non-destructive geophysical detection technology, transmits high-frequency electromagnetic waves underground and receives reflected echoes from interfaces with different dielectric properties. This enables continuous and rapid detection of the pavement's internal structure, making it the mainstream technology for detecting internal road defects.

[0003] Existing ground-penetrating radar (GPR) methods for detecting road cracks are mainly divided into two categories: identification methods and inversion methods. Identification methods use machine learning or deep learning algorithms to extract hyperbolic anomaly echo features from radar B-scan images to locate cracks. However, they can only determine the approximate planar location of cracks and cannot obtain their three-dimensional geometric dimensions and dielectric physical parameters. Furthermore, they are susceptible to radar echo ambiguity and noise interference, resulting in high false detection and false negative rates. Inversion methods aim to infer the dielectric constant distribution of the underground medium by observing electromagnetic wave signals, thereby reconstructing the spatial morphology of cracks. Compared to identification methods, they have clearer physical meaning and higher detection accuracy, representing the current development direction of GPR defect detection technology.

[0004] Among existing inversion-related technologies, prior art document CN118586067A discloses a method, system, equipment, and storage medium for inverting the depth of road surface cracks. This method simultaneously acquires visible light image data of road surface cracks and single-channel two-dimensional echo data from ground-penetrating radar. It extracts geometric features such as the length and width of the surface cracks, as well as waveform features such as the two-way travel time, peak amplitude, and phase shift of the radar echoes. A gradient boosting regression model with multi-feature fusion is then constructed to achieve rapid estimation of the surface crack depth. This method improves the automation level of surface crack depth detection to a certain extent and reduces the subjective error of manual interpretation.

[0005] However, the aforementioned comparative documents have the following significant technical limitations: First, this method is only applicable to visible cracks that have extended to the road surface and cannot detect hidden cracks that are not penetrating inside the road surface. These early cracks are the core control targets of preventive road maintenance. Second, this method can only output the single-point depth estimate of the crack and cannot reconstruct the three-dimensional spatial distribution of the crack and the differences in its internal dielectric properties, making it difficult to comprehensively assess the crack's expansion trend and the degree of structural hazard. Third, this method relies heavily on visible light images to extract surface crack features. Under complex engineering conditions such as oil stains, marking obstructions, water accumulation, and dust on the road surface, the accuracy of image feature extraction drops significantly, directly leading to the failure of the inversion results.

[0006] In addition, existing ground-penetrating radar (GPR) fracture inversion technologies still suffer from several common problems: Physics-driven full-waveform inversion methods, which construct inversion models based on Maxwell's equations and possess strong electromagnetic interpretability, suffer from exponentially increasing computational costs for three-dimensional full-waveform inversion, requiring tens of hours to process data per kilometer, thus failing to meet the needs of large-scale engineering surveys; Data-driven deep learning inversion methods, while computationally efficient, rely on local convolutional neural networks to expand the receptive field layer by layer, resulting in insufficient modeling of long-range dependencies across scan lines and time windows, making it difficult to accurately characterize the overall connectivity of fractures; Standard Transformer models, while possessing global context modeling capabilities, have a large number of parameters and high training sample requirements, making them prone to overfitting under small-sample conditions in GPR measurements. Furthermore, existing deep learning inversion methods often employ purely data-driven mean square error or L1 loss functions, lacking the collaborative constraints of electromagnetic propagation laws and fracture geometric priors, leading to severe ambiguity and spurious responses in the inversion results, poor physical interpretability, and difficulty in direct application to engineering practice.

[0007] Therefore, there is an urgent need to develop a three-dimensional inversion method for internal cracks in asphalt pavement that can simultaneously detect hidden cracks inside the pavement, achieve high-precision reconstruction of the three-dimensional spatial morphology of cracks, and balance inversion efficiency and physical interpretability, so as to meet the actual needs of preventive maintenance and accurate detection of structural defects in road engineering. Summary of the Invention

[0008] The purpose of this invention is to provide a three-dimensional inversion method, system, device and storage medium for internal cracks in asphalt pavement, so as to solve the problems mentioned in the background art, such as the inability of existing technologies to detect hidden internal cracks, strong inversion ambiguity, poor physical interpretability and insufficient engineering applicability.

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

[0010] A three-dimensional inversion method for internal cracks in asphalt pavement includes the following steps:

[0011] S1. Obtain three-dimensional ground-penetrating radar detection data of the internal crack area of ​​asphalt pavement, and decouple the three-dimensional ground-penetrating radar detection data into a multi-channel two-dimensional field strength matrix;

[0012] S2. Based on the crack region location results, obtain the approximate distribution range of the cracks, and construct the approximate distribution range of the cracks as geometric prior information;

[0013] S3. The multi-channel two-dimensional field strength matrix is ​​preprocessed and then reconstructed according to the antenna channel dimension to obtain the three-dimensional field strength data volume.

[0014] S4. Construct a three-dimensional dielectric constant inversion network GPR Trans-InvNet. The network includes an encoder, a bridging module, and a decoder. The encoder is composed of multiple CAE-Blocks. The bridging module is a lightweight Transformer bridging module. The decoder includes multiple upsampling units, an Attention Gate gating unit, and a CAE-Block fusion unit.

[0015] S5. The three-dimensional dielectric constant inversion network is trained using a composite loss function, which includes at least electromagnetic interpretability loss, geometric similarity loss, and numerical consistency loss.

[0016] S6. Output the three-dimensional distribution of the equivalent dielectric constant of the crack, which corresponds one-to-one with the spatial location of the three-dimensional field strength data volume.

[0017] Preferably, the three-dimensional ground-penetrating radar detection data in step S1 is acquired by 24 antenna channels, and the center frequency of the ground-penetrating radar is 800 MHz;

[0018] In step S2, the approximate distribution range of the cracks is determined by the coordinates of the four corner points output by the existing crack region localization network, forming the enclosing region ROI in the two-dimensional field strength matrix;

[0019] The three-dimensional field strength data volume in step S3 consists of a time sampling dimension, a scan line dimension along the detection direction, and an antenna channel dimension; the size of the three-dimensional field strength data volume obtained in step S3 is 512×304×24; the preprocessing in step S3 includes at least direct wave removal processing and normalization processing.

[0020] Preferably, the CAE-Block in step S4 includes at least: a coordinate convolution unit for constructing a normalized coordinate field and concatenating it with the input field strength features; an anisotropic direction convolution unit for extracting direction-sensitive features along the time axis, scan line axis, and antenna channel axis, respectively; an antenna axis attention unit for adaptive recalibration of the responses of different antenna channels; and a residual fusion and multi-stage feature concatenation unit for fusing the current layer features with the residual branch features and outputting the current layer's encoded or decoded features.

[0021] The lightweight Transformer bridging module in step S4 flattens the 3D feature volume output by the encoder into a token sequence, performs global context modeling through at least two Transformer Encoder Layers, and then maps it back to the 3D spatial domain. The lightweight Transformer bridging module does not have additional explicit position encoding; it achieves position awareness by relying on the 3D coordinate information injected by the encoder.

[0022] In step S4, the decoder performs spatial gating filtering on the high-resolution features output by the corresponding encoder layer through the Attention Gate at each decoding stage to obtain gated skip connection features. The gated skip connection features are then concatenated with the upsampled decoding features in the channel dimension and input into the CAE-Block fusion unit for feature fusion and detail reconstruction.

[0023] As a preferred embodiment, the electromagnetic interpretability loss in step S5 constructs the damped wave equation residuals between the observed electric field and the predicted dielectric constant for the two-dimensional spatiotemporal profiles corresponding to each antenna channel, and applies Huber loss as a constraint on the internal region after removing the boundaries. The calculation formula is as follows:

[0024] The damped wave equation is defined as:

[0025]

[0026] In the formula, E is the scalar electric field component in the two-dimensional spatiotemporal domain; t is the electromagnetic wave propagation time; x, y are the two-dimensional spatial coordinates; ε(t,x,y) is the spatiotemporally dependent dielectric constant; ∇ 2 x,y are two-dimensional Laplace operators describing the spatial diffusion and propagation characteristics of the electric field; σ is the conductivity of the medium; s(t,x,y) are reflection source and boundary-related terms;

[0027] The residual of the damped wave equation is defined as:

[0028]

[0029] In the formula, r(t,x,y) is the residual of the damped wave equation; E(t,x,y) is the electric field intensity component in the spatiotemporal domain. This represents the predicted dielectric constant obtained from network inversion;

[0030] The partial derivatives in the above residuals are all calculated on the discrete grid using second-order central difference, and the calculation formula is as follows:

[0031]

[0032]

[0033]

[0034]

[0035] In the formula, ∆t is the time step, and ∆x and ∆y are the spatial step sizes in the x and y directions, respectively;

[0036] Within the spatiotemporal computational domain after boundary removal, the Huber function is used to constrain the discrete-form damped wave equation residuals, thereby constructing an electromagnetic interpretability loss function:

[0037]

[0038] In the formula, L FDTD This is the physical constraint loss function constructed based on the FDTD discrete scheme; ∑ is the total number of sampling points in the spatiotemporal computational domain after removing boundaries; (t,x,y)∈Ω is the summation operator for all spatiotemporal points within the computational domain; 𝜌(⋅) is the Huber loss function, used to reduce the impact of abnormal residuals on the training process.

[0039] Preferably, the geometric similarity loss in step S5 is calculated by mapping the predicted dielectric constant to a differentiable soft mask, and combining it with the geometric prior mask corresponding to the approximate distribution range of the crack to construct an external penalty term and an internal coverage term for the ROI. The calculation formula is as follows:

[0040] For each channel index y, the geometric truth mask for the crack extent is defined as:

[0041]

[0042] In the formula, M roi (t,x,y) represents the geometric ground truth mask of the crack region in the GPR image; (t,x) represents the temporal-spatial coordinates of the GPR profile; M roi (t,x,y)=1 indicates that the coordinate point is located within the calibrated crack area; M roi (t,x,y)=0 indicates that the coordinate point is located outside the crack region;

[0043] A differentiable soft mask is generated using Cauchy mapping, and the calculation formula for the differentiable soft mask is as follows:

[0044]

[0045] In the formula, P(t,x,y) is a differentiable soft mask obtained by mapping the predicted dielectric constant; ε c τ is the reference center value of the crack dielectric constant, used to construct the geometric probability mapping; τ is the adjustment parameter that controls the weight decay rate.

[0046] Define an ROI-out penalty loss to suppress the expansion of the predicted crack region beyond the actual crack extent:

[0047]

[0048] In the formula, L out ε is the external penalty loss for ROI; ε is a smoothing term to prevent the denominator from being zero.

[0049] Define an in-ROI coverage loss to prevent the network from circumventing penalties by lowering the overall prediction probability:

[0050]

[0051] In the formula, L in This refers to the coverage loss within the ROI;

[0052] Ultimately, the geometric similarity loss consists of the two losses mentioned above:

[0053]

[0054] In the formula, L shape For geometric similarity loss; λ out For and λ in Let λ be the weighting coefficient for the two ROIs. out ≥λ in Prioritize suppressing spurious responses outside the ROI.

[0055] Preferably, the numerical consistency loss in step S5 is a regionalized L1 loss calculated only within the area defined by the crack voxel mask, and the calculation formula is as follows:

[0056] Crack Voxel Mask crack The set of voxels used to mark the cracks in the dielectric truth is defined as:

[0057]

[0058] In the formula, M crack (t,x,y) represents the voxel mask for the crack; ε gt (t,x,y) represents the target dielectric constant of the crack region; εc δ is the reference center value for the crack dielectric constant; δ is the tolerance threshold used to distinguish crack voxels. This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise.

[0059] The numerical consistency loss is defined as:

[0060]

[0061] In the formula, L num The numerical consistency loss for the crack region applies only to crack voxels and does not impose constraints on background voxels. The background region is constrained by both electromagnetic consistency loss and geometric consistency loss. ε represents the predicted dielectric constant obtained from network inversion; gt (t,x,y) represents the target dielectric constant of the crack region; the numerator is the sum of the absolute errors between the predicted dielectric constant and the target dielectric constant within the crack region, and the denominator is the total number of voxels in the crack region, used to normalize the loss value and eliminate the influence of crack size on the loss calculation.

[0062] Preferably, in step S6, a 1×1×1 three-dimensional convolution operator is first used to map the multi-channel features into a single-channel initial predicted dielectric constant volume; then, the initial predicted dielectric constant volume is corrected in three dimensions by trilinear interpolation, and finally, the three-dimensional distribution of the equivalent dielectric constant of the crack is output, which corresponds one-to-one with the three-dimensional field strength data volume in spatial location.

[0063] On the other hand, the present invention also provides a three-dimensional ground-penetrating radar inversion system for internal cracks in asphalt pavement, used in the aforementioned three-dimensional ground-penetrating radar inversion method for internal cracks in asphalt pavement, comprising:

[0064] The data acquisition and decoupling module is used to acquire three-dimensional ground-penetrating radar detection data of the crack area inside the asphalt pavement, and decouple the detection data into a multi-channel two-dimensional field strength matrix.

[0065] The geometric prior generation module is used to obtain the approximate distribution range of cracks based on the crack region location results, and to construct the approximate distribution range of cracks as geometric prior information.

[0066] The data preprocessing and reconstruction module is used to perform direct wave removal and normalization processing on the multi-channel two-dimensional field strength matrix, and reconstruct it into a three-dimensional field strength data volume according to the antenna channel dimension.

[0067] Network coding module: used to extract multi-scale features from the three-dimensional field strength data volume using an encoder composed of multi-level CAE-Blocks;

[0068] Network bridging module: Used to perform global context modeling of encoder output features using a lightweight Transformer bridging module;

[0069] Network Decoding and Output Correction Module: Used to decode the bridging features step by step using a decoder including an Attention Gate gating unit, and output the three-dimensional distribution of the equivalent dielectric constant of the crack corresponding to the spatial position of the three-dimensional field strength data volume through 1×1×1 three-dimensional convolution and trilinear interpolation.

[0070] Composite loss training module: used to train the three-dimensional dielectric constant inversion network using a composite loss function that includes at least electromagnetic interpretability loss, geometric similarity loss and numerical consistency loss.

[0071] Furthermore, the present invention provides an electronic device employing the above-mentioned three-dimensional ground-penetrating radar inversion system for internal cracks in asphalt pavement, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the above-mentioned three-dimensional inversion method for internal cracks in asphalt pavement.

[0072] In addition, the present invention provides a computer-readable storage medium storing a computer program employing the above-described three-dimensional ground-penetrating radar inversion system for internal cracks in asphalt pavement, wherein the computer program, when executed by a processor, implements the above-described three-dimensional inversion method for internal cracks in asphalt pavement.

[0073] Compared with the prior art, the beneficial effects of the present invention are:

[0074] 1. This invention, by incorporating coordinate convolution, anisotropic directional convolution, and antenna axis attention mechanisms within a CAE-Block, enables the network to simultaneously utilize spatial location information, directional variation information, and multi-antenna channel difference information in the field strength response for feature extraction. Specifically, coordinate convolution establishes the correlation between reflection features and location coordinates; anisotropic directional convolution characterizes the response patterns along the time direction, scan line direction, and antenna direction, respectively; and the antenna axis attention mechanism adaptively weights the features of different antenna channels, thus enhancing the ability to characterize crack-related reflection features.

[0075] 2. This invention utilizes a lightweight Transformer bridging module to perform global context modeling of the deep features output by the encoder. Because this module can establish cross-positional information interaction relationships between the flattened token sequences, it can further supplement long-range dependency information across time windows and scan lines based on the results extracted by local convolution, thereby helping to improve the representation ability of the overall distribution relationship and continuity features of cracks.

[0076] 3. This invention introduces a gating and filtering mechanism for skip connection features during the decoding stage, ensuring that the high-resolution features output by the encoder are filtered for target relevance before being fused with the decoding end. Based on this filtering process, the impact of background noise and irrelevant layered reflections directly entering the decoding and fusion process can be reduced, thereby helping to balance high-resolution detail recovery and non-target response suppression.

[0077] 4. This invention employs a composite loss function consisting of electromagnetic interpretability loss, geometric similarity loss, and numerical consistency loss. Electromagnetic interpretability loss constrains the consistency between the predicted results and the electromagnetic propagation laws; geometric similarity loss constrains the consistency between the predicted results and the prior geometric range of the crack; and numerical consistency loss constrains the consistency between the predicted dielectric constant value and the target value within the crack region. Therefore, it provides joint constraints on the inversion process from three aspects: physical laws, spatial structure, and local numerical values. Attached Figure Description

[0078] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0079] Figure 1 This is a flowchart illustrating a specific method for three-dimensional inversion of internal cracks in asphalt pavement, including a method, system, equipment, and storage medium, as described in an embodiment of the present invention.

[0080] Figure 2 This is a schematic diagram of the dataset construction process in an embodiment of the present invention, illustrating the data acquisition, data decoupling and preprocessing, crack region location, crack dielectric constant assignment, background dielectric constant assignment, and dataset construction process.

[0081] Figure 3 This is a schematic diagram of the overall network framework of GPR Trans-InvNet in an embodiment of the present invention, showing a three-dimensional ground-penetrating radar inversion network structure consisting of a multi-level CAE-Block encoder, a lightweight Transformer bridging module, and an Attention Gate decoder.

[0082] Figure 4 This is a schematic diagram of the CAE-Block structure in an embodiment of the present invention, illustrating a combination of three-dimensional coordinate convolution, anisotropic directional convolution, antenna axis attention mechanism, residual fusion, and multi-stage feature stitching.

[0083] Figure 5 This is a schematic diagram of the lightweight Transformer bridging module structure in an embodiment of the present invention, showing the process of flattening a three-dimensional feature volume into a token sequence, performing global context modeling through multi-head self-attention, and then mapping it back to the three-dimensional spatial domain.

[0084] Figure 6 This is a schematic diagram of the Attention Gate-based decoder structure in an embodiment of the present invention, illustrating the process of fusing and reconstructing upsampled features with gated skip connection features. Detailed Implementation

[0085] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0086] like Figure 1 As shown, this invention provides a method, system, device, and storage medium for three-dimensional inversion of internal cracks in asphalt pavement. The method first collects three-dimensional ground-penetrating radar detection data of the target road section, decouples the collected data into a multi-channel two-dimensional field strength matrix, and reconstructs it into a three-dimensional field strength data volume according to the antenna channel dimension after preprocessing. Then, the approximate distribution range of cracks is obtained as a geometric prior based on the crack area location results. Next, the three-dimensional field strength data volume is input into a three-dimensional dielectric constant inversion network, and after encoding, bridging, and decoding, the equivalent dielectric constant of the crack is output as a three-dimensional prediction volume. Finally, the training is completed through a composite loss function, and the final inversion result is output.

[0087] like Figure 2 As shown, this invention uses a three-dimensional ground-penetrating radar (GPR) to detect target road sections. The GPR used has 24 antenna channels and a center frequency of 800 MHz. First, the acquired three-dimensional GPR data is decoupled into 24 single-channel measurement lines, each of which is a two-dimensional field strength matrix with a size of 512×304. Subsequently, the 24 two-dimensional field strength matrices are stacked and reconstructed according to the antenna channel dimension to form a three-dimensional field strength data volume with a size of 512×304×24, which is used as the input of the inversion network.

[0088] In this invention, crack region localization is performed using an existing internal crack size detection network. First, the localization network outputs the coordinates of the four corner points of the crack region on the two-dimensional field strength matrix of each channel, thereby determining the Region of Interest (ROI) enclosing the crack within the two-dimensional field strength matrix. This ROI serves as a geometric prior, used to constrain the search space for dielectric inversion.

[0089] In this invention, the crack region location result can be obtained by a pre-trained internal crack size detection network. This internal crack size detection network is not the core content claimed in this application; its function is to output the coordinates of the four corner points of the crack region for the two-dimensional field strength matrix corresponding to each antenna channel, thereby determining the approximate distribution area (ROI) of the crack. The ROI serves as a geometric prior, used to limit the search space for dielectric inversion.

[0090] like Figure 4 As shown, the encoder of this invention employs a four-level downsampling structure, with each level using the CAE-Block for feature extraction. First, three-dimensional coordinate convolution is performed, constructing a normalized coordinate field and concatenating it with the input field strength data along the channel dimension to enhance the network's ability to perceive absolute position. Second, anisotropic directional convolution is performed, using three-dimensional directional convolutions with kernel sizes of (3, 1, 1), (1, 3, 1), and (1, 1, 3) to model the time direction, scan line direction, and antenna channel direction respectively. Third, antenna axis attention recalibration is performed, establishing local dependencies between adjacent antenna channels through one-dimensional convolution, generating channel weight vectors, and applying them to the original features. Finally, residual fusion and multi-stage feature concatenation are performed to form the output features of the current layer's CAE-Block.

[0091] like Figure 5 As shown, this invention employs a lightweight Transformer bridging module based on a self-attention mechanism to globally model the deepest layer output features of the encoder. In this invention, the bridging module achieves global information exchange of the token sequence through two Transformer EncoderLayers and four attention heads. First, the 3D feature volume output from the deepest layer of the encoder is flattened into a token sequence; then, this sequence is passed sequentially through two Transformer Encoder Layers; finally, the updated token sequence is remapped back to the 3D spatial domain, resulting in bridging features enhanced with global context. Since 3D coordinate information has already been explicitly injected at the encoder entry point through coordinate convolution, the bridging module does not need to introduce additional explicit positional encoding.

[0092] like Figure 6As shown, the decoder of this invention includes a multi-level upsampling unit, an Attention Gate gating unit, and a CAE-Block fusion unit. The 3D features output by the bridging module are first upsampled using a 3D transposed convolution with a kernel size of 2×2×2 and a stride of 2 to restore spatial resolution and enhance detail reconstruction. At each decoding stage, the upsampled decoded features are fused with the high-resolution skip connection features of the corresponding encoder layer. To avoid introducing a large amount of background noise and layered reflections unrelated to cracks directly into the decoding process, spatial gating is performed on the skip connection features using an Attention Gate before fusion. Subsequently, the gated skip features are concatenated with the upsampled features along the channel dimension and fed into the current-level CAE-Block for feature fusion and detail reconstruction. After multi-level upsampling and gating fusion, a 1×1×1 3D convolution is used at the top-level decoder output to map the multi-channel features into a single-channel dielectric constant prediction volume, and size correction is performed using trilinear interpolation.

[0093] This invention employs a composite loss function with electromagnetic-geometric-numerical coupling constraints. Since the three-dimensional field strength data volume is formed by stacking two-dimensional profiles corresponding to multiple antenna channels along the channel dimension, the electromagnetic interpretability loss in this invention is implemented using a channel-by-channel two-dimensional physical constraint approach. Specifically, damped wave equation residuals are constructed for each two-dimensional spatiotemporal profile corresponding to each antenna channel, and Huber loss is calculated in the internal region after boundary removal. The geometric similarity loss utilizes the ROI output by the crack location network as a geometric prior, and maps the continuous dielectric constant prediction values ​​to a differentiable soft mask through Cauchy mapping. After obtaining the differentiable soft mask P obtained from the predicted dielectric constant mapping and the geometric truth mask Mroi, the geometric similarity loss is constructed by introducing two complementary constraints: an external penalty loss and an internal coverage loss. The numerical consistency loss calculates the L1 error between the predicted dielectric constant and the reference dielectric constant only within the region defined by the crack voxel mask.

[0094] In this invention, the dataset construction includes the following steps: First, three-dimensional GPR measured data are collected; second, decoupling and preprocessing are performed; then, the Region of Interest (ROI) is obtained based on the crack area localization network; next, based on the amplitude and phase reversal characteristics of the crack reflected waves, the internal environment of the crack is judged manually and empirically, and divided into two typical states: dry and water-saturated; within the crack ROI, the equivalent dielectric constant of the crack is assigned values ​​of 1 and 81, respectively; for the background area outside the crack, non-uniform random values ​​are assigned according to the road structure layers, with the 99.7% confidence interval for the dielectric constant of the asphalt surface layer being (4.5610, 5.7476) and the 99.7% confidence interval for the dielectric constant of the base layer being (8.5583, 11.5025). Finally, the two-dimensional field strength matrix is ​​reconstructed into a three-dimensional data volume of size 512×304×24 as the network input, and the corresponding three-dimensional dielectric constant distribution is used as the label. Ultimately, 200 sets of three-dimensional samples are constructed and divided into training, validation, and test sets in a 6:2:2 ratio.

[0095] In this invention, the network is implemented and trained using PyTorch 1.9.1, with a training platform consisting of a single NVIDIA GeForce RTX 4090D GPU and 128 GB of system memory. The optimizer employs the Adam algorithm, with a batch size of 1 and an initial learning rate of 1×10⁻⁶. -3 And gradually decay to 1×10 according to the cosine annealing strategy. -5 .

[0096] The following embodiments are all based on the three-dimensional inversion method for internal cracks in asphalt pavement and the GPR Trans-InvNet network described above, covering three typical engineering scenarios: preventive inspection of newly built roads, general survey of structural defects in old roads, and emergency defect detection after rain, to verify the applicability and reliability of the present invention under different working conditions.

[0097] Example 1: Preventive Detection of Early Microcracks in Newly Built Asphalt Pavements of Expressways

[0098] Application Background:

[0099] A newly built six-lane, two-way expressway (design speed 120 km / h) has been open to traffic for six months. While there are no visible cracks on the road surface, uneven compaction in some sections during construction may have led to early internal microcracks (<2 mm in width). Traditional ground-penetrating radar (GPR) identification methods are insensitive to the echo characteristics of these microcracks, resulting in a high rate of missed detections and failing to meet the needs of preventative maintenance.

[0100] Detection equipment and parameters: A 24-channel 800MHz three-dimensional ground-penetrating radar system is used. Detection parameters are set as follows: time sampling interval 0.1ns, number of sampling points per channel 512; scan line spacing 2cm, single segment detection length 6.08m (corresponding to 304 scan lines); detection vehicle speed 5km / h, lateral coverage width 46cm, continuous detection along the wheel track of the driving lane.

[0101] Implementation process

[0102] 1. Collect 3D ground-penetrating radar data for a total of 300m from K123+400 to K123+700, decouple it into a 24-channel 2D field strength matrix, and complete the preprocessing of removing direct waves and Z-score normalization.

[0103] 2. An improved YOLOv8 crack localization network is used to locate microcracks in each channel's two-dimensional profile, outputting the coordinates of the four corner points of the crack ROI, and constructing a three-dimensional geometric prior mask.

[0104] 3. Input the preprocessed three-dimensional field strength data volume (512×304×24) into the trained GPRTrans-InvNet network, and use a two-stage training strategy to complete the model fine-tuning (based on 5 sets of borehole calibration data for this road section).

[0105] 4. Output the three-dimensional distribution of the equivalent dielectric constant of the crack, extract the crack voxels by threshold segmentation (cracks with a dielectric constant < 3 are judged as dry microcracks), and calculate the crack length, width, depth and spatial orientation.

[0106] Inversion Results and Analysis: A total of 17 internal microcracks were detected, including 11 cracks with a length of 0.5-1.0m and 6 cracks with a length of 1.0-1.5m, with an average width of 1.2mm and an average depth of 6.8cm. Three cracks were randomly selected for core drilling verification. The prediction errors for crack location and depth were both less than 3%, and the prediction error for width was less than 5%. The traditional B-scan manual identification method only detected 7 cracks, resulting in a false negative rate of 58.8%.

[0107] Engineering application value:

[0108] Early detection of micro-cracks inside the pavement allows maintenance units to be guided to use crack sealing treatment, preventing cracks from expanding into structural defects, extending the service life of the pavement by 3 to 5 years, and reducing the total life cycle maintenance cost by about 40%.

[0109] Example 2: Survey of Structural Cracks in Asphalt Pavements of Old Trunk Highways

[0110] Application Background: A provincial trunk highway, 12 years old, has developed numerous longitudinal and transverse surface cracks, with some sections exhibiting network cracking. Internal cracks have extended to the base layer, severely impacting the road's load-bearing capacity. Traditional full-waveform inversion methods involve extremely high 3D computation costs, requiring over 24 hours to process data per kilometer, making them unsuitable for large-scale surveys.

[0111] Detection equipment and parameters: The same model of 24-channel 800MHz three-dimensional ground penetrating radar is used. Detection parameters: time sampling interval 0.1ns, number of sampling points per channel 512; scan line spacing 2cm, single segment detection length 6.08m; detection vehicle speed 8km / h, 2 detection lines per lane (wheel track + lane center), covering the entire lane width.

[0112] Implementation process

[0113] 1. Complete the acquisition of 5km of three-dimensional ground-penetrating radar data for the road surface from K35+000 to K40+000, and perform batch decoupling and preprocessing in blocks of 6.08m each.

[0114] 2. The crack localization network is run in batches to generate geometric prior masks for each data block. GPU parallel computing technology is used to process the inversion task of 8 data blocks at the same time.

[0115] 3. Input the GPRTrans-InvNet network to perform batch 3D inversion, output the 3D distribution of crack dielectric constant for each data block, and stitch them together to generate a 3D crack distribution map of the entire road section.

[0116] 4. Differentiate crack types based on dielectric constant values ​​(dielectric constant 1-3 indicates dry cracks, 70-90 indicates water-saturated cracks), and count the number and distribution density of cracks of different depths and types.

[0117] Inversion Results and Analysis: A total of 213 internal structural cracks were detected, including 89 cracks penetrating the asphalt surface layer, 37 cracks extending to the base layer, and 42 water-saturated cracks. The total data processing time per kilometer was approximately 1.2 hours, representing a 95% improvement in efficiency compared to the traditional full-waveform inversion method. Compared with the results from 15 borehole core samples, the average error in crack depth prediction was 2.9%, and the accuracy rate for identifying base layer cracks reached 92%.

[0118] Engineering application value: It can quickly complete a large-scale survey of internal cracks in the pavement, accurately locate structural defects in the base layer, and provide data support for maintenance units to formulate targeted treatment plans such as milling and repaving and base layer reinforcement, thus avoiding the waste of funds caused by blind maintenance.

[0119] Example 3: Emergency Detection of Water-Saturated Cracks in Asphalt Pavements After Rain

[0120] Application Background: After 24 hours of continuous heavy rainfall, some sections of a city's expressway experienced pavement pumping, suspected to be caused by water accumulation in internal cracks leading to softening of the base layer. Damage inspection needs to be completed within 24 hours of the rain to determine the location and extent of the water-filled cracks, and to guide emergency drainage and repair work to prevent pavement subsidence and damage.

[0121] Detection equipment and parameters: A 24-channel 800MHz three-dimensional ground-penetrating radar was used. Detection parameters: time sampling interval 0.1ns, number of sampling points per channel 512; scan line spacing 2cm, single segment detection length 6.08m; detection vehicle speed 6km / h, to carry out full coverage detection of the pumping section and the surrounding 500m range.

[0122] Implementation process:

[0123] 1. Collect three-dimensional ground-penetrating radar data of the pumping section, quickly complete data decoupling and preprocessing, and focus on preserving the phase reversal characteristics of the reflected waves.

[0124] 2. A crack localization network is used to quickly generate ROI geometric priors. The three-dimensional field strength data volume is input into the pre-trained GPRTrans-InvNet network (which already contains water-saturated crack samples).

[0125] 3. Output the three-dimensional distribution of the dielectric constant of the crack, extract the voxels of the water-filled crack by threshold segmentation (cracks with a dielectric constant > 70 are judged to be water-saturated), and calculate the water volume and distribution range.

[0126] 4. Generate a plan view and depth profile of the water-saturated cracks, and mark high-risk water accumulation areas.

[0127] Inversion Results and Analysis: A total of 23 water-saturated cracks were detected, including 7 high-risk cracks with a water volume greater than 0.1 m³, and a maximum water depth of 32 cm. The total time from data acquisition to generating a complete detection report was approximately 3.5 hours, meeting the time requirements for emergency detection. On-site excavation verification showed that the prediction accuracy of the location and extent of water-saturated cracks reached 100%, and the prediction error of water volume was less than 10%.

[0128] Engineering application value: It can quickly and accurately locate water accumulation cracks after rain, guide maintenance units to take timely emergency measures such as drilling for drainage and grouting for sealing, avoid road subsidence accidents in 3 high-risk sections, and ensure the traffic safety of urban expressways.

[0129] This invention, by incorporating coordinate convolution, anisotropic directional convolution, and antenna axis attention mechanisms within a CAE-Block, enables the network to simultaneously utilize spatial location information, directional variation information, and multi-antenna channel difference information for feature extraction. Coordinate convolution establishes the correlation between reflection features and position coordinates; anisotropic directional convolution characterizes response patterns along the time direction, scan line direction, and antenna direction, respectively; and the antenna axis attention mechanism adaptively weights features from different antenna channels, thus enhancing the representation of crack-related reflection features. A lightweight Transformer bridging module is used to perform global context modeling of the deep features output by the encoder. Since this module can establish cross-positional information interaction relationships between the flattened token sequences, it can further supplement long-range dependency information across time windows and scan lines based on the local convolution extraction results, thereby improving the representation of the overall distribution relationship and continuity features of cracks. By introducing a gating and filtering mechanism for skip connection features during the decoding stage, the high-resolution features output by the encoder undergo target relevance filtering before being fused with the decoding end. Based on this screening process, the impact of background noise and irrelevant layered reflections directly entering the decoding and fusion process can be reduced, thus helping to balance high-resolution detail recovery and non-target response suppression. This invention employs a composite loss function composed of electromagnetic interpretability loss, geometric similarity loss, and numerical consistency loss. Specifically, electromagnetic interpretability loss constrains the consistency between the prediction results and electromagnetic propagation laws; geometric similarity loss constrains the consistency between the prediction results and the prior geometric range of the crack; and numerical consistency loss constrains the consistency between the predicted and target values ​​of the dielectric constant within the crack region. Therefore, a joint constraint can be formed on the inversion process from three aspects: physical laws, spatial structure, and local numerical values.

[0130] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional inversion method for internal cracks in asphalt pavement, characterized in that, Includes the following steps: S1. Obtain three-dimensional ground-penetrating radar detection data of the internal crack area of ​​asphalt pavement, and decouple the three-dimensional ground-penetrating radar detection data into a multi-channel two-dimensional field strength matrix; S2. Based on the crack region location results, obtain the approximate distribution range of the cracks, and construct the approximate distribution range of the cracks as geometric prior information; S3. The multi-channel two-dimensional field strength matrix is ​​preprocessed and then reconstructed according to the antenna channel dimension to obtain the three-dimensional field strength data volume. S4. Construct a three-dimensional dielectric constant inversion network GPR Trans-InvNet. The network includes an encoder, a bridging module, and a decoder. The encoder is composed of multiple CAE-Blocks. The bridging module is a lightweight Transformer bridging module. The decoder includes multiple upsampling units, an Attention Gate gating unit, and a CAE-Block fusion unit. S5. The three-dimensional dielectric constant inversion network is trained using a composite loss function, which includes at least electromagnetic interpretability loss, geometric similarity loss, and numerical consistency loss. S6. Output the three-dimensional distribution of the equivalent dielectric constant of the crack, which corresponds one-to-one with the spatial location of the three-dimensional field strength data volume.

2. The three-dimensional inversion method for internal cracks in asphalt pavement according to claim 1, characterized in that, The three-dimensional ground-penetrating radar detection data mentioned in step S1 was acquired by 24 antenna channels, and the center frequency of the ground-penetrating radar was 800 MHz; In step S2, the approximate distribution range of the cracks is determined by the coordinates of the four corner points output by the existing crack region localization network, forming the enclosing region ROI in the two-dimensional field strength matrix; The three-dimensional field strength data volume in step S3 consists of a time sampling dimension, a scan line dimension along the detection direction, and an antenna channel dimension; the size of the three-dimensional field strength data volume obtained in step S3 is 512×304×24; the preprocessing in step S3 includes at least direct wave removal processing and normalization processing.

3. The three-dimensional inversion method for internal cracks in asphalt pavement according to claim 1, characterized in that, The CAE-Block in step S4 includes at least: a coordinate convolution unit for constructing a normalized coordinate field and concatenating it with the input field strength features; an anisotropic direction convolution unit for extracting orientation-sensitive features along the time axis, scan line axis, and antenna channel axis, respectively; an antenna axis attention unit for adaptive recalibration of the responses of different antenna channels; and a residual fusion and multi-stage feature concatenation unit for fusing the current layer features with the residual branch features and outputting the current layer's encoded or decoded features. The lightweight Transformer bridging module in step S4 flattens the 3D feature volume output by the encoder into a token sequence, performs global context modeling through at least two Transformer Encoder Layers, and then maps it back to the 3D spatial domain. The lightweight Transformer bridging module does not have additional explicit position encoding; it achieves position awareness by relying on the 3D coordinate information injected by the encoder. In step S4, the decoder performs spatial gating filtering on the high-resolution features output by the corresponding encoder layer through the Attention Gate at each decoding stage to obtain gated skip connection features. The gated skip connection features are then concatenated with the upsampled decoding features in the channel dimension and input into the CAE-Block fusion unit for feature fusion and detail reconstruction.

4. The three-dimensional inversion method for internal cracks in asphalt pavement according to claim 1, characterized in that, In step S5, the electromagnetic interpretability loss is calculated by constructing the damped wave equation residuals between the observed electric field and the predicted dielectric constant for each antenna channel's corresponding two-dimensional spatiotemporal profile, and then using Huber loss to constrain the internal region after removing the boundaries. The calculation formula is as follows: The damped wave equation is defined as: ； In the formula, E is the scalar electric field component in the two-dimensional spatiotemporal domain; t is the electromagnetic wave propagation time; x, y are the two-dimensional spatial coordinates; ε(t,x,y) is the spatiotemporally dependent dielectric constant; ∇ 2 x,y are two-dimensional Laplace operators describing the spatial diffusion and propagation characteristics of the electric field; σ is the conductivity of the medium; s(t,x,y) are reflection source and boundary-related terms; The residual of the damped wave equation is defined as: ； In the formula, r(t,x,y) is the residual of the damped wave equation; E(t,x,y) is the electric field intensity component in the spatiotemporal domain. This represents the predicted dielectric constant obtained from network inversion; The partial derivatives in the above residuals are all calculated on the discrete grid using second-order central difference, and the calculation formula is as follows: ； ； ； ；； In the formula, ∆t is the time step, and ∆x and ∆y are the spatial step sizes in the x and y directions, respectively; Within the spatiotemporal computational domain after boundary removal, the Huber function is used to constrain the discrete-form damped wave equation residuals, thereby constructing an electromagnetic interpretability loss function: ； In the formula, L FDTD This is the physical constraint loss function constructed based on the FDTD discrete scheme; ∑ is the total number of sampling points in the spatiotemporal computational domain after removing boundaries; (t,x,y)∈Ω is the summation operator for all spatiotemporal points within the computational domain; 𝜌(⋅) is the Huber loss function, used to reduce the impact of abnormal residuals on the training process.

5. The three-dimensional inversion method for internal cracks in asphalt pavement according to claim 1, characterized in that, The geometric similarity loss in step S5 is calculated by mapping the predicted dielectric constant to a differentiable soft mask and combining it with the geometric prior mask corresponding to the approximate distribution range of the crack to construct an external penalty term and an internal coverage term for the ROI. The calculation formula is as follows: For each channel index y, the geometric truth mask for the crack extent is defined as: ； In the formula, M roi (t,x,y) represents the geometric ground truth mask of the crack region in the GPR image; (t,x) represents the temporal-spatial coordinates of the GPR profile; M roi (t,x,y)=1 indicates that the coordinate point is located within the calibrated crack area; M roi (t,x,y)=0 indicates that the coordinate point is located outside the crack region; A differentiable soft mask is generated using Cauchy mapping, and the calculation formula for the differentiable soft mask is as follows: ； In the formula, P(t,x,y) is a differentiable soft mask obtained by mapping the predicted dielectric constant; ε c τ is the reference center value of the crack dielectric constant, used to construct the geometric probability mapping; τ is the adjustment parameter that controls the weight decay rate. Define an ROI-out penalty loss to suppress the expansion of the predicted crack region beyond the actual crack extent: ； In the formula, L out ε is the external penalty loss for ROI; ε is a smoothing term to prevent the denominator from being zero. Define an in-ROI coverage loss to prevent the network from circumventing penalties by lowering the overall prediction probability: ； In the formula, L in This refers to the coverage loss within the ROI; Ultimately, the geometric similarity loss consists of the two losses mentioned above: ； In the formula, L shape For geometric similarity loss; λ out For and λ in Let λ be the weighting coefficient for the two ROIs. out ≥λ in Prioritize suppressing spurious responses outside the ROI.

6. The three-dimensional inversion method for internal cracks in asphalt pavement according to claim 1, characterized in that, The numerical consistency loss in step S5 is the regionalized L1 loss calculated only within the area defined by the crack voxel mask, and the calculation formula is as follows: Crack Voxel Mask crack The set of voxels used to mark the cracks in the dielectric truth is defined as: ； In the formula, M crack (t,x,y) represents the crack voxel mask; ε gt (t,x,y) represents the target dielectric constant of the crack region; ε c δ is the reference center value for the crack dielectric constant; δ is the tolerance threshold used to distinguish crack voxels. This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise. The numerical consistency loss is defined as: ； In the formula, L num The numerical consistency loss for the crack region applies only to crack voxels and does not impose constraints on background voxels. The background region is constrained by both electromagnetic consistency loss and geometric consistency loss. ε represents the predicted dielectric constant obtained from network inversion; gt (t,x,y) represents the target dielectric constant of the crack region; the numerator is the sum of the absolute errors between the predicted dielectric constant and the target dielectric constant within the crack region, and the denominator is the total number of voxels in the crack region, used to normalize the loss value and eliminate the influence of crack size on the loss calculation.

7. The three-dimensional inversion method for internal cracks in asphalt pavement according to claim 1, characterized in that, In step S6, a 1×1×1 three-dimensional convolution operator is first used to map the multi-channel features into a single-channel initial predicted dielectric constant volume; then, the initial predicted dielectric constant volume is corrected in three dimensions by trilinear interpolation, and finally the three-dimensional distribution of the equivalent dielectric constant of the crack is output, which corresponds one-to-one with the three-dimensional field strength data volume in spatial location.

8. A three-dimensional ground-penetrating radar inversion system for internal cracks in asphalt pavement, used to implement the three-dimensional ground-penetrating radar inversion method for internal cracks in asphalt pavement as described in any one of claims 1-7, characterized in that, include: The data acquisition and decoupling module is used to acquire three-dimensional ground-penetrating radar detection data of the crack area inside the asphalt pavement, and decouple the detection data into a multi-channel two-dimensional field strength matrix. The geometric prior generation module is used to obtain the approximate distribution range of cracks based on the crack region location results, and to construct the approximate distribution range of cracks as geometric prior information. The data preprocessing and reconstruction module is used to perform direct wave removal and normalization processing on the multi-channel two-dimensional field strength matrix, and reconstruct it into a three-dimensional field strength data volume according to the antenna channel dimension. Network coding module: used to extract multi-scale features from the three-dimensional field strength data volume using an encoder composed of multi-level CAE-Blocks; Network bridging module: Used to perform global context modeling of encoder output features using a lightweight Transformer bridging module; Network Decoding and Output Correction Module: Used to decode the bridging features step by step using a decoder including an Attention Gate gating unit, and output the three-dimensional distribution of the equivalent dielectric constant of the crack corresponding to the spatial position of the three-dimensional field strength data volume through 1×1×1 three-dimensional convolution and trilinear interpolation. Composite loss training module: used to train the three-dimensional dielectric constant inversion network using a composite loss function that includes at least electromagnetic interpretability loss, geometric similarity loss and numerical consistency loss.

9. An electronic device employing the three-dimensional ground-penetrating radar inversion system for internal cracks in asphalt pavement as described in claim 8, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the three-dimensional inversion method for internal cracks in asphalt pavement as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing thereon a computer program employing the three-dimensional ground-penetrating radar inversion system for internal cracks in asphalt pavement as described in claim 8, wherein the computer program, when executed by a processor, implements the three-dimensional inversion method for internal cracks in asphalt pavement as described in any one of claims 1 to 7.

Images