Road dielectric constant inversion and disease discrimination method based on neural network

Abstract

The invention relates to a road dielectric constant inversion and disease discrimination method based on a neural network. The method comprises the following steps: expanding the number of jump connections based on a TransUNet baseline, introducing a gating attention mechanism, and additionally arranging a multi-scale feature fusion module, an asymmetric depth-space attention module and a deep supervision solution terminal; an AdamW optimizer and a Warmup-cosine annealing learning rate scheduling strategy are adopted to construct an adaptive region weighted loss function so as to amplify dielectric constant mutation boundary loss weight. After a model is trained through a road GPR simulation data set and parameter adjustment of the set is verified to be convergent, a to-be-detected road GPR-B-scan image is input, dielectric constant distribution is output, and according to dielectric constant characteristics of different diseases, accurate discrimination of hidden diseases such as cavities, voids, loose bodies and water-rich bodies is achieved. By adopting the method, high-resolution dielectric constant end-to-end inversion can be realized, and multi-scale disease characteristics can be accurately captured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection, in particular to a road dielectric constant inversion and disease discrimination method based on a neural network. BACKGROUND

[0002] As a national key infrastructure, the structural safety and service life of the road are directly related to the traffic efficiency and public safety. In the long-term service process, the internal structure of the road is prone to hidden diseases such as cavities, voids, loose bodies and water-rich bodies due to repeated load action, environmental erosion and other factors. Such diseases are difficult to be found by visual inspection in the early stage, but they can significantly reduce the bearing capacity and durability of the road. If not treated in time, they can easily cause serious problems such as road collapse and cracking, so there is an urgent need for efficient and accurate detection technology for internal diseases of the road.

[0003] Ground penetrating radar (GPR) has become the mainstream technology for detecting the internal structure of the road due to its advantages of non-destructive, high resolution and fast imaging. Its principle is to generate a B-scan spectrum reflecting the internal structure of the road by transmitting high-frequency electromagnetic waves and receiving the reflected signals of the underground medium. However, the original GPR spectrum is only the strength profile of the electromagnetic wave reflection signal, which cannot directly represent the physical properties of the material. It needs to be inverted into the relative dielectric constant, which is the most sensitive parameter to indicate the health of the road. The dielectric constant of the road material is usually 3-10, the dielectric constant of water is as high as 81, and the dielectric constant of air is close to 1. The distribution of dielectric constant can directly reflect the distribution of internal diseases of the road.

[0004] Current dielectric constant inversion methods can be mainly divided into two categories: one is the physics-driven method, which realizes inversion by iteratively matching the observed and simulated waveforms. However, this method has the defects of high computational complexity, sensitivity to initial model and susceptibility to noise interference. Especially, three-dimensional FWI can improve the spatial coverage, but it is too time-consuming to meet the efficient detection needs of engineering. The second is the data-driven method. Traditional machine learning relies on manual feature design, and its prediction ability for complex underground structure dielectric constant is limited, making it difficult to achieve fine inversion. Although existing deep learning methods can extract deep features, CNN lacks the ability to model the global context, which can easily lead to "one picture with multiple interpretations" of GPR spectrum. Traditional skip connection can easily introduce redundant noise, and the multi-scale disease feature adaptation is insufficient, the boundary detail prediction is fuzzy, which cannot meet the engineering needs of accurate discrimination of road hidden diseases.

[0005] In summary, the existing GPR dielectric constant inversion technology has the bottlenecks of low computational efficiency, limited inversion accuracy, high manual intervention and poor adaptability to complex diseases, which makes it difficult to efficiently support road maintenance decisions. SUMMARY

[0006] Therefore, it is necessary to provide a neural network-based method for road dielectric constant inversion and disease identification that can achieve high-resolution end-to-end dielectric constant inversion and accurately capture multi-scale disease characteristics, in order to address the above-mentioned technical problems.

[0007] A method for road dielectric constant inversion and defect identification based on neural networks, the method comprising: Construct a GPR simulation dataset for roads; expand the number of skip connections in the TransUNet baseline model, introduce a gated attention mechanism at each skip connection, and design a multi-scale feature fusion module, an asymmetric depth-space attention module, and a deep supervised decoder head to form a GPR-TransUNet neural network model; AdamW is used as the optimizer, and the learning rate scheduling method is set in combination with the Warmup-cosine annealing strategy. An adaptive region-weighted loss function is constructed. The adaptive region-weighted loss function amplifies the loss weight of the dielectric constant change boundary through a dynamic weighting mechanism. The training set from the GPR simulation dataset is input into the GPR-TransUNet neural network model. The configured training strategy and adaptive region weighted loss function are used to train the model. The loss change and evaluation index fluctuation of the model are monitored through the validation set during the training process. The model parameters are adjusted until the model converges, and the trained GPR-TransUNet neural network model is obtained. The GPR-B-scan image of the road to be detected is input into the trained GPR-TransUNet neural network model, which outputs the dielectric constant distribution of the road to be detected. Based on the dielectric constant characteristics corresponding to different diseases in the dielectric constant distribution, the model identifies cavities, voids, loose bodies and water-rich bodies in the road to be detected, thus completing the identification of hidden road diseases.

[0008] The above-mentioned method for road dielectric constant inversion and defect identification based on neural networks first constructs a simulation dataset through the gprMax platform, which not only covers the real structural characteristics of roads, but also provides rich defect feature samples for the model, avoiding insufficient generalization ability due to single data, and laying a data foundation for subsequent accurate inversion. The number of skip connections in the TransUNet baseline model is increased, and a gated attention mechanism is introduced at each skip connection. A multi-scale feature fusion module, an asymmetric depth-space attention module, and a deep-supervised decoder are designed to construct the GPR-TransUNet neural network model. The asymmetric depth-space attention module decouples the physical features of the GPR map's vertical depth axis and horizontal scan axis, accurately capturing the correlation between layered structures and local anomalies. The multi-scale feature fusion module extracts features in parallel through multiple receptive fields, adapting to different scales of defects such as small cavities and large-scale voids; ablation experiments confirm its greatest contribution to accuracy improvement. The increased skip connections and the introduction of a gated attention mechanism filter effective information and redundant noise from the encoder's low-level detail features, addressing the interference introduced by traditional skip connections. The deep-supervised decoder guides the model to learn dielectric constant mapping rules at various scales through multi-layered supervision constraints, avoiding the loss of boundary details during decoding. These four components work together to enhance the model's ability to model global semantics and local details, providing structural support for high-resolution inversion. The adaptive region-weighted loss function is designed to dynamically amplify the loss weights at dielectric constant abrupt change boundaries, addressing the problem of traditional loss functions' insufficient attention to disease boundaries and improving edge characterization accuracy. The AdamW optimizer, combined with a Warmup-cosine annealing strategy, avoids local optima caused by inappropriate learning rates in the early stages of training and ensures convergence stability through smooth deceleration in later stages, guaranteeing the model's full learning of data features. Ultimately, the trained model can directly map the GPR-B-scan image to be detected as a dielectric constant distribution, achieving accurate discrimination by combining the dielectric characteristics of different diseases. Simulation verification shows that the inversion results are highly consistent with borehole endoscopy detection, confirming the effectiveness of this application in end-to-end inversion and multi-scale disease capture. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for road dielectric constant inversion and defect identification based on neural networks in one embodiment. Figure 2 This is a schematic diagram of the GPR-TransUNet neural network model in one embodiment; Figure 3 This is a schematic diagram of the network structure of the multi-scale feature fusion (MSF) module in one embodiment; Figure 4 This is a flowchart illustrating the gated attention mechanism (GSF) in another embodiment; Figure 5This is a flowchart illustrating the deep supervision mechanism (DS-Head) in one embodiment; Figure 6 This is a schematic diagram of the asymmetric depth-space attention mechanism (ASDS) structure in one embodiment; Figure 7 This is a schematic diagram of the changes in training loss and evaluation metrics of GPR-TransUNet in one embodiment; Figure 7 (a) is a schematic diagram of the loss change curve. Figure 7 (b) is a schematic diagram of the changes in the indicators; Figure 8 This is a schematic diagram of the dielectric inversion results of the mainstream model in one embodiment; Figure 9 This is a schematic diagram of the inversion results of GPR-TransUNet for different types of diseases in one embodiment; Figure 10 This is a schematic diagram of ground-penetrating radar measured spectrum inversion and endoscopic examination in one embodiment; Figure 10 (a) is a schematic diagram of road segment 1; Figure 10 (b) is a schematic diagram of the dielectric constant inversion results for section 1; Figure 10 (c) is a schematic diagram of endoscopic examination of section 1; Figure 10 (d) is a schematic diagram of road segment 2; Figure 10 (e) is a schematic diagram of the dielectric constant inversion results for section 2; Figure 10 (f) is a schematic diagram of endoscopic examination of section 2; Figure 10 (g) is a schematic diagram of road segment 3; Figure 10 (h) is a schematic diagram of the dielectric constant inversion results for section 3; Figure 10 (i) is a schematic diagram of endoscopic examination of section 3. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0011] In one embodiment, such as Figure 1 As shown, a method for road dielectric constant inversion and defect identification based on neural networks is provided, including the following steps: Step 102: Construct a road GPR simulation dataset; expand the number of skip connections in the TransUNet baseline model, introduce a gated attention mechanism at each skip connection, and design a multi-scale feature fusion module, an asymmetric depth-space attention module, and a deep supervised decoder head to form a GPR-TransUNet neural network model.

[0012] The road GPR simulation dataset is the foundation for model training. The dataset built through the gprMax 3D forward modeling platform can simulate the homogeneous and heterogeneous structures of real roads, and includes four typical defects: cavities, voids, loose bodies, and water-rich bodies, ensuring the diversity and authenticity of the data.

[0013] TransUNet employs three skip connections for feature fusion. This application extends this to four skip connections and introduces a gated attention mechanism at each skip connection, enabling the model to adaptively and selectively enhance key features. This focuses on more discriminative regions, improving the accuracy of detailed predictions.

[0014] Since the features output by the Transformer differ from those of the convolutional features in terms of spatial layout, this application designs an MSF module to reconstruct the global features of the Transformer into a convolution-friendly tensor representation, and supplements it with multi-scale local information before inputting it into the decoder.

[0015] like Figure 2 The diagram shows the structure of the GPR-TransUNet neural network model. While retaining the global modeling advantages of TransUNet, GPR-TransUNet improves the effectiveness of training signals through a hybrid loss function and deep supervision mechanism, optimizes the convergence of the training process through Warmup and cosine annealing, and enhances feature representation capabilities through attention enhancement and multi-scale fusion. The improved network can more accurately recover the underground dielectric constant distribution. The multi-scale feature fusion (MSF) module network structure is shown below. Figure 3 As shown.

[0016] The road GPR simulation dataset is the foundation for model training. The dataset built through the gprMax 3D forward modeling platform can simulate the homogeneous and heterogeneous structures of real roads, and includes four typical defects: cavities, voids, loose bodies, and water-rich bodies, ensuring the diversity and authenticity of the data.

[0017] The TransUNet baseline model combines the local feature extraction capabilities of CNNs with the global context modeling advantages of Transformers. Expanding the number of skip connections enhances the integrity of feature transmission. The addition of gated attention mechanism, multi-scale feature fusion module, asymmetric depth-space attention module and deep supervised decoding head solves problems such as noise interference, poor multi-scale disease adaptation, insufficient feature extraction targeting, and loss of boundary details in traditional models, and synergistically improves the inversion accuracy of the model.

[0018] Step 104: AdamW is used as the optimizer, and the learning rate scheduling method is set in combination with the Warmup-cosine annealing strategy to construct an adaptive region-weighted loss function. The adaptive region-weighted loss function amplifies the loss weight of the dielectric constant change boundary through a dynamic weighting mechanism.

[0019] The AdamW optimizer effectively reduces the risk of overfitting and improves the stability of parameter updates through weight decay. The Warmup-Cosine Annealing strategy linearly increases the learning rate in the early stages of training to prevent the model from converging to local optima prematurely, and then gradually decreases the learning rate using a cosine function to achieve smooth convergence. The adaptive region-weighted loss function addresses the characteristic of disease boundary pixels, which have a small percentage but significant physical meaning due to abrupt changes in dielectric constant. It dynamically amplifies the loss weights of boundary regions, solving the problem of blurred inversion results caused by insufficient attention to boundaries in traditional loss functions, and improving edge delineation accuracy.

[0020] Step 106: Input the training set from the GPR simulation dataset into the GPR-TransUNet neural network model, use the configured training strategy and adaptive region weighted loss function to train the model, monitor the loss changes and evaluation index fluctuations of the model during the training process through the validation set, adjust the model parameters until the model converges, and obtain the trained GPR-TransUNet neural network model.

[0021] The training set provides the model with sufficient learning samples, enabling it to grasp the mapping relationship between GPR-B-scan images and dielectric constant distribution. The validation set is used to monitor the model's training performance in real time. By observing the loss change curve and fluctuations in evaluation metrics (MAE, RMSE, R²), it is determined whether the model is overfitting or underfitting, and then the model parameters (such as learning rate, number of iterations, etc.) are adjusted to ensure that the model has good generalization ability and inversion accuracy after convergence.

[0022] Step 108: Input the GPR-B-scan image of the road to be detected into the trained GPR-TransUNet neural network model, and output the dielectric constant distribution of the road to be detected; based on the dielectric constant characteristics corresponding to different diseases in the dielectric constant distribution, identify cavities, voids, loose bodies and water-rich bodies in the road to be detected, and complete the identification of hidden road diseases.

[0023] The trained model can achieve end-to-end mapping from the original GPR signal to the dielectric constant distribution without manual intervention, thus improving detection efficiency. The dielectric constant characteristics corresponding to different defects are significantly different. Air (dielectric constant ≈ 1) corresponds to cavitation and delamination defects, water (dielectric constant ≈ 81) corresponds to water-rich defects, and road materials (dielectric constant 3-10) with uneven distribution correspond to loose material defects. Based on this characteristic, the type and distribution range of defects can be accurately identified.

[0024] The above-mentioned method for road dielectric constant inversion and defect identification based on neural networks first constructs a simulation dataset through the gprMax platform, which not only covers the real structural characteristics of roads, but also provides rich defect feature samples for the model, avoiding insufficient generalization ability due to single data, and laying a data foundation for subsequent accurate inversion. The number of skip connections in the TransUNet baseline model is increased, and a gated attention mechanism is introduced at each skip connection. A multi-scale feature fusion module, an asymmetric depth-space attention module, and a deep-supervised decoder are designed to construct the GPR-TransUNet neural network model. The asymmetric depth-space attention module decouples the physical features of the GPR map's vertical depth axis and horizontal scan axis, accurately capturing the correlation between layered structures and local anomalies. The multi-scale feature fusion module extracts features in parallel through multiple receptive fields, adapting to different scales of defects such as small cavities and large-scale voids; ablation experiments confirm its greatest contribution to accuracy improvement. The increased skip connections and the introduction of a gated attention mechanism filter effective information and redundant noise from the encoder's low-level detail features, addressing the interference introduced by traditional skip connections. The deep-supervised decoder guides the model to learn dielectric constant mapping rules at various scales through multi-layered supervision constraints, avoiding the loss of boundary details during decoding. These four components work together to enhance the model's ability to model global semantics and local details, providing structural support for high-resolution inversion. The adaptive region-weighted loss function is designed to dynamically amplify the loss weights at dielectric constant abrupt change boundaries, addressing the problem of traditional loss functions' insufficient attention to disease boundaries and improving edge characterization accuracy. The AdamW optimizer, combined with a Warmup-cosine annealing strategy, avoids local optima caused by inappropriate learning rates in the early stages of training and ensures convergence stability through smooth deceleration in later stages, guaranteeing the model's full learning of data features. Ultimately, the trained model can directly map the GPR-B-scan image to be detected as a dielectric constant distribution, achieving accurate discrimination by combining the dielectric characteristics of different diseases. Simulation verification shows that the inversion results are highly consistent with borehole endoscopy detection, confirming the effectiveness of this application in end-to-end inversion and multi-scale disease capture.

[0025] In one embodiment, the process of constructing a road GPR simulation dataset includes: Based on the gprMax 3D forward modeling platform, two types of road models, homogeneous and heterogeneous, were built. Four typical road defects, namely cavities, voids, loose bodies, and water-rich bodies, were randomly embedded in the two models to generate multiple B-scan images of the road GPR. The B-scan images were divided into training, validation, and test sets in an 8:1:1 ratio. At the same time, the true dielectric constant distribution corresponding to the B-scan images was extracted from the gprMax forward modeling results and mapped to uint16 grayscale images with the same size as the B-scan images as label data, thus completing the dataset construction.

[0026] Specifically, forward modeling of the road structure was performed using GprMax software. The simulation parameters of the road model are shown in Table 1, and the settings of the GprMax software are shown in Table 2. Ground-penetrating radar image datasets of various typical types of internal road defects were constructed.

[0027] The geometric and material parameters of the road structure are simplified based on actual pavement engineering construction. The model consists of a 20cm asphalt surface layer and an 80cm cement-stabilized crushed stone base layer, with a total thickness of 1.0m. A PLM layer is set around the model to absorb electromagnetic waves propagating to the model boundary and prevent them from being reflected back to the calculation area. To enhance data diversity, two basic models were designed: a homogeneous material model, in which the materials inside the asphalt layer and the water-stabilized layer are uniform; and a heterogeneous material model, in which aggregates of different sizes are distributed inside the asphalt layer and the water-stabilized layer according to gradation requirements, to more realistically simulate the heterogeneous structure inside the road. Based on this, four basic defects—voids, voids, loose bodies, and water-rich bodies—are randomly embedded in both the homogeneous and heterogeneous models. Their location, scale, and combination are all randomly set, thereby enriching the diversity of defect manifestations.

[0028] Table 1

[0029] Table 2

[0030] A total of 2000 Bscan images were generated, including 1000 homogeneous images and 1000 heterogeneous images. Following the conventional partitioning principles for deep learning tasks, the data was randomly divided into training, validation, and test sets in a ratio of 8:1:1, with 1600 images in the training set, 200 images in the validation set, and 200 images in the test set.

[0031] The label data in this study comes directly from the dielectric constant distribution generated by gprMax forward modeling. For each B-scan image, its corresponding true dielectric constant field is mapped to a uint16 grayscale image with the same size as the input, where the grayscale value of each pixel represents the dielectric constant at that location. During model training, the uint16 is converted to float32 without changing the pixel values ​​themselves. This pixel real-value matrix is ​​used as the supervision signal, and the network output is a continuous dielectric constant distribution of the same size, achieving end-to-end pixel-level regression.

[0032] To facilitate result presentation and intuitive comparison, color mapping was used in the experiment to convert the predicted dielectric constant values ​​into pseudo-color images, making different value ranges visually represent gradient color bands. This strategy not only preserves the continuous characteristics of the true dielectric constant but also improves the image's recognizability, helping to evaluate the model's prediction accuracy at interlayer transitions, defect boundaries, and local details.

[0033] In one embodiment, the implementation process of the gating attention mechanism includes: The high-level semantic features X output by the decoder and the low-level detail features S output by the encoder in the GPR-TransUNet neural network model are concatenated along the channel dimension to obtain the joint feature Z. ; The joint feature Z is compressed by the first 1×1 convolutional layer, followed by batch normalization and ReLU activation function operation. Then, it is operated by the second 1×1 convolutional layer and Sigmoid activation function to generate gating weights with the same number of channels as the low-level detail features. The gating weights are multiplied element-wise with the low-level detail features to obtain the weighted filtering features, thereby enhancing the effective features and suppressing the noisy features in the skip connections. The gating weights that generate the same number of low-level detail feature channels are: ; in, Indicates the gating weight, These are the weight parameters for the first 1×1 convolutional layer. These are the weight parameters for the second 1×1 convolutional layer. For the Sigmoid function, For ReLU function, This is for batch normalization operations.

[0034] Specifically, a lightweight gating unit is embedded in the skip path, and the skip features are adaptively recalibrated using a learnable channel-level weight vector to achieve "useful information enhancement and noise component suppression," thereby improving the sensitivity and robustness of dielectric constant inversion to key scattering details. Its structural diagram is shown below. Figure 5 As shown.

[0035] The high-level semantic features X output by the decoder contain global contextual information, while the low-level detail features S output by the encoder contain local texture and edge information. The joint feature Z formed by concatenating these two features along the channel dimension combines the advantages of both types of features. The first 1×1 convolutional layer is used to compress the number of channels in the joint feature Z, reducing computation and fusing feature information. Batch normalization accelerates model training convergence, and the ReLU activation function enhances the non-linear expressive power of the features. The second 1×1 convolutional layer maps the features to a vector with the same number of channels as the low-level detail features S. After passing through the Sigmoid activation function, a gate weight G is obtained. The closer the G value is to 1, the more important the feature at the corresponding position; the closer it is to 0, the more redundant the feature at that position is noise. Multiplying the gate weight G element-wise with the low-level detail features S preserves and enhances useful detail features, filters redundant noise, and solves the noise introduction problem caused by the semantic gap in traditional skip connections. The beneficial effect is that this mechanism achieves adaptive selection of skip connection features, improves the effectiveness of feature propagation, makes the model more focused on key features related to the disease, and thus improves the accuracy of dielectric constant inversion and the clarity of boundary details.

[0036] In one embodiment, the implementation of the asymmetric depth-space attention module includes: The module input feature map is linearly transformed along the channel dimension, and a Value feature map with the same size as the input feature map is generated through a 1×1 convolutional layer. We construct a deep branch and a spatial branch to capture features of different dimensions. The deep branch uses a k×1 size convolutional kernel to extract the correlation features of the vertical axis of the GPR map, thus obtaining a deep attention map. The spatial branch uses a 1×k size convolution kernel to extract the correlation features of the horizontal axis of the GPR map, thus obtaining the spatial attention map; The deep attention map and the spatial attention map are added together, and then compressed to the range [0,1] using the Sigmoid activation function to generate the final attention gating; The attention gating is multiplied element-wise with the Value feature map to obtain the recalibrated output feature map, which enables precise focusing on vertical layered features and lateral anomaly features in the GPR map.

[0037] Specifically, Asymmetric Depth-Spatial Attention (ASDS) is a customized attention module proposed in this invention for the physical characteristics of GPR (Ground Penetrating Radar) B-scan maps. Traditional convolutional neural networks typically use symmetric convolution kernels when processing images. This design assumes that the features of the image are isotropic in all spatial directions. However, GPRB-scan is physically anisotropic, and the physical meanings of its two dimensions are completely different. The vertical axis represents the propagation time (depth) of electromagnetic waves. Features on this axis mainly reflect layered structures (such as road surfaces, base layers, and voids) and signal attenuation. These features typically have strong continuity in the horizontal direction. The horizontal axis represents the radar's scan trajectory. Features on this axis are mainly used to identify local anomalies.

[0038] Using symmetric convolution kernels can confuse these two completely different physical features. It forces the network to use the same tools to process both lateral and longitudinal signals, which violates the physical imaging principle of GPR, resulting in low feature extraction efficiency and weak ability to distinguish between different types of diseases.

[0039] The ADSA module solves the above problem by decoupling the two orthogonal physical axes of the GPR graph, specifically by the following steps: Step 1: Linear Transformation The linear transformation Convvalue unit is used to transform input features. Figure X A linear transformation along the channel dimension is performed to generate a feature map V (called the Value feature map) with the same size as the input. The convolution operation of step 1 performs a linear recombination and refinement of the contents of the input x, generating a feature map V to be processed: ; Step 2: Asymmetric Attention Mechanism The depth branch is used to capture the correlation of the vertical axis (depth / time axis) in the GPR map. It uses asymmetric (k... 1) A convolutional kernel whose receptive field can only move and sense in the vertical direction, and is insensitive to changes in the horizontal direction. This is suitable for extracting electromagnetic wave reflection features of vertical defects in the road surface and subgrade from GPR maps. The feature extraction process of deep branches is as follows: ; Spatial branch, this branch is used to capture the correlation of the horizontal axis (spatial / survey axis) in the GPR map. Using asymmetric (1... A convolutional kernel of type k) has a receptive field that can only move and sense horizontally, and is insensitive to changes in the vertical direction. It is suitable for identifying reflection patterns with a certain horizontal size caused by local anomalies. The feature extraction process for spatial branching is as follows: ; Step 3: Attention Gated Fusion This step generates the final attention gating M, a fusion method that allows M to simultaneously activate regions with both "salient vertical features" and "salient lateral features." Sigmoid activation is used to compress the fused attention map to the [0,1] range, which transforms it into a gating unit. The final attention gating generation process is as follows: ; Step 4: Feature Recalibration This step is the final application of the attention mechanism. The attention gate M generated in step 3 is multiplied element-wise with the value V generated in step 1. Features in V corresponding to high-value regions of M will be fully preserved, while features in V corresponding to low-value regions of M will be effectively suppressed. The final output feature map Y is: ; This module decouples the vertical and horizontal physical features of the GPR map, avoids the confusion of different dimensional features by the symmetric convolution kernel, improves the targeting and effectiveness of feature extraction, and enables the model to more accurately capture the layered structure of roads and local anomalies, providing support for high-resolution dielectric constant inversion.

[0040] In one embodiment, the construction process of the multi-scale feature fusion module includes: Four parallel feature extraction branches are set up, and each branch uses convolution kernels of size 1×1, 3×3, 5×5 and 7×7 respectively for convolution operation. Padding is set during the convolution process to ensure that the spatial resolution of the output feature map of each branch is consistent with the input feature map. After the convolution operation in each branch, batch normalization and ReLU activation function operation are performed sequentially to enhance the nonlinear representation of features. The output features of the four branches are concatenated along the channel dimension. The concatenated features are then compressed and integrated using a 1×1 convolutional layer to restore the number of output channels to the number of channels in the input feature map. Batch normalization and ReLU activation function operations are then performed again. By introducing a learnable residual scaling factor, the integrated features are weighted and fused with the module input features to obtain a multi-scale feature fusion result.

[0041] Specifically, the multi-scale feature fusion module designed in this application explicitly extracts multi-level features within the receptive field range of 1×1 to 7×7 through a parallel multi-branch structure, and integrates them in an adaptive manner to enhance the model's ability to represent complex underground structures at multiple scales, thereby improving the accuracy and robustness of dielectric constant inversion. Its structural diagram is shown below. Figure 4As shown, this module extracts contextual information from different receptive fields through a parallel multi-branch convolutional structure, thereby achieving effective integration of cross-scale features. The module comprises four parallel branches, each employing standard convolutional operations with different kernel sizes (1×1, 3×3, 5×5, and 7×7), and introducing appropriate pendding while maintaining spatial resolution. Each convolution is followed by batch normalization and ReLU activation. The outputs of the four branches are concatenated along the channel dimension. To further compress the channel dimension and enhance feature interaction, a 1×1 convolutional layer is introduced to integrate the concatenated features, restoring the channel count to C, and then batch normalization and ReLU activation are applied again. Finally, the module introduces a learnable residual scaling factor to weightedly fuse the fused features with the original input. This module, through parallel extraction from multiple receptive fields and adaptive fusion, enables the model to adapt to different scales of road defects, enhancing its multi-scale representation ability of complex underground structures. Ablation experiments confirm that it contributes the most to improving inversion accuracy.

[0042] In another embodiment, in traditional encoder-decoder networks, the supervision signal typically only applies to the output of the last layer. This single-point supervision mechanism easily leads to insufficient gradient propagation in intermediate layers, thus affecting the sufficiency of feature learning. Simultaneously, because deep features gradually lose detailed information during progressive upsampling, the model often exhibits fuzziness and inaccuracy in predicting object boundaries. Therefore, a deep supervision mechanism is introduced to constrain the decoding process at multiple scales. The specific structure is as follows... Figure 6 As shown, the output of each decoding layer is passed through a lightweight segmentation head, and a base loss is calculated for each auxiliary output, with weights of 0.5, 0.3, and 0.2 respectively. These are then added to the main output loss to obtain the final loss. By introducing supervision into output layers at different scales, the model can simultaneously learn local details and global semantics, effectively constraining low-level features containing rich spatial information and improving the network's predictive ability in boundary and fine-grained regions.

[0043] In one embodiment, the calculation process of the adaptive region-weighted loss function includes: Spatial gradient information is extracted from the true dielectric constant distribution of the tag data, and the approximate gradient values ​​in the horizontal and vertical directions of the true dielectric constant distribution are calculated. The gradient magnitude map is synthesized by calculating the square root of the sum of the squares of the approximate gradient values ​​in the horizontal and vertical directions pixel by pixel, and restoring the original size by padding with zeros. The gradient magnitude map is input into an auxiliary detection network consisting of two convolutional layers, a ReLU nonlinear activation layer, and a Sigmoid output layer to generate a dynamic key region weighted mask with the same size as the predicted dielectric constant map output by the GPR-TransUNet neural network model. The fundamental loss is calculated based on the global pixel-level error between the predicted dielectric constant map and the actual dielectric constant distribution. The loss in the critical region is calculated based on the predicted dielectric constant map, the actual dielectric constant distribution, and the dynamic critical region weighted mask. An adaptive region weighted loss function is obtained by weighting and summing the basic loss and the critical region loss using preset coefficients.

[0044] In one embodiment, the underlying loss is calculated based on the global pixel-level error between the predicted dielectric constant map and the true dielectric constant distribution. ; in, For feature map height, The width of the feature map. To predict the pixel value at position (i,j) of the dielectric constant map, The pixel value represents the actual dielectric constant at position (i,j).

[0045] In one embodiment, the critical region loss is calculated based on the predicted dielectric constant map, the actual dielectric constant distribution, and a dynamic critical region weighted mask, including: Based on the predicted dielectric constant map, the actual dielectric constant distribution, and the dynamic critical region weighted mask, the critical region loss is calculated as follows: ; in, For feature map height, The width of the feature map. To predict the dielectric constant plot in ( i, j The pixel value at position ) The true dielectric constant is distributed in ( i, j The pixel value at position ) Weighted mask for dynamic key regions in ( i, j The pixel value at the location.

[0046] Specifically, the core idea of ​​the adaptive region-weighted loss function is to utilize a trainable auxiliary network to automatically learn and amplify the loss weights for dielectric constant abrupt changes. Standard loss functions (MSE or L1) treat all pixels in an image equally. In the GPR dielectric constant inversion task, background pixels constitute the vast majority, while the boundary pixels of lesions, though fewer in number, are physically significant. Traditional loss functions cause the model to ignore the precision of these boundaries in order to reduce the overall average error, resulting in blurred boundaries, loss of detail, and inability to accurately delineate the extent of lesions. The adaptive region-weighted loss function addresses these issues through a dynamic weighting mechanism. It identifies which pixels represent key boundaries of dielectric constant abrupt changes and then applies a higher penalty to the prediction error of these pixels.

[0047] From the actual dielectric constant label image Spatial gradient information is extracted to locate regions where the dielectric constant changes abruptly. This step yields a gradient map. Used only as an auxiliary detection network The input is treated as a constant during training.

[0048] Calculate separately Gradient approximations in the horizontal and vertical directions and .

[0049] ; ; Calculated pixel by pixel and The square root of the sum of squares is used to synthesize the gradient magnitude map by restoring the original size by padding with zeros. And set it to prevent gradient backpropagation.

[0050] ; The generated gradient map without gradients Input to a trainable auxiliary detection network , It consists of a lightweight convolutional neural network with two convolutional layers, a ReLU non-linear activation layer, and a sigmoid output layer. Forward propagation is performed, and the predicted dielectric constant map is automatically generated based on the saliency of the region, compared with the output of the master inversion network. and Dynamic key region weighted mask of the same size . It is a spatial attention map where the value of each pixel is between [0,1], and the value represents the probability or weight of the corresponding position being the boundary of the disease.

[0051] Basic loss calculate and Global pixel-level error between Calculation by mask The critical area error of the disease boundary is adjusted. This calculation is performed by... Simultaneously acting on and To achieve, ; ; This step is equivalent to... and Pixel-wise squared error Apply a squared weight This square-weighted mechanism amplifies Error penalty in areas with higher values, while suppressing Error impact in areas with lower values.

[0052] By preset coefficients and Will and The final total loss is obtained by weighted summation. .

[0053] This loss function addresses the problem of traditional loss functions not paying enough attention to the disease boundary. Through a dynamic weighting mechanism, the model focuses more on key boundary regions during training, improving the edge fidelity of dielectric constant inversion and providing clearer boundary criteria for accurate disease identification.

[0054] In one embodiment, the learning rate is set to... ; in, The learning rate for the current round. To minimize the learning rate, The initial learning rate, For the current training round, This refers to the total number of training rounds.

[0055] In a specific implementation, the optimizer AdamW is used to reduce the risk of overfitting and improve the stability of parameter updates. The learning rate scheduling employs a Warmup + Cosine Annealing strategy. In the early stages of training, linear Warmup is used to gradually increase the learning rate, preventing the model from converging to a local optimum prematurely. Later, cosine annealing is used to gradually decrease the learning rate, resulting in a smoother convergence process. This combined scheduling strategy demonstrates a more stable training curve and better final accuracy in experiments. This learning rate scheduling strategy balances the model's global search capability and local refinement capability, avoiding gradient explosion caused by an excessively high learning rate in the early stages of training or slow convergence caused by an excessively low learning rate in the early stages, ensuring that the model can stably converge to the optimal state and improving inversion accuracy.

[0056] In one embodiment, the GPR-B-scan image of the road to be detected is input into a trained GPR-TransUNet neural network model, which outputs the dielectric constant distribution of the road to be detected, including: After inputting the GPR-B-scan image of the road to be detected into the trained GPR-TransUNet, the asymmetric depth-spatial attention module is first applied to the encoder. The layered features of the vertical depth axis of the GPR image are captured by k×1 convolution, and the local abnormal features of the horizontal scanning axis are extracted by 1×k convolution. Then, the features are fused to generate attention gating, resulting in the enhanced feature map of the encoding stage.

[0057] The enhanced feature map is optimized by a multi-scale feature fusion module. Different receptive field features are extracted in parallel by multi-size convolutional kernels. After the output of each branch is processed by BN-ReLU, a residual scaling factor is introduced after concatenation in the channel dimension to weight and sum the fused features with the original high-dimensional features to obtain the feature map after feature fusion.

[0058] During the feature transfer to the decoder, the gated attention module concatenates the low-level detail features of the encoder with the high-level semantic features of the decoder, compresses the dimension through 1×1 convolution-BN-ReLU, and then generates gated weights through Sigmoid. The gated weights are multiplied pixel by pixel with the original features of the encoder to filter out the detail features useful for dielectric inversion, and the filtered features are passed into the corresponding layer of the decoder. The deep supervised decoding head applies supervisory constraints to the output of each decoding layer, and an auxiliary prediction is generated by a lightweight segmentation head. The decoder output is then subjected to channel compression and linear activation to map the pixel values ​​of the feature map to the actual dielectric constant values, forming a dielectric constant distribution consistent with the size of the input image.

[0059] Specifically, the GPR-B-scan image of the road to be detected first undergoes standardization preprocessing. Each pixel value is normalized according to the mean and standard deviation of the training set images to ensure the input feature distribution is consistent with the training phase. In the encoder phase, the preprocessed image is downsampled through three layers of "convolution-BN-ReLU" convolutional blocks, gradually increasing the number of channels from the initial 3 to 256, capturing local reflection features. The asymmetric depth-spatial attention module enhances the output of each convolutional block, generating an enhanced feature map for the encoding phase. The multi-scale feature fusion module optimizes the 256-channel high-dimensional features output from the last layer of the encoder. It extracts multiple receptive field features through four parallel branches (1×1, 3×3, 5×5, and 7×7 convolutional kernels), concatenates them, compresses the number of channels to 256 through a 1×1 convolution, and then weights the fusion using a residual scaling factor to obtain the fused feature map. In the decoder stage, three layers of deconvolution (stride 2) are used to upsample the fused feature map. After each upsampling layer, it is concatenated with the encoder feature map (same size, same number of channels) filtered by the gated attention module. Then, through "convolution-BN-ReLU" processing, it is gradually restored to the input image size. The deep-supervised decoder head sets a lightweight segmentation head (1×1 convolution + BN) at the output of each decoding layer to generate an auxiliary prediction map. Although the inference stage does not rely on the auxiliary output, the multi-layer supervision during training has enabled the decoder to have accurate feature mapping capabilities. Finally, the 3-channel feature map output by the decoder is compressed to 1 channel by a 1×1 convolution. A linear activation function maps the pixel values ​​to dielectric constant values ​​in the range of 1 to 81, forming a dielectric constant distribution matrix consistent with the input image size. The beneficial effects are that this process achieves end-to-end mapping from the original GPR signal to the dielectric constant distribution without manual intervention, improving detection efficiency. The synergistic effect of each core module ensures high resolution and high accuracy of the dielectric constant distribution, providing reliable physical parameter support for subsequent disease identification.

[0060] In a specific embodiment, a unified software and hardware experimental platform was constructed. The experiment relied on the Windows 11 operating system, the PyTorch deep learning framework, and the Python programming language to complete the algorithm improvement and development. The detailed configuration parameters of the software and hardware environment are shown in Table 3.

[0061] During the model training phase, the total number of training epochs was set to 200; the batch size was set to 4; the initial learning rate (lr0) was set to 0.001; and the optimizer was Warmup + cosine annealing, as shown in Table 3, to ensure that the construction from environment to policy is reproducible.

[0062] Table 3

[0063] In the fields of statistics and machine learning, evaluating the predictive performance of regression models is a core step in the model validation process. To achieve objective and quantitative evaluation, a series of mathematical indicators are typically used to measure the consistency between model predictions and actual observations. The mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) used in this study are three of the most representative indicators, each revealing the model's performance characteristics from different dimensions.

[0064] MAE is the arithmetic mean of the absolute errors of all individual observations. It quantifies the average magnitude of the model's predicted residuals. Its mathematical expression is: ; In the formula, n is the sample size. Representing the Actual value of each sample This represents the predicted value of the i-th sample.

[0065] The MAE (Modulation Error Analysis) is based on the L1 norm and directly reflects the expected magnitude of the prediction error. Its units are consistent with the dependent variable, making it highly intuitive and interpretable. The core statistical property of the MAE lies in its robustness. Because it linearly penalizes the error, it is not highly sensitive to outliers in the data. This means that individual extreme errors will not disproportionately affect the overall result, allowing it to more stably reflect the model's general performance on a majority of samples.

[0066] The root mean square error (RMSE) is the square root of the mean of the squares of the prediction errors. It represents the standard deviation of the prediction residuals. Its mathematical expression is: ; In the formula, n is the sample size. Representing the Actual value of each sample This represents the predicted value of the i-th sample.

[0067] RMSE is based on the L2 norm and its dimensions are restored to the same level as the dependent variable through a square root operation for easier interpretation. Unlike MAE, RMSE applies a double penalty to the error. This characteristic makes it highly sensitive to large prediction errors. Specifically, large residuals are squared and amplified during the calculation, thus significantly affecting the final value of RMSE.

[0068] R² is a statistic that measures the goodness of fit of a model. It represents the proportion of the total variance of the dependent variable that can be explained by the model. Its mathematical expression is: ; In the formula, RSS (Residual Sum of Squares) is the residual sum of squares, and TSS (Total Sum of Squares) is the total sum of squares. n is the sample size. Representing the Actual value of each sample Let represent the predicted value of the i-th sample, and y represent the arithmetic mean of all actual observations.

[0069] The core value of R² lies in providing a relative performance benchmark. It evaluates the magnitude of a model's improvement in predictions compared to a benchmark, rather than the absolute magnitude of the prediction error. Therefore, it is independent of the scale of the dependent variable, facilitating comparisons of explanatory power across different datasets or models.

[0070] To verify the independent and synergistic effects of the proposed Multi-Scale Feature Fusion (MSF), Gated Skip Connection (GSF), Asymmetric Deep-Spatial Attention (ASDS), and Deep Supervision Mechanism (DS-Head) in GPR-TransUNet, ablation experiments were conducted on each of the proposed improved modules. Using the base TransUNet as a control, the experiments introduced MSF, GSF, ASDS, and DSHead under the same training configuration and data partitioning, respectively, and investigated their impact on the accuracy of dielectric constant inversion.

[0071] The ablation schemes are shown in Table 4. Model 1 is the baseline without model improvements, named as follows: Models 2-5 each individually incorporate MSF, GSF, ASDS, and DSHear; Models 6-11 combine MSF, GSF, ASDS, and DSHear modules in pairs; Models 12-15 combine MSF, GSF, ASDS, and DSHear modules in three combinations; and Model 16 is the complete GPR-TransUNet with all four modules fused. The model evaluation metrics are RMSE, MAE, and the coefficient of determination R. 2 .

[0072] Table 4

[0073] As can be seen from the ablation experiment results in Table 4, the original TransUNet (Model 1) had the lowest RMSE, MAE, and R² scores, indicating that the baseline model has limitations in handling complex ground-penetrating radar inversion tasks. After introducing the improved modules one by one, the model performance was improved to varying degrees.

[0074] In the single-module introduction stage, the model that introduced MSF alone (Model 2) performed the best, with RMSE decreasing from 2.927 to 2.845 and R² increasing to 0.652, indicating that multi-scale features can effectively enhance adaptability to the spatial scale of underground structures. Notably, the model that introduced ASDS alone (Model 4) also showed significant advantages, with RMSE decreasing to 2.860 and R² reaching 0.648, outperforming the models that introduced GSF alone (Model 3) or DS-Head alone (Model 5). This result confirms that the asymmetric depth-spatial attention mechanism (ASDS) plays an independent and crucial role in suppressing background noise and focusing on effective target features. In contrast, without robust feature extraction or attention guidance, relying solely on gating or deep supervision mechanisms (Model 3 and Model 5) resulted in relatively limited performance improvements.

[0075] Further analysis of the synergistic effect of the dual modules reveals a qualitative leap in the combination of ASDS with other modules. In particular, Model 8 (MSF+ASDS) shows a significant reduction in RMSE to 2.795 and an increase in R² to 0.665, significantly outperforming Model 7 (RMSE 2.848) which combines MSF and GSF. This indicates that the rich multi-scale features provided by MSF, under the attention weighting of ASDS, can more accurately locate and reconstruct regions of dielectric constant anomalies, demonstrating a strong complementarity between the two.

[0076] In the stage of deep integration of multiple modules, the role of ASDS becomes even more prominent. Comparing the three-module combination Model14 (MSF+GSF+DS, RMSE 2.840) without ASDS with the combination Model12 (MSF+GSF+ASDS, RMSE 2.770) with ASDS, the latter has significantly better performance, proving that ASDS is a key component for improving the upper limit of the model.

[0077] Ultimately, Model16, integrating all four modules, reached the global optimum. Its RMSE decreased to 2.750, MAE stabilized at 0.893, and R² reached a high of 0.674. This result clearly demonstrates that the multi-scale capture of MSF, the spatial attention focus of ASDS, the feature gating of GSF, and the deep gradient supervision of DS-Head form a closed-loop synergy in GPR-TransUNet. This synergy not only enhances the network's ability to represent features of complex subsurface media but also effectively solves the gradient vanishing and feature redundancy problems in deep network training, thereby achieving high-precision inversion of the dielectric constant of ground-penetrating radar.

[0078] Figure 7 This demonstrates the changes in loss and evaluation metrics of the proposed GPR-TransUNet over 200 training epochs.Figure 7 (a) It can be seen that the training set and validation set losses decrease rapidly in the initial stage and gradually stabilize after about 50 cycles; the convergence trends of the validation curve and the training curve are basically consistent, indicating that the model does not have obvious overfitting and the overall convergence is stable.

[0079] Figure 7 (b) The results show that RMSE and MAE decreased rapidly from their initial high values ​​and then entered a slow decline phase around 60 cycles. 2 After a rapid initial rise, the value gradually approaches 0.5, eventually maintaining smooth fluctuations throughout the training process. In the mid-to-late stages, the amplitude of each metric curve decreases significantly, exhibiting near-smooth changes. This is related to the cosine annealing combined with the Warmup learning rate scheduling strategy employed. This strategy alleviates gradient oscillations in the early training phase through the Warmup stage, and then uses cosine annealing to gradually reduce the learning rate, resulting in a smooth decrease in the parameter update step size, thereby effectively suppressing later-stage loss and metric fluctuations. Here, RMSE represents the root mean square error, MAE represents the mean absolute error, R² represents the coefficient of determination, and Epoch represents the training epoch.

[0080] The smooth convergence of loss and evaluation metrics not only verifies the training stability of GPR-TransUNet, but also shows that the learning rate strategies of cosine annealing and Warmup play a positive role in ensuring global search and subsequent refinement.

[0081] Table 5

[0082] To verify the effectiveness of the proposed model, it was compared with three powerful mainstream network models: U-Net++, DeepLabV3+, and ConvNeXt. The models were compared on the same ground-penetrating radar dielectric constant inversion dataset. Table 5 lists the key evaluation metrics for each model. The results show that the proposed model achieves the best overall performance with an RMSE of 2.750, a MAE of 0.893, and an R² of 0.674. In contrast, U-Net++ has an RMSE of 3.337 and a MAE of 1.173, with an R² of 0.525, indicating significant deficiencies in continuous value regression and detail recovery. Although DeepLabV3+ and ConvNeXt show some improvement over U-Net++, their RMSE, MAE, and R² still lag behind the proposed model. Figure 8 The node constant inversion images of radar images are obtained by using three mainstream models, U-Net++, DeepLabV3+, and ConvNeXt, and the GPR-TransUNet model. Figure 8 The results show that the proposed improved network has stronger expressive and generalization capabilities in capturing the spatial continuity and local details of the dielectric constant.

[0083] To further verify the practical performance of the proposed GPR-TransUNet in ground-penetrating radar dielectric constant inversion tasks, this section presents a visual comparison of typical test samples. The examples include complete profiles of asphalt layers and water-stabilized layers, covering typical defects such as cavities, loose bodies, water-rich bodies, and voids, to examine the model's ability to reproduce defects in complex media environments. Figure 9 This section presents the inversion results of GPR-TransUNet for different types of defects. Overall, the GPR-TransUNet model accurately recovers the spatial distribution characteristics of the dielectric constant under various defect scenarios. Comparing the inversion results for different typical defects, it can be seen that the local details of voids, delamination, and water-rich areas are almost identical to the labels, indicating that the model has good capabilities in recovering interface structure and dielectric features. However, for porous regions, the predicted results are slightly blurry compared to the true labels. This is mainly because the convolution operator suffers from spatial detail loss during feature extraction, making it impossible to preserve the texture. The error map further reveals the spatial distribution characteristics of the prediction bias. Overall, the errors are mainly concentrated at the defect boundaries, reflecting a large dielectric constant gradient in the interface transition region, indicating that the model still faces some difficulty in representing subtle edges. For porous regions, the errors exhibit a more diffuse overall distribution, rather than being limited to the boundaries. This phenomenon suggests that convolutional feature extraction may not be able to fully capture the continuously and rapidly changing dielectric information of highly textured porous structures, leading to an increase in overall prediction bias. To verify the engineering applicability of the GPR-TransUNet model in real-world road distress detection, this study conducted radar measurements and borehole sampling on a national and provincial highway in Province X. Figure 10 The images shown are a Bscan image from a ground-penetrating radar scan with a center frequency of 900MHz, an image retrieved from the dielectric constant of the GPR-TransUNet model, and an endoscopic image taken after drilling (right image). Figure 10 (a) is a schematic diagram of road segment 1; Figure 10 (b) is a schematic diagram of the dielectric constant inversion results for section 1; Figure 10 (c) is a schematic diagram of endoscopic examination of section 1; Figure 10 (d) is a schematic diagram of road segment 2; Figure 10 (e) is a schematic diagram of the dielectric constant inversion results for section 2; Figure 10 (f) is a schematic diagram of endoscopic examination of section 2; Figure 10 (g) is a schematic diagram of road segment 3; Figure 10 (h) is a schematic diagram of the dielectric constant inversion results for section 3; Figure 10 (i) is a schematic diagram of endoscopic examination of section 3.

[0084] The inversion results showed that large areas of low dielectric constant anomalies appeared in all three road sections. The average relative dielectric constant within these areas was close to 1, and they were distributed in a banded or blocky pattern, highly consistent with typical voiding characteristics. To further verify the reliability of the model's inferences, borehole sampling was carried out in the anomaly areas, and direct observation was performed using borehole endoscopy imaging. On-site endoscopic results confirmed that all anomaly areas contained obvious voiding and cavity structures, perfectly matching the model predictions.

[0085] This example demonstrates that GPR-TransUNet can not only accurately reconstruct the dielectric constant distribution under laboratory conditions, but also stably capture the electrical anomalies of large-scale defects in real road environments. Compared with traditional manual interpretation relying solely on radar images, the model in this application can directly provide a quantitative dielectric constant profile, significantly improving the objectivity and accuracy of defect identification and providing reliable technical support for road structure health monitoring and rapid decision-making.

[0086] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated in this application, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Furthermore, Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0087] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0088] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for road dielectric constant inversion and defect identification based on neural networks, characterized in that, The method includes: Construct a GPR simulation dataset for roads; expand the number of skip connections in the TransUNet baseline model, introduce a gated attention mechanism at each skip connection, and design a multi-scale feature fusion module, an asymmetric depth-space attention module, and a deep supervised decoder head to form a GPR-TransUNet neural network model; AdamW is used as the optimizer, and the learning rate scheduling method is set in combination with the Warmup-cosine annealing strategy. An adaptive region-weighted loss function is constructed. The adaptive region-weighted loss function amplifies the loss weight of the dielectric constant change boundary through a dynamic weighting mechanism. The training set from the GPR simulation dataset is input into the GPR-TransUNet neural network model. The model is trained using the configured training strategy and the adaptive region weighted loss function. The loss change and evaluation index fluctuation of the model are monitored through the validation set during the training process. The model parameters are adjusted until the model converges, and the trained GPR-TransUNet neural network model is obtained. The GPR-B-scan image of the road to be detected is input into the trained GPR-TransUNet neural network model, which outputs the dielectric constant distribution of the road to be detected. Based on the dielectric constant characteristics corresponding to different diseases in the dielectric constant distribution, the model identifies cavities, voids, loose bodies and water-rich bodies in the road to be detected, thus completing the identification of hidden road diseases.

2. The method for road dielectric constant inversion and defect identification according to claim 1, characterized in that, The process of constructing a road GPR simulation dataset includes: Based on the gprMax 3D forward modeling platform, two types of road models, homogeneous and heterogeneous, were built. Four typical road defects, namely cavities, voids, loose bodies, and water-rich bodies, were randomly embedded in the two models to generate multiple B-scan images of the road GPR. The B-scan images were divided into training, validation, and test sets in an 8:1:1 ratio. At the same time, the true dielectric constant distribution corresponding to the B-scan images was extracted from the gprMax forward modeling results and mapped to uint16 grayscale images with the same size as the B-scan images as label data, thus completing the dataset construction.

3. The method for road dielectric constant inversion and defect identification according to claim 1, characterized in that, The implementation process of the gated attention mechanism includes: The high-level semantic features X output by the decoder and the low-level detail features S output by the encoder in the GPR-TransUNet neural network model are concatenated along the channel dimension to obtain the joint feature Z. The joint feature Z is compressed by the first 1×1 convolutional layer, followed by batch normalization and ReLU activation function operation. Then, it is operated by the second 1×1 convolutional layer and Sigmoid activation function to generate gating weights with the same number of channels as the low-level detail features. The gating weights are multiplied element-wise with the low-level detail features to obtain the weighted filtering features, thereby enhancing the effective features and suppressing the noisy features in the skip connections. The gating weights that generate the same number of low-level detail feature channels are: in, Indicates the gating weight, These are the weight parameters for the first 1×1 convolutional layer. These are the weight parameters for the second 1×1 convolutional layer. For the Sigmoid function, For ReLU function, For batch normalization operations, Z represents joint features.

4. The method for road dielectric constant inversion and defect identification according to claim 1, characterized in that, The implementation process of the asymmetric depth-space attention module includes: The module input feature map is linearly transformed along the channel dimension, and a Value feature map with the same size as the input feature map is generated through a 1×1 convolutional layer. We construct a deep branch and a spatial branch to capture features of different dimensions. The deep branch uses a k×1 size convolutional kernel to extract the correlation features of the vertical axis of the GPR map, thus obtaining a deep attention map. The spatial branch uses a 1×k size convolution kernel to extract the correlation features of the horizontal axis of the GPR map, thus obtaining the spatial attention map; The deep attention map and the spatial attention map are added together, and then compressed to the range [0,1] using the Sigmoid activation function to generate the final attention gating; The attention gating is multiplied element-wise with the Value feature map to obtain the recalibrated output feature map, which enables precise focusing on vertical layered features and lateral anomaly features in the GPR map.

5. The method for road dielectric constant inversion and defect identification according to claim 1, characterized in that, The construction process of the multi-scale feature fusion module includes: Four parallel feature extraction branches are set up, and each branch uses convolution kernels of size 1×1, 3×3, 5×5 and 7×7 respectively for convolution operation. Padding is set during the convolution process to ensure that the spatial resolution of the output feature map of each branch is consistent with the input feature map. After the convolution operation in each branch, batch normalization and ReLU activation function operation are performed sequentially to enhance the nonlinear representation of features. The output features of the four branches are concatenated along the channel dimension. The concatenated features are then compressed and integrated using a 1×1 convolutional layer to restore the number of output channels to the number of channels in the input feature map. Batch normalization and ReLU activation function operations are then performed again. By introducing a learnable residual scaling factor, the integrated features are weighted and fused with the module input features to obtain a multi-scale feature fusion result.

6. The method for road dielectric constant inversion and defect identification according to claim 1, characterized in that, The calculation process of the adaptive region-weighted loss function includes: Spatial gradient information is extracted from the true dielectric constant distribution of the tag data, and the approximate gradient values ​​in the horizontal and vertical directions of the true dielectric constant distribution are calculated. The gradient magnitude map is synthesized by calculating the square root of the sum of the squares of the approximate gradient values ​​in the horizontal and vertical directions pixel by pixel, and restoring the original size by padding with zeros. The gradient magnitude map is input into an auxiliary detection network consisting of two convolutional layers, a ReLU nonlinear activation layer, and a Sigmoid output layer to generate a dynamic key region weighted mask with the same size as the predicted dielectric constant map output by the GPR-TransUNet neural network model. The fundamental loss is calculated based on the global pixel-level error between the predicted dielectric constant map and the actual dielectric constant distribution. The loss in the critical region is calculated based on the predicted dielectric constant map, the actual dielectric constant distribution, and the dynamic critical region weighted mask. An adaptive region weighted loss function is obtained by weighting and summing the basic loss and the critical region loss using preset coefficients.

7. The method for road dielectric constant inversion and defect identification according to claim 6, characterized in that, The fundamental loss is calculated based on the global pixel-level error between the predicted dielectric constant map and the actual dielectric constant distribution. in, For feature map height, The width of the feature map. To predict the pixel value at position (i,j) of the dielectric constant map, The pixel value represents the actual dielectric constant at position (i,j).

8. The method for road dielectric constant inversion and defect identification according to claim 6, characterized in that, The loss in the critical region is calculated based on the predicted dielectric constant map, the actual dielectric constant distribution, and a dynamic critical region weighted mask, including: Based on the predicted dielectric constant map, the actual dielectric constant distribution, and the dynamic critical region weighted mask, the critical region loss is calculated as follows: in, For feature map height, The width of the feature map. To predict the dielectric constant plot in ( i,j The pixel value at position ) The true dielectric constant is distributed in ( i,j The pixel value at position ) Weighted mask for dynamic key regions in ( i,j The pixel value at the location.

9. The method for road dielectric constant inversion and defect identification according to claim 1, characterized in that, Set the learning rate in, The learning rate for the current round. To minimize the learning rate, The initial learning rate, For the current training round, This refers to the total number of training rounds.

10. The method for road dielectric constant inversion and defect identification according to claim 1, characterized in that, The GPR-B-scan image of the road to be detected is input into the trained GPR-TransUNet neural network model, which outputs the dielectric constant distribution of the road to be detected, including: After inputting the GPR-B-scan image of the road to be detected into the trained GPR-TransUNet, the asymmetric depth-space attention module is first applied to the encoder. The layered features of the vertical depth axis of the GPR image are captured by k×1 convolution, and the local abnormal features of the horizontal scanning axis are extracted by 1×k convolution. Then, the features are fused to generate attention gating, and the enhanced feature map of the encoding stage is obtained. The enhanced feature map is optimized by a multi-scale feature fusion module. Different receptive field features are extracted in parallel by multi-size convolutional kernels. After the output of each branch is processed by BN-ReLU, a residual scaling factor is introduced after concatenation in the channel dimension to weight and sum the fused features with the original high-dimensional features to obtain the feature map after feature fusion. During the feature transfer to the decoder, the gated attention module concatenates the low-level detail features of the encoder with the high-level semantic features of the decoder, compresses the dimension through 1×1 convolution-BN-ReLU, and then generates gated weights through Sigmoid. The gated weights are multiplied pixel by pixel with the original features of the encoder to filter out the detail features useful for dielectric inversion, and the filtered features are passed into the corresponding layer of the decoder. The deep supervised decoding head applies supervisory constraints to the output of each decoding layer, and an auxiliary prediction is generated by a lightweight segmentation head. The decoder output is then subjected to channel compression and linear activation to map the pixel values ​​of the feature map to the actual dielectric constant values, forming a dielectric constant distribution consistent with the size of the input image.

Images