Thermal insulation layer surface penetrating radar data enhancement method based on generative adversarial network

Abstract

The invention discloses a thermal insulation layer surface penetrating radar data enhancement method based on a generative adversarial network, belongs to the technical field of radar data, is used for penetrating radar data enhancement, and comprises the steps that a thermal insulation layer image is input into a radar data enhancement network, and the radar data enhancement network comprises a generator and a discriminator; 100-dimensional random noise is input into the generator and passes through the seven up-sampling middle feature layers in sequence to obtain a generated image, the generated image and a heat preservation layer image are input into an adaptive discriminator enhancement module of the discriminator together, the generated image and the heat preservation layer image pass through the seven down-sampling middle feature layers in sequence, and finally a processing result of the heat preservation layer image is obtained. According to the method, the defect of insufficient distribution of real data is overcome, and the detection robustness of the model on the weak reflection signal of the thermal insulation layer is remarkably improved; although the local detail fidelity of the generated sample is limited, the time-frequency disturbance feature enhances the analysis capability of the model on the cross-scale signal feature, and verifies the effectiveness of the data enhancement strategy.

Description

Technical Field

[0001] This invention discloses a method for enhancing radar data penetrating the surface of thermal insulation layers based on generative adversarial networks, belonging to the field of radar data technology. Background Technology

[0002] As a crucial component of building envelope, the exterior wall insulation layer is susceptible to defects such as hollowness, cracks, and water seepage, which can significantly reduce insulation performance and induce structural safety hazards. Surface-penetrating radar, due to its non-contact and high-resolution characteristics, has become a core technology for defect detection. However, it faces two major bottlenecks in practical engineering: (1) The cost of acquiring actual measurement data is high. High-altitude detection requires the erection of scaffolding or the use of drones to carry equipment, with the cost of a single detection exceeding 5,000 yuan. (2) There are insufficient samples for surface penetration radar detection. The types of defects in the insulation layer are complex and sparsely distributed. The signal characteristics of different defects and radar frequencies are significantly different, making it difficult for existing data to cover all detection scenarios. At the same time, SPR data usually requires high-resolution samples (such as 256×256 and above) for training deep learning models, while traditional generative adversarial networks (GANs) are unstable when generating high-resolution samples, which limits the effect of data augmentation. How to effectively expand the dataset and improve the generalization ability of the model is a problem that urgently needs to be solved.

[0003] Current deep learning-based radar detection methods heavily rely on large-scale labeled data, but the construction industry lacks publicly available datasets, resulting in insufficient model generalization ability. Therefore, constructing low-cost, high-fidelity simulation data augmentation methods has become a key breakthrough in promoting the intelligent transformation of building inspection. In forward modeling methods, gprMax, as a mainstream electromagnetic simulation tool, has been used to generate SPR training data. However, existing research is mostly limited to modeling regular geometries (such as rectangular voids and circular holes), failing to characterize the irregular bifurcation morphology of real cracks. Furthermore, single-model modeling requires manually writing .in files, taking over 30 minutes, which is insufficient for batch data requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a data enhancement method for radar data penetrating the surface of thermal insulation layers based on generative adversarial networks, so as to solve the problem of poor data enhancement effect in the prior art when processing radar data penetrating the surface of thermal insulation layers.

[0005] A method for enhancing radar data that penetrates the surface of thermal insulation layer based on generative adversarial networks includes inputting an image of the thermal insulation layer into a radar data enhancement network, wherein the radar data enhancement network includes a generator and a discriminator. 100-dimensional random noise is input into the generator and passed through 7 upsampled intermediate feature layers to obtain the generated image (Generate data). The generated image and the insulation layer image (Real data) are then input into the adaptive discriminator enhancement module (ADA) of the discriminator and passed through 7 downsampled intermediate feature layers to obtain the final processed insulation layer image.

[0006] The seven upsampling intermediate feature layers sequentially include a first upsampling convolutional layer, a second upsampling convolutional layer, a first multi-scale self-attention layer (MSA), a third upsampling convolutional layer, a second multi-scale self-attention layer, a fourth upsampling convolutional layer, and a fifth upsampling convolutional layer.

[0007] The upsampling convolutional layer includes a transposed convolutional layer (ConvTranspose2d), a spectral normalization layer (SpectralNormalization), an instance normalization layer (Instance Normalization), and a ReLU activation function layer.

[0008] The input feature of the multi-scale self-attention layer has C channels and is divided into three branches. The first branch is input into a two-dimensional convolutional layer to obtain a channel count of C. The features are reshaped using the Softmax activation function to obtain the Q matrix. The second feature is then input into a 2D convolutional layer to obtain the number of channels. The features are reshaped using the Softmax activation function to obtain a channel number of . The K matrix, the third feature input to the 2D convolutional layer, yields a channel number of... The characteristics of the reshaping result in a channel number of . The V matrix, K matrix, and V matrix are multiplied by their eigenvalues, and the result is then multiplied by the Q matrix by its eigenvalues ​​to obtain the number of channels. The features are then reshaped to obtain the output features.

[0009] The two-dimensional convolutional layer includes a two-dimensional convolutional module and a spectral normalization layer.

[0010] Sample augmentation probability of the adaptive discriminator augmentation module yes: ; ; ; ; ; In the formula, It is discriminator loss. It is the gradient norm of the discriminator parameters. It is the gradient norm of the generator parameters. It is to prevent positive numbers with a denominator of zero. It is the generator loss. These are discriminator parameters. These are generator parameters. It is the expected value of the insulation layer image. It is the expected value of the generated image. It is an image generated from random noise. It is an image Perform discrimination processing.

[0011] The seven downsampling intermediate feature layers sequentially include a first downsampling convolutional layer, a second downsampling convolutional layer, a third multi-scale self-attention layer, a third downsampling convolutional layer, a fourth multi-scale self-attention layer, a fourth downsampling convolutional layer, and a fifth downsampling convolutional layer.

[0012] The downsampling convolutional layer includes a 2D convolutional layer, a spectral normalization layer, an instance normalization layer, a LeakyReLu activation function layer, and a dropout layer.

[0013] Compared with existing technologies, the present invention has the following advantages: the generated data significantly improves the performance of the detection model, and exhibits better consistency in depth prediction under complex media scenarios; the multipath scattering noise introduced by the generated data simulates the propagation characteristics of real electromagnetic waves, making up for the deficiency of real data distribution, and significantly improving the robustness of the model in detecting weak reflection signals of the insulation layer; although the local detail fidelity of the generated samples is limited, its time-frequency perturbation characteristics enhance the model's ability to analyze cross-scale signal features, verifying the effectiveness of the data augmentation strategy. Attached Figure Description

[0014] Figure 1 Diagram of radar data augmentation network structure; Figure 2 This is a diagram of the multi-scale self-attention layer structure; Figure 3 These are the original simulation data with a structural similarity of 0.75; Figure 4 This is the generated data of the present invention with a structural similarity of 0.75; Figure 5 These are the original simulation data with a structural similarity of 0.62; Figure 6 This is the generated data of the present invention with a structural similarity of 0.62; Figure 7 The original simulation data has a structural similarity of 0.57. Figure 8 This is the generated data of the present invention with a structural similarity of 0.57. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0016] A method for enhancing radar data penetrating the surface of thermal insulation layers based on generative adversarial networks, such as Figure 1 This includes inputting an image of the insulation layer into a radar data augmentation network, which includes a generator and a discriminator. 100-dimensional random noise is input into the generator and passed through 7 upsampled intermediate feature layers to obtain the generated image (Generate data). The generated image and the insulation layer image (Real data) are then input into the adaptive discriminator enhancement module (ADA) of the discriminator and passed through 7 downsampled intermediate feature layers to obtain the final processed insulation layer image.

[0017] The seven upsampling intermediate feature layers sequentially include a first upsampling convolutional layer, a second upsampling convolutional layer, a first multi-scale self-attention layer (MSA), a third upsampling convolutional layer, a second multi-scale self-attention layer, a fourth upsampling convolutional layer, and a fifth upsampling convolutional layer.

[0018] The upsampling convolutional layer includes a transposed convolutional layer (ConvTranspose2d), a spectral normalization layer (SpectralNormalization), an instance normalization layer (Instance Normalization), and a ReLU activation function layer.

[0019] like Figure 2 The input feature of the multi-scale self-attention layer has C channels and is divided into three branches. The first branch is input into a two-dimensional convolutional layer to obtain a channel count of C. The features are reshaped using the Softmax activation function to obtain the Q matrix. The second feature is then input into a 2D convolutional layer to obtain the number of channels. The features are reshaped using the Softmax activation function to obtain a channel number of . The K matrix, the third feature input to the 2D convolutional layer, yields a channel number of... The characteristics of the reshaping result in a channel number of . The V matrix, K matrix, and V matrix are multiplied by their eigenvalues, and the result is then multiplied by the Q matrix by its eigenvalues ​​to obtain the number of channels. The features are then reshaped to obtain the output features.

[0020] The two-dimensional convolutional layer includes a two-dimensional convolutional module and a spectral normalization layer.

[0021] Sample augmentation probability of the adaptive discriminator augmentation module yes: ; ; ; ; ; In the formula, It is discriminator loss. It is the gradient norm of the discriminator parameters. It is the gradient norm of the generator parameters. It is to prevent positive numbers with a denominator of zero. It is the generator loss. These are discriminator parameters. These are generator parameters. It is the expected value of the insulation layer image. It is the expected value of the generated image. It is an image generated from random noise. It is an image Perform discrimination processing.

[0022] The seven downsampling intermediate feature layers sequentially include a first downsampling convolutional layer, a second downsampling convolutional layer, a third multi-scale self-attention layer, a third downsampling convolutional layer, a fourth multi-scale self-attention layer, a fourth downsampling convolutional layer, and a fifth downsampling convolutional layer.

[0023] The downsampling convolutional layer includes a 2D convolutional layer, a spectral normalization layer, an instance normalization layer, a LeakyReLu activation function layer, and a dropout layer.

[0024] Currently, the mainstream external insulation systems for residential buildings are: EPS / XPS board thin-plaster external insulation systems, inorganic lightweight aggregate mortar external insulation systems, polystyrene particle insulation mortar external insulation systems, and rock wool thin-plaster external insulation systems. Among these, the electrical parameters of the insulation layer in EPS / XPS board thin-plaster external insulation systems are close to those of air, resulting in weak echoes from delamination and high detection difficulty. Therefore, subsequent research focuses on EPS / XPS board thin-plaster external insulation systems, from the outside to the inside: Finishing layer: ceramic tile / paint (3-5mm); Protective layer: crack-resistant mortar + fiberglass mesh (8-10mm); Insulation layer: EPS / XPS board (50-100mm); Adhesive layer: Polymer mortar (5-8mm); Base layer: reinforced concrete (≥200mm); Anchor bolts: galvanized steel anchor bolts (penetrating the insulation layer).

[0025] This invention constructs a refined multi-layered media model based on a typical EPS / XPS thin-plastered external insulation system. The external wall structure typically includes multiple layers such as an exterior finish layer, an insulation layer (e.g., EPS, XPS), a plastering mortar composite fiberglass mesh, and a base layer. During the simulation, the electrical parameters (relative permittivity, relative permeability, conductivity, and permeability) of each layer need to be set. The electromagnetic parameters of each layer are significantly affected by various factors such as environmental moisture content in practice. All electrical parameters used in the following simulations are approximate values ​​based on engineering applications; precise values ​​need to be calibrated using material testing data, as shown in Table 1.

[0026] Table 1 Dielectric constants of materials ; The simulation data was validated, and the accuracy of the model was verified using Paraview visualization. In the layer structure verification, the thickness error of each material layer was ≤0.5mm (JGJ / T 177-2009 allowable error); in the anchor bolt positioning accuracy, the horizontal spacing was 20cm±0.5cm, and the vertical positioning error was ≤1cm; for defect detectability, when the diameter of the hollow area was ≥5cm, the B-Scan signal attenuation was ≤-15dB. The morphological diversity of defects (such as disc-shaped, crack-shaped, and irregular cluster-shaped) and their embedment depth and horizontal distribution in the wall have a significant impact on the time-frequency characteristics of radar echoes. This invention constructs four typical parameterized models of voids using gprMax: disc-shaped voids correspond to local debonding scenarios, with their echo amplitude increasing nonlinearly with increasing diameter; crack-like defects simulate interlayer cracking, exhibiting hyperbolic phase reversal characteristics in the time domain; irregular polygonal voids (with an area equivalent to disc-shaped voids) generate high-frequency oscillations due to edge scattering effects; and the multi-void cluster model suffers from time-domain aliasing due to multiple reflections, resulting in a periodic trailing signal after the main reflection peak. These characteristics provide a quantitative basis for constructing decision trees in defect classification algorithms. The morphological diversity of defects (such as disc-shaped, crack-shaped, and irregular cluster-shaped) and their location (depth, horizontal offset) within the wall significantly affect the time-frequency characteristics of radar echoes. Void defects, as a typical problem in building exterior walls, usually appear at the boundary between the insulation layer and other layers. In the simulation, a void can be considered as a cavity discontinuous with the surrounding material medium. Initially, the void is thin and has a small area, so a cuboid thin sheet can be used to approximate the location and morphology of the void defect. Based on a parametric simulation model, the electromagnetic response patterns under different defect scenarios are quantitatively analyzed, providing a theoretical basis for defect classification.

[0027] Four typical defect morphology models are constructed, and the geometric parameters are set as follows: Type A (crack-like): Long axis 50mm-150mm, short axis 5mm-15mm, simulating interlayer cracking; Type C (irregular polygon): The area is equivalent to Type A, simulating construction defects; Type D (multiple hollow clusters): 3-5 hollow clusters are randomly distributed with a spacing of 10mm-30mm, simulating a large area of ​​hollow clusters; In addition, there are also hollow areas at the adhesive joints. Hollow areas are divided into normal hollow areas and abnormal hollow areas. Normal hollow areas are the gaps between the interface mortar and the polystyrene board at the bonding point. This is a normal hollow area and will definitely occur. Abnormal hollow areas refer to the gaps that occur at the bonding point between the polystyrene board and the interface mortar. This is an abnormal hollow area.

[0028] The depth and horizontal offset of defects significantly alter radar signal characteristics through propagation path loss and wavefront curvature effects. Simulation studies show that: 1) Depth impact: As the defect depth increases, the round-trip path of electromagnetic waves lengthens, leading to a linear increase in time delay, the rate of which is related to the average dielectric constant of the path. Simultaneously, electromagnetic waves experience energy attenuation during propagation in multi-layered media, with the amplitude decreasing exponentially with depth; at a depth of 5cm, the echo amplitude attenuates by 78% compared to 1cm. Furthermore, deep defects (>5cm) cause resonant frequency shifts due to dielectric constant gradient changes, resulting in broadening of the frequency domain energy distribution. 2) Horizontal offset: Horizontal displacement of the defect alters the curvature of the wavefront reaching the antenna, directly affecting the interchannel position of the hyperbola apex in the radar image. When the offset exceeds the antenna aperture, the lateral offset rate of the hyperbola apex can reach 82% (measured positioning error ±2.3cm), accompanied by sidelobe energy diffusion and phase distortion (center phase angle offset ≥45°). Horizontal offset weakens interchannel signal correlation and increases the false alarm probability of defect edge detection (by 21%). This characteristic provides a key constraint for beamforming optimization of multi-channel radar arrays.

[0029] GANs generate data by training two adversarial networks (Generator and Discriminator), such as... Figure 8 As shown, the Generator receives random input and generates generated data, i.e., fake data Xfalse. The Discriminator is responsible for determining the authenticity of the generated data Xfalse and the real data Xtrue. The Generator's optimization goal is to make Xfalse be considered real by the Discriminator, while the Discriminator's goal is the opposite. The competition between the Generator and the Discriminator iteratively occurs during training, eventually reaching a dynamic equilibrium. When input is noisy, a well-trained Generator can generate highly realistic data.

[0030] In the inspection task, the resolution of the images collected by the actual sampling and forward simulation of the insulation layer was 512×512. In order to ensure the uniformity of the image quality generated by the model and the resolution of the dataset, as well as the effectiveness of the model in extracting defect features, the convolutional layers of the GAN network generator and discriminator were redesigned.

[0031] The generator takes 100-dimensional random noise as input and then upsamples the feature map through transposed convolution to generate intermediate feature layers with scales of 32×32×512, 64×64×256, 128×128×128, and 256×256×64. To further enhance the model's ability to generate details in high-resolution images, considering the complex correlation between the local features of small cracks and defects in B-Scan images and the overall insulation layer structure, an improved MSA self-attention module was added after the last two intermediate feature layers (128×128×128, 256×256×64). By introducing the self-attention module at this specific level, the model can generate richly detailed local features while also ensuring accurate reproduction of the overall insulation layer structure, ensuring that the generated image reflects the characteristics of surface-penetrating radar data in a detailed and accurate manner. Furthermore, the 128 and 256 feature maps capture feature representations from coarse to detailed, respectively. The improved self-attention mechanism enables the model to pay more attention to the global dependencies and interactions between features at these levels, thereby improving the accuracy of defect feature reproduction. Finally, the feature map is expanded to 512×512×32 through an additional deconvolution step, and a 3×3 3-channel convolution kernel is used to further adjust the feature map. All intermediate convolutional layers are processed with ReLU activation function and double normalization to improve training stability and efficiency. Finally, a 512×512×3 image is output through tanh activation function.

[0032] The discriminator takes a 512×512×3 defect image as input and performs downsampling through Conv2d convolution, generating intermediate feature layers with scales of 256×256×32, 128×128×64, 64×64×128, and 32×32×256 respectively. An improved self-attention module (MSA) is added after the 64×64×128 and 32×32×256 intermediate feature layers. The 256×256 layer focuses on more macroscopic layout and structural features, while the 128×128 layer focuses on finer defect features and texture information. This fusion of multi-scale features allows the discriminator to consider both macroscopic structure and microscopic details when judging image authenticity, enhancing the model's discriminative ability. Furthermore, adding a self-attention mechanism to the mid-to-high resolution layers helps improve the discriminator's generalization ability, enabling it to better handle B-Scan images in various scenarios. Generalization ability is a crucial evaluation factor for generative adversarial networks, directly affecting the model's performance on unseen data. All intermediate feature layers use the LeakyReLU activation function, with the hyperparameter alpha set to 0.1 to coordinate the learning rates of the generator and discriminator. Finally, a fully connected layer is connected to output a probability value between 0 and 1, which is normalized using the spectral norm. This is then used for error calculation of the ground truth label and backpropagation to optimize the network. Through these improvements, the radar data augmentation network SPR-GAN of this invention can more effectively generate 512×512 images that match the resolution of actual insulation layer defect images, while ensuring high-quality generated images and stable model training.

[0033] To ensure stable training and performance of SPR-GAN under different conditions, three key designs are employed in the SPR-GAN network: dual normalization, the ADA module, and the MSA module. Dual normalization is used to control the stability of the network structure and enhance its adaptability to high-resolution generation tasks; the ADA module enables the network to train stably on small datasets; and the MSA module enables SPR-GAN to generate complex or structurally large surface-penetrating radar defect data.

[0034] Normalization layers are an essential component of advanced deep learning networks. They normalize the outputs of convolutional and transposed convolutions to a standard normal distribution to overcome gradient vanishing and gradient explosion problems in deep networks and accelerate training. SPR-GAN employs instance normalization (IN) to constrain the outputs of convolutional and transposed convolutional layers in the network. Besides limiting output, Generative Adversarial Networks (GANs) are prone to early training collapse. In the early stages of training, low-quality samples obtained by the generator under Gaussian noise often lead the discriminator to easily determine whether a sample is real or fake and converge quickly, causing the adversarial relationship between the generator and discriminator to end prematurely. Theoretically, the loss function of a GAN can constrain the difference in JS divergence between generated and real data. In the early stages of GAN training, the distributions of low-quality generated samples are usually disjoint from those of real samples, and the JS divergence is a constant value, failing to provide gradients. Therefore, the discriminator converges rapidly in the early stages of training, causing the adversarial training to lose balance and leading to GAN instability. The distribution of high-resolution samples is more complex, and the higher the task resolution, the stronger the instability of GANs.

[0035] Dual normalization includes spectral normalization (SN) and instance normalization (IN). SN constrains abrupt changes in model parameters, while IN constrains abrupt changes in the outputs of convolutions and transposed convolutions. Under the traditional batch normalization scheme, the generator loss curve exhibits significant divergence: although it briefly decreases in the initial stage, it begins to oscillate and climb continuously after about 150 iterations, far exceeding the initial loss value by the 500th iteration, indicating that the parameter update direction is out of control. The discriminator loss, while converging rapidly in the early stages, fluctuates violently in sync after the generator collapses, reflecting the complete breakdown of the adversarial balance. In contrast, the dual normalization architecture exhibits drastically different dynamic characteristics—the generator and discriminator loss curves maintain a stable coupled downward trend, simultaneously entering a plateau period after about 300 iterations, preserving the equilibrium of their gradient game. This difference stems from dual normalization's targeted optimization of the adversarial training dynamics: by establishing independent normalization channels in the generator and discriminator, it alleviates the gradient conflicts caused by shared statistics in traditional schemes. Specifically, the generator's feature mapping retains its sensitivity to the discriminator's feedback during the normalization process, while the discriminator's batch statistics calculation avoids interference from the mixing of generated and real samples. Experiments demonstrate that dual normalization, by decoupling the normalization paths of the two types of networks, effectively suppresses gradient drift caused by statistical coupling in traditional methods, transforming the convergence trajectory of adversarial training in the loss space from divergent oscillation to asymptotic convergence, thus providing a new technical paradigm for the stable optimization of deep GAN models.

[0036] In the SPR-GAN framework, the discriminator overfitting problem stems from the model's excessive sensitivity to local noise and random patterns in the training samples on small datasets, leading to an ill-conditioned contraction of the decision boundary outside the real data distribution. To break this deadlock, the proposed ADA mechanism reconstructs the discriminator's feature learning space through a dual mechanism of physical constraint enhancement and dynamic coupling feedback.

[0037] GANs are prone to discriminator overfitting on small datasets, leading to training instability. Since it's difficult to determine the specific dataset required for stable training of a SPR-GAN, the ADA module applies adaptive probabilistic sample augmentation to both real and generated samples to prevent instability on small datasets. The sample augmentation probability of the ADA module is: ; when The following physical constraint transformation is applied to the input samples: a virtual medium layer is constructed, and the characteristic impedance of the real image is... Statistical analysis from training data (e.g., concrete) The dielectric impedance of the generated image The reflection coefficient is implicitly encoded by the generator. for: .

[0038] An exponential decay is applied to the signal amplitude, and random perturbations are introduced into the signal phase in the Short-Time Fourier Transform (STFT) domain: the signal is divided into frames (64 frames, 50% overlap), the phase spectrum of each frame is calculated, and 10%~20% of frequency points are randomly selected by probability to add phase shifts; the signal is reconstructed by inverse STFT to generate enhanced samples with multipath interference characteristics. To be compatible with end-to-end training of generative adversarial networks, all enhancement operations are designed to be differentiable: the impedance attenuation coefficient Γ is used as an auxiliary output of the generator, and gradient backpropagation is achieved through Gumbel-Softmax reparameterization; the phase perturbation is directionally shifted by real-imaginary part differentiable modulation of the complex domain STFT to avoid direct manipulation of the non-differentiable phase; wavelet packet decomposition is implemented using PyTorch differentiability to ensure that the energy entropy constraint term can be backpropagated. This improved scheme employs a triple coupling of physical field perturbation injection, energy entropy constraints, and gradient game modulation. This allows the discriminator to combat overfitting while implicitly learning the partial differential equation constraints of electromagnetic wave propagation, thereby guiding the generator to output radar signals conforming to Maxwell's equations under limited data. Compared to the original ADA module's simple strategy of relying solely on geometric enhancement, this method upgrades data augmentation from low-level pixel perturbations to high-level physical law encoding, achieving domain-adaptive stabilization of the generative adversarial process.

[0039] In SPR-GAN with the ADA module, the loss curves of the generator (srcG) and discriminator (srcD) exhibit highly coordinated dynamic characteristics: both start from high initial values, enter a stable descent channel synchronously after about 200 iterations, and eventually converge to a low value range. The generator loss shows minimal fluctuations throughout, indicating stable parameter update direction; the discriminator loss decreases synchronously and smoothly without abnormal fluctuations, reflecting the equilibrium of the adversarial game. This stable convergence mode stems from the dual-path control mechanism of the ADA module—by decoupling the calculation of feature statistics between the generator and discriminator, it eliminates gradient conflicts caused by shared statistics in traditional schemes. In contrast, in the model without the ADA module, the generator loss (noADAG) quickly enters a divergent state after a brief decrease, rising to several times its initial value with violent oscillations, revealing that the parameter space collapses due to gradient instability; the discriminator (noADAD), although quickly converging to near zero in the early stages, fluctuates violently synchronously after the generator goes out of control. This vicious cycle stems from the statistical coupling in traditional architectures: the generator and discriminator share feature distribution statistics, leading to mutual interference between gradient updates of generated and real samples. The ADA module addresses this by establishing independent normalization channels, allowing the generator to retain its sensitivity to discriminator feedback while ensuring the discriminator updates its parameters based on a clean data distribution. Experiments demonstrate that this design transforms the training dynamics from uncontrolled divergence to asymptotic convergence, providing a crucial solution for stable training of generative models in small-sample scenarios.

[0040] Generative adversarial networks (GANs) tend to generate smaller targets, such as rivets and cracks, but struggle to represent more complex and large structures. Because convolutional operations are localized, localized features are more easily propagated in fully convolutional models, making it challenging to capture and generate complex and large structures that depend on long-range dependencies. Self-attention (SA) has proven to be an effective method for addressing long-range dependencies in deep learning. However, the SA module consumes a lot of resources on large feature maps, with a complexity of O(n log n). , where N represents the size of the feature map. For surface penetration radar missions, where Self-attention MSA (Self-Attention Image Generation) has a significantly higher complexity than other low-resolution image generation tasks. This invention designs an improved self-attention MSA module with a complexity of O(n log n). ,in This refers to the number of channels. The MSA module enables SPR-GAN to handle complex geometries without consuming excessive computational resources. Ablation experiments using the MSA module demonstrate the results of SPR-GAN generating external wall insulation layer samples. Without the MSA module, the layer structure of the external wall insulation layer is not obvious, and the layer phase axes are blurred. However, with the MSA module, richer interlayer details can be clearly observed.

[0041] To assess the difference between the generated data and the original data, this invention uses the Structural Similarity Index (SSIM) as an evaluation metric. Additionally, the F1 score is used to measure the detection performance.

[0042] This invention employs Python-driven gprMax batch modeling and simulation technology to construct a multiphysics coupled model of a building exterior wall insulation layer system, generating the simulation dataset required for SPR-GAN training. The model strictly adheres to the layered structure design in actual engineering. Through a parametric modeling algorithm, three typical defects are randomly injected: hollow defects (5cm-20cm irregular debonding zones), crack defects (1mm-5mm wide bifurcated cracks), and water seepage defects (high dielectric regions with a moisture content ≥30%), generating a total of 400 sets of B-Scan images (512×512 pixels resolution). Each image contains a time-domain impulse response signal and superimposed quantization noise (…). ), where the hollow defect has a dielectric constant close to that of air ( The reflected wave amplitude is attenuated by about 15dB compared to the base layer signal, forming a characteristic aliasing with the rivet interference signal, which highly reproduces the low signal-to-noise ratio detection environment in the measured data. This dataset was visualized and verified by Paraview to verify the layer thickness error (≤0.5mm) and defect location accuracy (±2.5mm), providing high-fidelity training samples for SPR-GAN.

[0043] The simulated dataset was divided into a training set (368 images) and a test set (32 images). The training set was used for the SPR-GAN generation task, resulting in an equal number of generated images (368 images). The generated and simulated data were difficult to distinguish visually, and exhibited significant differences in structure and morphology. The original simulation data with a structural similarity of 0.75 is shown below. Figure 3 Generate data such as Figure 4 The original simulation data with a structural similarity of 0.62 are as follows: Figure 5 Generate data such as Figure 6 The original simulation data with a structural similarity of 0.57 are as follows: Figure 7 Generate data such as Figure 8 To further investigate the impact of SPR-GAN-generated data on the performance of detection models in deep learning, three datasets were constructed, consisting of generated samples and real samples, as shown in Table 2. The dataset `real` comprises real simulation data used to train SPR-GAN, the dataset `gan` comprises SPR-GAN-generated data, and `both` is the sum of `real` and `gan`. The datasets `real`, `gan`, and `both` were used to train Faster R-CNN and SSD, respectively, to evaluate the impact of SPR-GAN-generated data on deep learning detection models.

[0044] Table 2 Dataset partitioning in simulated data detection network testing ; Three different detection models were trained using different datasets to produce typical results. In terms of confidence, all three models improved under training with Both. The gap between false negatives and false positives was prevalent in Real and Gann, but decreased under training with Both.

[0045] The improvement in defect detection using SPR-GAN-generated data is limited. Surface-penetrating radar (SPR) is based on the theory of electromagnetic wave propagation; rivets made of metal completely reflect electromagnetic wave energy, creating strong clutter interference in the SPR profile. Defect signals are easily shielded by the rivet layer, becoming relatively weak, making it extremely difficult for defect detection models to distinguish them. Adding SPR-GAN-generated data to the detection model improved seepage detection performance, resulting in an F1 score improvement of over 7.18% on Faster R-CNN and SSD.

[0046] The training and testing environment for SPR-GAN was a Windows system equipped with an NVIDIA TESLA P40 24.0 GB GPU. SPR-GAN was written using the PyTorch framework, and the experimental hyperparameters are shown in Table 3.

[0047] Table 3 Hyperparameters of SPR-GAN ; This invention addresses the challenges of scarce SPR (Surface Physical Reflection) inspection data for building exterior walls and the coupling of defect features by proposing an innovative method that combines physical simulation with constraint generative adversarial networks (GANs) for collaborative enhancement, achieving the following core results: The batch modeling and simulation system is constructed. Based on Python and gprMax, a parametric modeling framework is developed to realize the random generation of geometric features of defects such as hollowness, cracks and water seepage. The system integrates equipment quantization noise and environmental interference pulses (peak intensity 5%-20%) to generate 400 sets of standardized datasets containing 4 types of defect combinations.

[0048] The SPR-GAN architecture innovates by proposing a dual normalization strategy (spectral normalization + instance normalization) to stabilize the generation process of high-resolution (256×256) data. It captures the topological correlations of crack edges through a multi-scale self-attention module (MSA) to generate long-range dependency features of defects. An adaptive discriminator enhancement (ADA) module dynamically balances differences in data distribution. Even with the highest-paired SSIM data, the SPR-GAN-generated data and real data still differ in reflection morphology, number of holes, and hole location.

[0049] SPR-GAN-generated data significantly improved the performance of the detection model. When training with a mixture of real and generated data, the model's stability in target localization was enhanced, especially in depth prediction consistency under complex media conditions. The multipath scattering noise introduced by the generated data simulated the propagation characteristics of real electromagnetic waves, compensating for the insufficient distribution of real data and significantly improving the model's robustness in detecting weakly reflected signals from insulation layers. Although the local detail fidelity of the generated samples was limited, their time-frequency perturbation features strengthened the model's ability to resolve cross-scale signal features, validating the effectiveness of the data augmentation strategy.

[0050] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for enhancing radar data penetrating the surface of thermal insulation layers based on generative adversarial networks, characterized in that, This includes inputting an image of the insulation layer into a radar data augmentation network, which includes a generator and a discriminator. 100-dimensional random noise is input into the generator and passed through 7 upsampled intermediate feature layers to obtain the generated image (Generate data). The generated image and the insulation layer image (Real data) are then input into the adaptive discriminator enhancement module (ADA) of the discriminator and passed through 7 downsampled intermediate feature layers to obtain the final processed insulation layer image.

2. The method for enhancing radar data penetrating the surface of thermal insulation layers based on generative adversarial networks according to claim 1, characterized in that, The seven upsampling intermediate feature layers sequentially include a first upsampling convolutional layer, a second upsampling convolutional layer, a first multi-scale self-attention layer (MSA), a third upsampling convolutional layer, a second multi-scale self-attention layer, a fourth upsampling convolutional layer, and a fifth upsampling convolutional layer.

3. The method for enhancing radar data penetrating the surface of an insulation layer based on generative adversarial networks according to claim 2, characterized in that, The upsampling convolutional layer includes a transposed convolutional layer (ConvTranspose2d), a spectral normalization layer (SpectralNormalization), an instance normalization layer (Instance Normalization), and a ReLU activation function layer.

4. The method for enhancing radar data penetrating the surface of thermal insulation layer based on generative adversarial networks according to claim 3, characterized in that, The input feature of the multi-scale self-attention layer has C channels and is divided into three branches. The first branch is input into a two-dimensional convolutional layer to obtain a channel count of C. The features are reshaped using the Softmax activation function to obtain the Q matrix. The second feature is then input into a 2D convolutional layer to obtain the number of channels. The features are reshaped using the Softmax activation function to obtain a channel number of . The K matrix, the third feature input to the 2D convolutional layer, yields a channel number of... The characteristics of the channel were reshaped to obtain the number of channels. The V matrix, K matrix, and V matrix are multiplied by their eigenvalues, and the result is then multiplied by the Q matrix by its eigenvalues ​​to obtain the number of channels. The features are then reshaped to obtain the output features.

5. The method for enhancing radar data penetrating the surface of thermal insulation layers based on generative adversarial networks according to claim 4, characterized in that, The two-dimensional convolutional layer includes a two-dimensional convolutional module and a spectral normalization layer.

6. The method for enhancing radar data penetrating the surface of an insulation layer based on generative adversarial networks according to claim 5, characterized in that, Sample augmentation probability of the adaptive discriminator augmentation module yes: ； ； ； ； ； In the formula, It is discriminator loss. It is the gradient norm of the discriminator parameters. It is the gradient norm of the generator parameters. It is to prevent positive numbers with a denominator of zero. It is the generator loss. These are discriminator parameters. These are generator parameters. It is the expected value of the insulation layer image. It is the expected value of the generated image. It is an image generated from random noise. It is an image Perform discrimination processing.

7. The method for enhancing radar data penetrating the surface of thermal insulation layers based on generative adversarial networks according to claim 6, characterized in that, The seven downsampling intermediate feature layers sequentially include a first downsampling convolutional layer, a second downsampling convolutional layer, a third multi-scale self-attention layer, a third downsampling convolutional layer, a fourth multi-scale self-attention layer, a fourth downsampling convolutional layer, and a fifth downsampling convolutional layer.

8. The method for enhancing radar data penetrating the surface of an insulation layer based on generative adversarial networks according to claim 7, characterized in that, The downsampling convolutional layer includes a 2D convolutional layer, a spectral normalization layer, an instance normalization layer, a LeakyReLu activation function layer, and a dropout layer.

Images