A method for quantitatively diagnosing a cavity in a GPR tunnel lining by suppressing noise interference

Abstract

The application discloses a kind of GPR tunnel lining cavity quantitative diagnosis methods of noise interference suppression.The method comprises: obtaining target radar data, the target radar data is the tunnel section corresponding radar data of the disease part of target tunnel;Image reconstruction is carried out based on the target radar data, and target radar image is obtained;The target radar image is analyzed, and the cavity condition diagnosis report of the target tunnel is obtained.The GPR tunnel lining cavity quantitative diagnosis method of noise interference suppression provided by the application can solve the problem that there is no method in the prior art that can accurately quantify the size of the cavity in the tunnel based on the de-noised radar image.Accurately reconstruct the radar reflection pattern in the tunnel, and accurately identify the state of the cavity in the tunnel.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a quantitative diagnostic method for voids in GPR tunnel lining with noise interference suppression. Background Technology

[0002] Subway tunnel linings are prone to cracks and voids, posing a threat to system safety. Voids are the most dangerous type of defect, often caused by insufficient grouting or soil erosion, and occur between lining layers, requiring different maintenance strategies. Small voids can be repaired locally, while large voids require extensive intervention. Therefore, accurately diagnosing the location and size of voids is crucial.

[0003] Ground-penetrating radar (GPR) is widely used in tunnel lining inspection due to its advantages such as being non-destructive and having high resolution. The difference in dielectric properties between cavities and the lining causes radar wave reflection or diffraction, and the reflected signal pattern varies with the size of the cavity, which is crucial for diagnosis. However, GPR data interpretation relies on professional knowledge, and the results are variable, affecting accuracy. The expansion of the subway tunnel network makes manual interpretation time-consuming and labor-intensive, necessitating efficient and reliable automated interpretation methods.

[0004] In existing technologies, deep learning methods are typically used to suppress rebar interference in GPR data to address the challenges. This study relies on generative models to reconstruct radar images and exclude rebar signals that enhance the reflection of defects.

[0005] However, diagnosing tunnel lining cavities from denoised radar images remains challenging, as the size of the cavities cannot be accurately quantified. Therefore, current technologies have made limited progress in quantitatively assessing tunnel lining defects using GPR data.

[0006] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0007] To address the aforementioned shortcomings of existing technologies, a quantitative diagnostic method for voids in GPR tunnel lining with noise interference suppression is provided, aiming to solve the problem that there is no existing method that can accurately quantify the size of voids in tunnels based on denoised radar images.

[0008] A first aspect of the present invention provides a method for quantitative diagnosis of voids in GPR tunnel lining with noise interference suppression, comprising:

[0009] Acquire target radar data, which is the radar data corresponding to the tunnel cross-section of the damaged part of the target tunnel;

[0010] Image reconstruction is performed based on the target radar data to obtain the target radar image;

[0011] The radar image of the target is analyzed to obtain a cavity condition diagnosis report for the target tunnel.

[0012] In one implementation, acquiring target radar data includes:

[0013] Acquire the initial radar data of the target tunnel, preprocess the initial radar data, and obtain the initial radar image;

[0014] An initial classification model is constructed, with EfficientNet V2s as the base model and the EMA module as the deep feature extraction module.

[0015] The initial classification model is pre-trained based on the target dataset to obtain the target classification model;

[0016] The radar data of the target tunnel is classified based on the target classification model to obtain the target radar data.

[0017] In one implementation, the step of reconstructing the image based on the target radar data to obtain a target radar image includes:

[0018] Construct a target CUT model, which includes a generator and a discriminator;

[0019] Based on the target CUT model, the interference of steel bars in the target radar data is eliminated to perform image reconstruction, thereby obtaining the target radar image.

[0020] In one implementation, constructing the target CUT model includes:

[0021] A lightweight attention mechanism, CBAM, is introduced in the decoding stage. CBAM consists of a channel attention module and a spatial attention module.

[0022] Obtain the target baseline loss function and the target hybrid loss function, wherein the target hybrid loss function is a hybrid function obtained by fusing MS-SSIM and L1 loss;

[0023] The target loss function is obtained by fusing the target baseline loss function and the target hybrid loss function;

[0024] In one implementation, eliminating the interference of reinforcing bars in the target radar data based on the target CUT model includes:

[0025] Weighted features are obtained based on the channel attention module;

[0026] Key spatial enhancement features are obtained based on the spatial attention module;

[0027] The interference of reinforcing bars in the target radar data is eliminated based on the weighted features, the spatial enhancement features, and the target loss function.

[0028] In one implementation, obtaining the weighted features based on the channel attention module includes:

[0029] Based on the channel attention module, global average pooling and global max pooling are applied to the target radar data to obtain a first feature vector and a second feature vector.

[0030] The first feature vector and the second feature vector are processed by MLP and then added together to obtain the channel attention map;

[0031] A weighted output feature map is obtained based on the target radar data and the channel attention map.

[0032] In one implementation, obtaining key spatial enhancement features based on the spatial attention module includes:

[0033] The target radar data is subjected to global average pooling and global max pooling based on the spatial attention module to obtain the first pooling feature and the second pooling feature.

[0034] After concatenating the first pooling feature and the second pooling feature, a target spatial attention map is obtained through a 7×7 convolutional layer;

[0035] The target space attention map is normalized to obtain the key space enhancement features.

[0036] In one implementation, the step of analyzing the target radar image to obtain a cavity condition diagnosis report for the target tunnel includes:

[0037] Based on Fast-RCNN, the target radar image is used to perform hole state diagnosis, and multiple target suggestions are obtained;

[0038] The target suggestion is projected onto the radar image of the corresponding tunnel section to obtain a cavity condition diagnosis report for the target tunnel.

[0039] In a second aspect, the present invention provides a terminal comprising: a processor and a storage medium communicatively connected to the processor, the storage medium being adapted to store a plurality of instructions, and the processor being adapted to invoke the instructions in the storage medium to execute the steps of the method for quantitative diagnosis of GPR tunnel lining voids that implements noise interference suppression as described in any of the preceding claims.

[0040] A third aspect of the present invention provides a storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the quantitative diagnosis method for noise interference suppression in GPR tunnel lining as described in any of the preceding claims.

[0041] Beneficial Effects: Compared with existing technologies, this invention provides a quantitative diagnostic method for voids in GPR tunnel linings with noise interference suppression. In this method, target radar data is first acquired, which is the radar data corresponding to the tunnel cross-section of the defective portion of the target tunnel. Then, image reconstruction is performed based on the target radar data to obtain a target radar image. Finally, the target radar image is analyzed to obtain a void status diagnostic report for the target tunnel. This method, by reconstructing radar data within the tunnel and analyzing the reconstructed radar image, solves the problem in existing technologies of lacking a method to accurately quantify the size of voids within the tunnel based on denoised radar images. It accurately reconstructs the radar reflection pattern within the tunnel and precisely identifies the state of voids within the tunnel. Attached Figure Description

[0042] Figure 1 A flowchart illustrating an embodiment of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining provided by the present invention;

[0043] Figure 2 An overall implementation flowchart of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention;

[0044] Figure 3 A schematic diagram of the classification process in an embodiment of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining provided by the present invention;

[0045] Figure 4 A classification model structure diagram in an embodiment of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining voids provided by the present invention;

[0046] Figure 5 A schematic diagram of the EMA module in an embodiment of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining provided by the present invention;

[0047] Figure 6 A schematic diagram of the CUT model structure in an embodiment of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining provided by the present invention.

[0048] Figure 7 In an embodiment of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining voids provided by the present invention, the location and structure of CBAM are shown.

[0049] Figure 8 Tunnel diagram in an experimental case of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention;

[0050] Figure 9 An example diagram of a cavity model in an experimental case of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining provided by the present invention.

[0051] Figure 10 Radar image from an experimental example in the embodiment of the quantitative diagnostic method for noise interference suppression in GPR tunnel lining voids provided by the present invention;

[0052] Figure 11 Experimental results in the experimental cases of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention. Figure 1 ;

[0053] Figure 12 Experimental results in the experimental cases of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention. Figure 2 ;

[0054] Figure 13 Experimental results in the experimental cases of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention. Figure 3 ;

[0055] Figure 14 Experimental results in the experimental cases of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention. Figure 4 ;

[0056] Figure 15 Experimental results in the experimental cases of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention. Figure 5 ;

[0057] Figure 16 Experimental results in the experimental cases of the method for quantitative diagnosis of GPR tunnel lining voids for noise interference suppression provided by the present invention. Figure 6 ;

[0058] Figure 17 A schematic diagram of the structure of an embodiment of the terminal provided by the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0060] The present invention provides a quantitative diagnostic method for GPR tunnel lining voids with noise interference suppression, which can be applied to a terminal with computing capabilities. The terminal can execute the quantitative diagnostic method for GPR tunnel lining voids with noise interference suppression provided by the present invention to diagnose tunnel lining voids.

[0061] Example 1

[0062] Currently, aging subway tunnel linings are prone to defects such as cracks, voids, delamination, and separation, posing risks to system safety. Among these, voids are the most threatening, usually formed due to insufficient grouting or soil erosion, and typically occur between lining layers. Since the required maintenance strategies vary, accurately diagnosing the location and size of voids in the tunnel lining is crucial. Small voids can be addressed with localized repairs (such as directional grouting), while larger voids may require more extensive interventions, such as relining or structural reinforcement. Therefore, a reliable method for diagnosing the condition of tunnel lining voids is essential to ensuring effective maintenance and avoiding unnecessary repairs (leading to subway operation disruptions or the neglect of dangerous defects).

[0063] Ground-penetrating radar (GPR) has been widely used in tunnel lining inspection due to its advantages of being non-destructive, high-resolution, efficient, and non-contact. Compared to the tunnel lining itself, the air in the cavities causes differences in dielectric properties, resulting in reflection or diffraction of radar waves at the defect boundaries. As the cavity size increases, the reflected signals exhibit different patterns in the radar image, serving as a key indicator for defect diagnosis. However, GPR data interpretation is challenging, often relying on specialized knowledge, leading to variability in results and potentially affecting detection accuracy. With the expansion of subway tunnel networks, the interpretation of manual GPR data is becoming increasingly time-consuming and labor-intensive, urgently requiring an efficient and reliable automated interpretation method.

[0064] Deep learning (DL) has emerged as a promising solution for large-scale quantitative interpretation of GPR (Gross Residue Permeability) data, such as estimation of rebar radius, classification of underground facilities, pavement thickness, and void dimensions. Recent research has developed robust models for identifying tunnel defects, primarily based on architectures such as region-based convolutional neural networks (RCNN), residual networks (ResNet), and YouOnly Look Once (YOLO), which are specifically designed to manage the complex patterns and variations encountered in tunnel environments. These models have significantly improved the accuracy and reliability of defect identification. However, identifying and diagnosing targets from tunnel GPR data presents greater challenges compared to other GPR applications. Infrastructure within tunnels, particularly rebar and utilities, generates diffraction and superimposed reflections of the GPR, complicating and distorting the reflections of target defects, resulting in noisy and blurred patterns.

[0065] To address these challenges, many researchers have focused on deep learning methods to suppress the interference of rebar in GPR data. These include the Re2-AttUnet network, equipped with residual and attention modules, which removes enhanced clutter and amplifies defect signals in GPR b-scan images. The RCE-GAN method, combining extended convolution and attention modules, suppresses enhanced signals at different intervals, improving the accuracy of void detection. However, current research also relies on generative models capable of reconstructing radar images to exclude rebar signals and enhance defect reflections. Nevertheless, diagnosing voids in tunnel linings from denoised radar images remains challenging.

[0066] (1) The effectiveness of suppression and reconstruction is highly dependent on the availability of sufficient training pairs, which are difficult to obtain in real-world engineering applications, thus limiting the generalization ability of the model.

[0067] (2) In scenarios where there are significant differences between tunnels with defects and normal tunnels, the performance of the reconstruction model will decrease. Since most tunnels are without defects, the suppression and reconstruction process may introduce artifacts or false defects, which may lead to misjudgment of tunnel lining voids.

[0068] (3) Although interference suppression models improve the accuracy of hole detection, these models still face difficulties in dealing with distorted or noisy reflection signal features.

[0069] Tunnel lining cavities vary in geometry and location, making it difficult to establish representative patterns for differentiation. These models may produce similar and distorted patterns, affecting the accurate assessment of cavity condition and hindering the precise quantification of cavity size. Therefore, current technology has made limited progress in quantitatively assessing tunnel lining defects using GPR data.

[0070] Based on this, this embodiment provides a quantitative diagnostic method for GPR tunnel lining voids to suppress noise interference and solve the above-mentioned problems. For example... Figure 1 As shown, the quantitative diagnostic method for noise interference suppression in GPR tunnel lining provided in this embodiment includes the following steps:

[0071] S100. Acquire target radar data, wherein the target radar data is the radar data corresponding to the tunnel cross-section of the damaged part of the target tunnel.

[0072] The acquisition of target radar data includes:

[0073] Acquire the initial radar data of the target tunnel, preprocess the initial radar data, and obtain the initial radar image;

[0074] An initial classification model is constructed, with EfficientNet V2s as the base model and the EMA module as the deep feature extraction module.

[0075] The initial classification model is pre-trained based on the target dataset to obtain the target classification model;

[0076] The radar data of the target tunnel is classified based on the target classification model to obtain the target radar data.

[0077] Reference Figure 2 In this embodiment, there are three stages in total. The first stage is to classify the target tunnel into scenes and extract the defect sections. The second stage is to suppress interference. The third stage is to diagnose voids.

[0078] Specifically, in the first stage, the GPR measurement results of the tunnel lining are classified. In this stage, radar segments with defects in the target tunnel are identified, while normal radar segments are excluded to obtain the target radar data. This significantly reduces the number of radar images that need to be reconstructed, significantly reduces workload, and improves diagnostic efficiency.

[0079] First, initial radar data of the target tunnel is acquired and preprocessed to obtain initial radar images. Specifically, GPR data of the target tunnel is collected as the initial radar data. This data typically exists in the form of radar images and includes different states of the lining (normal and defective). Then, the collected initial radar data is cleaned to remove noise and invalid data.

[0080] Then, a classification model is constructed. In this embodiment, the provided classification model is built upon the robust foundation of EfficienetNetV2s, leveraging the advantages of EMA and transfer learning to improve its performance, helping to avoid artifacts and spurious noise that may be introduced during reconstruction in normal tunnel segments. Specifically, the target classification model structure is as follows: Figure 3 As shown.

[0081] EfficientNet is a highly efficient neural network that overcomes the limitations of traditional convolutional networks (such as VGG, GoogleNet, and ResNet), which typically scale along a single dimension (such as width, depth, or resolution). By employing a compound scaling strategy, EfficientNet adjusts all three dimensions simultaneously when scaling the network, significantly improving both accuracy and training speed.

[0082] The EfficientNetV2 mentioned is a variant, with "s" representing a smaller network size. EfficientNet V2s have fewer parameters and lower computational complexity, making them suitable for lightweight tasks and applications requiring faster inference speeds. Its core structure employs inverse bottleneck convolutions, incorporating techniques such as depthwise separable convolutions, expanded convolutions, and pointwise convolutions, reducing computational complexity while maintaining high accuracy. In shallower layers, fuse-MBConv is an optimized version of MBConv, further improving the speed and performance of EfficientNetV2 by merging expanded convolutions and depthwise convolutions into a single operation, thus simplifying the architecture and reducing computational steps. Due to these advantages, EfficientNetV2s is used in the radar image scene classification stage. The structure of EfficientNet V2s, including MBConv and fuse-MBConv, is as follows... Figure 4 As shown.

[0083] Unlike optical images, tunneled GPR data is often affected by significant noise, making it difficult to distinguish weak signals. Attention mechanisms are widely used to enhance weak feature extraction. Among commonly used attention mechanisms, the EMA module stands out. The EMA module encodes the input feature map to recalibrate the channel weights within each parallel branch and group the channel size into multiple sub-features. By enabling cross-dimensional interactions, it captures pixel-level relationships, facilitating richer feature fusion. This enhancement improves the extraction of spatial and channel information, thereby increasing the model's ability to understand complex features, while also reducing parameter requirements and computational costs.

[0084] like Figure 5 As shown, in this embodiment, the deep feature extraction module employs an EMA module. The EMA module operates on a given input feature map, dividing the input feature X into sub-feature groups of RC (number of channels) × H (height) × W (width), and using three parallel processing paths to extract attention weight descriptors from these grouped feature maps. Two of these paths use 1×1 convolutional branches, which apply global average pooling in both spatial dimensions to encode channel information. The third path uses a 3×3 convolutional branch, designed to capture multi-scale feature representations. Subsequently, the spatial attention weights generated by these parallel paths are used to integrate the output feature maps of each group. These integrated output feature maps are processed by a sigmoid function, designed to capture pixel-level correlations, thereby enhancing the global contextual information of all pixels.

[0085] In this embodiment, placing the EMA module in deeper layers of EfficientNetV2s is crucial because these layers capture more abstract, high-level features that convey global information and semantic content, enabling them to perform complex classification tasks more effectively. This enhances multi-scale attention to these features, improving global consistency and context awareness. In contrast, placing the EMA module in shallow layers that focus on low-level details such as edges and textures may cause the model to overemphasize local features while ignoring more important information. Furthermore, while the EMA module improves feature representation, it also increases computational cost. Placing it in deeper layers ensures that these computational resources are used effectively to handle complex global features. Using the EMA module in shallow layers may result in unnecessary computational work because shallow features provide limited semantic information for more complex tasks.

[0086] Furthermore, due to the lack of real-world ground data and the scarcity of excavation verification in tunnel inspections, this embodiment introduces transfer learning to improve performance on small datasets. Transfer learning leverages knowledge gained from large-scale datasets to solve new tasks. This approach reduces training time for new tasks, minimizes the need for large amounts of labeled data, improves generalization, and reduces the risk of overfitting.

[0087] Currently, few pre-trained models possess features similar to radar charts, resulting in relatively low data similarity. Therefore, model-based transfer learning is more suitable for this study. To achieve transfer learning, this embodiment uses the ImageNet-21k dataset for pre-training. These datasets were chosen because they contain a wide range of visual features, which facilitates the learning of generalized representations.

[0088] Considering the relatively small size of the GPR tunnel lining defect dataset, this embodiment also includes fine-tuning the lower layers of the pre-trained model. This allows for the reuse of general features such as gradients and edge detection learned in the initial layers, while radar-specific features are extracted and fine-tuned in later layers of the model. Thus, by pre-training the initial classification model based on the target dataset ImageNet-21k, the target classification model can be obtained.

[0089] S200. Based on the target radar data, image reconstruction is performed to obtain the target radar image.

[0090] The process of reconstructing the target radar image based on the target radar data includes:

[0091] S210. Construct a target CUT model, which includes a generator and a discriminator.

[0092] The construction of the target CUT model includes:

[0093] S211. A lightweight attention mechanism CBAM is introduced in the decoding stage. The CBAM consists of a channel attention module and a spatial attention module.

[0094] S212, By fusing MS-SSIM and L1 loss, the target hybrid loss function is obtained;

[0095] S213. Construct the target CUT model based on the CBAM and the target hybrid loss function.

[0096] S220. Based on the target CUT model, the interference of steel bars in the target radar data is eliminated to perform image reconstruction, thereby obtaining the target radar image.

[0097] The process of eliminating interference from reinforcing bars in the target radar data based on the target CUT model includes:

[0098] S221. Obtain weighted features based on the channel attention module;

[0099] The acquisition of weighted features based on the channel attention module includes:

[0100] Based on the channel attention module, global average pooling and global max pooling are applied to the target radar data to obtain a first feature vector and a second feature vector.

[0101] The first feature vector and the second feature vector are processed by MLP and then added together to obtain the channel attention map;

[0102] A weighted output feature map is obtained based on the target radar data and the channel attention map.

[0103] S222. Obtain key spatial enhancement features based on the spatial attention module;

[0104] The acquisition of key spatial enhancement features based on the spatial attention module includes:

[0105] The target radar data is subjected to global average pooling and global max pooling based on the spatial attention module to obtain the first pooling feature and the second pooling feature.

[0106] After concatenating the first pooling feature and the second pooling feature, a target spatial attention map is obtained through a 7×7 convolutional layer;

[0107] The target space attention map is normalized to obtain the key space enhancement features.

[0108] S223. Based on the weighted features, the spatial enhancement features, and the target hybrid loss function, eliminate the interference of reinforcing bars in the target radar data.

[0109] Specifically, after the defective tunnel lining diagram, i.e. the target radar data, is extracted, in the second stage of this embodiment, the target radar data is input into the improved CUT model to suppress the steel reinforcement that interferes with the defective section, thereby enabling the reconstruction of the reflection pattern related to the shape and size of the cavity.

[0110] In this stage, by improving the CUT algorithm and combining it with CBAM to enhance feature extraction and the fusion loss function, fast and stable convergence was achieved. The overall structure of the improved CUT is as follows: Figure 6 As shown.

[0111] Specifically, CUT is a novel unsupervised method based on the concept of translating images from one domain to another without requiring paired examples, unlike traditional supervised methods. CUT employs a simpler architecture, requiring only a single network to perform the transformation from domain a to domain B, which reduces the number of model parameters and thus accelerates training convergence. Its effectiveness has been demonstrated in tasks involving the translation of optical, thermal, X-ray, and MRI images. However, applying the CUT algorithm to noisy and blurred GPR data presents significant challenges. The lack of backprojection limits its ability to maintain global structural consistency, making it prone to semantic inconsistencies during noise suppression, further impacting the reliability of defect detection in tunnel lining analysis.

[0112] In existing technologies, CUT is an unsupervised image translation method based on contrastive learning. Its core framework is a generative adversarial network, consisting of a generator and a discriminator. In this embodiment, the main objective is to construct the target CUT model based on the CUT algorithm to train enhanced and unenhanced signals on unpaired radar images to suppress rebar noise.

[0113] Specifically, the generator in the target CUT model employs an encoder-decoder architecture, consisting of four convolutional layers, nine residual blocks, and one additional convolutional layer. The encoder extracts high-level features from the input image and then passes them through nine sequentially connected residual blocks. Each residual block contains skip connections and batch normalization layers, which helps accelerate training and enhance stability. These residual blocks capture non-linear relationships between domains, learning the mapping between the input image and the target domain while mitigating the vanishing gradient problem. Finally, the decoder gradually increases the data dimensionality by using transposed convolutional layers, progressively restoring the feature mapping to the output image, thereby generating the final image in the target domain.

[0114] The discriminator in the target CUT model uses the PatchGAN architecture, consisting of five convolutional layers. After each convolutional operation, instance normalization is applied, followed by the non-linear activation function LeakyReLU. The final layer does not use a fully connected layer but outputs a 70×70 two-dimensional feature map to determine whether the image is real or fake. The convolutional layers are responsible for extracting local features from the input data, such as edges and textures, which helps the model recognize the content of the image. Instance normalization normalizes each sample to ensure that the feature map has a mean of 0 and a variance of 1, reducing the differences between samples and enhancing the consistency of the generated images. When the input is less than 0, the LeakyReLU activation function allows small negative outputs, preventing the "neuron death" problem and improving the model's representational power.

[0115] Specifically, since the CUT does not perform an inverse transformation, image feature information is often lost during the encoding and decoding process, thus limiting the similarity and quality between the generated image and the original image, and introducing irrelevant pseudo-signals in regions unrelated to the task. To address this issue, in this embodiment, a lightweight attention mechanism, CBAM, is introduced in the decoding stage to invariant information while maintaining processing speed and model size. CBAM consists of a channel attention module and a spatial attention module, which together enhance the extraction of semantic information in the channel and spatial dimensions, improving feature representation capabilities. Specifically, the location and structure of CBAM are as follows: Figure 7 As shown.

[0116] First, weighted features are obtained based on the channel attention module.

[0117] The acquisition of weighted features based on the channel attention module includes:

[0118] Based on the channel attention module, global average pooling and global max pooling are applied to the target radar data to obtain a first feature vector and a second feature vector.

[0119] The first feature vector and the second feature vector are processed by MLP and then added together to obtain the channel attention map;

[0120] A weighted output feature map is obtained based on the target radar data and the channel attention map.

[0121] The channel attention mechanism first applies global average pooling and global max pooling to the input feature map, generating two feature maps, each with a dimension of C×1×1. These two feature maps are then passed through a shared multilayer perceptron (MLP), where the output channel size of the first layer is C, and the output channel size of the second layer is C / rC. After processing by the MLP, the two feature maps are added together, and a channel attention map is generated using a sigmoid activation function. Finally, the original feature map is multiplied by the channel attention map to generate a weighted output feature map, thereby enhancing important features, suppressing less relevant features, and improving feature representation and model performance.

[0122] Specifically, the weighted features are obtained based on a weighted feature formula, which is:

[0123]

[0124] Where F is the input feature map, AvgPool is the global average pooling operation applied to the feature map, and MaxPool is the global max pooling operation. MLP represents a multilayer perceptron, σ is the activation function, and W0 and W1 are the weights of the first and second fully connected layers, respectively. This represents the vector obtained after global average pooling. This represents the vector obtained after global max pooling.

[0125] Then, key spatial enhancement features are obtained based on the spatial attention module.

[0126] The acquisition of key spatial enhancement features based on the spatial attention module includes:

[0127] The target radar data is subjected to global average pooling and global max pooling based on the spatial attention module to obtain the first pooling feature and the second pooling feature.

[0128] After concatenating the first pooling feature and the second pooling feature, a target spatial attention map is obtained through a 7×7 convolutional layer;

[0129] The target space attention map is normalized to obtain the key space enhancement features.

[0130] Specifically, in this embodiment, the spatial attention mechanism first applies global average pooling and global max pooling to the input feature map, resulting in two pooled feature maps. These feature maps are then concatenated and passed through a 7×7 convolutional layer to generate a spatial attention map. Finally, the attention map is normalized using a sigmoid activation function and multiplied with the input feature map to enhance features at key spatial locations.

[0131] Specifically, the spatial enhancement features are obtained based on the spatial enhancement feature formula, which is:

[0132]

[0133] Where f 7×7 This indicates that a 7×7 core is used for convolution operations.

[0134] Finally, the interference of steel bars in the target radar data is eliminated based on the weighted features, the spatial enhancement features, and the target loss function.

[0135] Before proceeding, the target loss function needs to be obtained, specifically including:

[0136] Obtain the target baseline loss function and the target hybrid loss function, wherein the target hybrid loss function is a hybrid function obtained by fusing MS-SSIM and L1 loss; fuse the target baseline loss function and the target hybrid loss function to obtain the target loss function.

[0137] The total loss function of the target CUT model combines adversarial loss and contrastive loss to jointly optimize the performance of the generator and discriminator. By balancing the contributions of different loss components, the target CUT model can generate more realistic images while preserving structure and details. In this embodiment, the overall target baseline loss function is:

[0138]

[0139]

[0140] In the formula, G is the generator, D is the discriminator, X and Y are the images before and after noise suppression, respectively, and H is the intermediate feature map in the generator; For the total loss, and These are adversarial loss and contrastive loss, respectively, λ X and λ Y These represent the weights within the constructive losses.

[0141] Furthermore, by fusing MS-SSIM and L1 loss, the target hybrid loss function is obtained.

[0142] Specifically, while contrastive loss in baseline models is effective in preserving local consistency in images, relying solely on local feature matching may not yield optimal results for more complex tasks such as style transfer, denoising, or reconstruction. In tasks like rebar signal suppression, contrastive loss can cause the model to overemphasize local features, leading the generator to focus excessively on small regions while neglecting the overall structure of the image. Consequently, the generated image may perform well in terms of local details but exhibit problems with global structure, such as scale imbalance or background disorder.

[0143] In existing technologies, MS-SSIM effectively preserves high-level information during image reconstruction, particularly edges and details. However, it often causes variations in brightness and color. To address the potential brightness and color distortion introduced by MS-SSIM, this embodiment incorporates an absolute error loss (L1 loss) as a compensation mechanism. By combining these two loss functions, the advantages of MS-SSIM in detail preservation can be fully utilized while ensuring the consistency of brightness in the reconstructed image, thereby improving overall image quality.

[0144] Specifically, the target hybrid loss function is:

[0145]

[0146] in, Here, α is the fusion loss function, and α is the weight. Gaussian filtering, The MS-SSIM loss function is... The absolute error loss function is also known as the L1 loss function.

[0147] Specifically, in this embodiment, the MS-SSIM loss function is introduced to ensure the structural and positional consistency between the reconstructed lesion image after removing enhancement noise and the original image with noise. MS-SSIM effectively preserves spatial information and detailed features by measuring image similarity at different scales, resulting in a more realistic and higher-quality image. The MS-SSIM loss function is as follows:

[0148]

[0149] Where μ x and μ y This represents the local mean of the x and y images at the current scale; and σ represents the local variance of the x and y images at the current scale. xy c1 and c2 are the covariances of the x and y images at the current scale; c1 and c2 are stability constants; β m and γ mThese are the weights for brightness similarity and contrast structure similarity at the m-scale, respectively.

[0150] The L1 loss function calculates the mean absolute difference between predicted and true values, demonstrating excellent robustness and sparsity. In tasks such as image denoising and reconstruction, L1 loss prioritizes preserving the main structure and details of the image, unaffected by significant single-pixel outliers, resulting in clearer and more realistic results. Furthermore, L1 loss promotes sparsity by driving feature weights to zero, which not only improves model interpretability but also enhances computational efficiency.

[0151] The L1 loss function is:

[0152]

[0153] Where N represents the total number of samples, x(p) is the pixel value in the generated image, and y(p) is the pixel value in the target image.

[0154] Finally, the target hybrid loss function is fused with the target baseline loss function to obtain the target loss function:

[0155]

[0156] In this way, the interference of steel bars in the target radar data can be eliminated based on the target CUT model to perform image reconstruction and obtain the target radar image.

[0157] After obtaining the target radar image, the process further includes the following steps:

[0158] S300. Analyze the radar image of the target to obtain a cavity condition diagnosis report of the target tunnel.

[0159] The analysis of the target radar image to obtain a cavity condition diagnosis report for the target tunnel includes:

[0160] Based on Fast-RCNN, the target radar image is used to perform hole state diagnosis, and multiple target suggestions are obtained;

[0161] The target suggestion is projected onto the radar image of the corresponding tunnel section to obtain a cavity condition diagnosis report for the target tunnel.

[0162] Specifically, in the study of using Fast-RCNN to diagnose holes in reconstructed ray maps, this embodiment employs a highly efficient and accurate object detection framework—Fast-RCNN. Compared to earlier R-CNN models, this framework offers improvements in both speed and accuracy. The core advantage of Fast-RCNN lies in its integration of region proposal generation and feature extraction into a single, simplified processing flow, thereby significantly optimizing detection efficiency.

[0163] Specifically, in the Fast-RCNN workflow, the input image is first processed by a convolutional neural network (CNN) to generate a feature map containing image feature information. Subsequently, the system identifies region proposals that may contain potential objects in the image and projects these region proposals onto the previously generated feature map. In this step, Region of Interest (ROI) pooling is used to extract features corresponding to these region proposals from the feature map.

[0164] Unlike traditional R-CNN methods, Fast-RCNN performs classification and bounding box regression tasks simultaneously in a single network. This design not only simplifies the processing flow but also significantly improves detection speed. More importantly, Fast-RCNN successfully achieves a significant improvement in detection speed while maintaining high accuracy by reducing redundancy in feature computation for overlapping regions.

[0165] To ensure the accuracy of cavity condition diagnosis, this embodiment constructs a sufficiently reliable and carefully prepared training dataset. Specifically, in this embodiment, cavity size is treated as discrete finite categories, with each category spaced 0.1 meters apart, thus providing a size diagnosis resolution of 0.1 meters, which is sufficient to meet the needs of actual tunnel maintenance work. Furthermore, eight size categories are defined, covering a range from 0.1 meters to 0.8 meters or even larger.

[0166] To construct a robust training dataset and address the data imbalance issue, this embodiment collects the same number of 100 samples for each hole size category. These samples are labeled using the LabelIMG tool, with precise annotations provided for the holes and their sizes in each sample. This approach not only helps balance the dataset but also enhances the model's accuracy and generalization performance in hole size diagnosis. With this carefully prepared dataset and the efficient Fast-RCNN framework, more accurate and reliable hole state diagnosis is expected.

[0167] The following are the experiments and results based on the solution provided in this embodiment:

[0168] The data for this experiment comes from the section of Metro Line 12 from station B to station C, using the ALA GPR system and configured with 800MHz and 500MHz shielded antennas. Figure 8 As shown. The time windows for the 800 MHz and 500 MHz shielded antennas were set to 50 ns and 80 ns, respectively, with a GPR step size of 0.01 m. A total of 2.4 km of radar images were collected. The collected GPR data underwent preprocessing, including typical signal processing techniques such as DC drift removal, gain adjustment, zero-time correction, bandpass filtering, and moving average.

[0169] Numerical simulation data: Due to the lack of real ground data, numerical simulation has become a popular method for studying the GPR response of cavities of different sizes. The finite-difference time-domain (FDTD) method is often used for numerical simulation of ground-penetrating radar wave propagation due to its relative simplicity and directness. This method involves discretizing the electromagnetic (EM) field on a spatial grid and employing time-stepping techniques. Furthermore, when studying the GPR response modes of tunnel lining defects, the structure and corresponding materials of the tunnel lining must be considered. In this embodiment, the DEM-FDTD numerical simulation method is used to generate realistic and universal simulated radar images.

[0170] According to the design standards for subway tunnel lining, the lining typically consists of two layers of steel mesh, with the second layer of concrete lining having a thickness of 0.3–0.4 m [3,40], the first layer of lining having a thickness of 0.2 m, and a layer of surrounding rock. Cavities are located at the interface between the first layer of lining and the surrounding rock, and between the first layer of lining and the secondary lining layer. The cavity size ranges from 0.1 to 0.8 m, with a step size of 0.1 m, and the number of cavities for each size is approximately equal. Furthermore, for each cavity model, a corresponding unreinforced double model is constructed to establish the cavity size model and verify the radar chart reconstruction. This model is identical to the original model in terms of cavity and lining layer structure, but does not include reinforcement. An example of a tunnel lining cavity model is shown below. Figure 9 As shown.

[0171] Data annotation and augmentation: The data annotation process includes three stages: defective tunnel segmentation, rebar suppression and void reconstruction, and void size diagnosis. Before conversion to the COCO dataset format, annotations were performed using LabelImg software following the PASCALVOC format. For the defective tunnel classification and void reconstruction tasks, the tunnel lining sections containing voids need to be labeled. First, enhanced radar images were collected and segmented into 2-meter sections. Then, sections containing voids were identified, extracted, and labeled as "void," while the remaining sections were labeled as "defect-free." Voids typically produce strong reverberant reflection signals, which usually appear as bright areas or high-amplitude waveforms in images, such as... Figure 10 As shown.

[0172] The dataset was organized into four groups: Group A: 137 measured radar images of tunnel lining with cavities; Group B: 490 measured radar images of normal tunnel lining; Group C: 1000 simulated radar images containing cavities and reinforcing bars; Group D: 1000 simulated radar images containing cavities but without reinforcing bars.

[0173] Since the proposed CD-GPR method is unsupervised and does not require paired samples, groups A and C are used as input data for the original domain, while group D is used as training data for the target domain. These radar images require no annotations except for hole size labels, which are applied to group D, forming the third part of the training dataset. Then, all four groups of data (consisting of 1137 hole images and 1490 normal tunnel lining images) are subjected to data augmentation techniques such as random mirroring, Gaussian blur, brightness adjustment, and scaling transformations, generating a total of 7533 defect images. As a result, the final dataset consists of 8370 images containing 10425 annotated holes with corresponding size labels.

[0174] Experimental setup and evaluation metrics:

[0175] (1) Experimental environment:

[0176] The experiments were conducted using a computing system equipped with an NVIDIA GeForce RTX A6000 graphics processing unit (GPU) with 48GB of video random access memory. The operating system used was Windows 10, and Python 3.8 was used as the software development framework. PyTorch library (version 1.1) and CUDA (version 12.2) were employed. The training parameters used in the experimental models are detailed in Table 1. These training parameters were carefully selected to optimize model performance and mitigate overfitting.

[0177] Table 1. Experimental Environment:

[0178]

[0179] (2) Evaluation criteria:

[0180] The performance of the proposed three-stage method is evaluated using different metrics for each stage.

[0181] Phase 1 - Tunnel Lining Defect Classification: For Phase 1 tunnel lining defect classification, standard evaluation metrics are used, including true positives, false positives, true negatives, and false negatives. These metrics are used to calculate the accuracy of the classification model, defined as the proportion of correctly identified defective and non-defective segments. Recall measures the model's ability to correctly identify defective segments, showing the degree to which the model identifies true defects. This is particularly important in tunnel defect detection, as missed defects (false negatives) can lead to safety risks. Precision refers to the proportion of correctly identified defective segments out of all segments identified as defective. This is crucial to ensuring that the model does not over-predict defects (potentially leading to unnecessary repairs). More importantly, the F1-score is applied, which represents the harmonic mean of Precision and Recall, providing a single metric that balances both, as shown in the following formula:

[0182]

[0183] The F1-Score is particularly useful in imbalanced datasets, such as tunnel lining classification, where the number of defective sections may be much smaller than the number of defect-free sections. This metric evaluates overall classification performance by measuring the ratio of correctly classified sections to the total number of sections.

[0184] The second stage – noise suppression and reconstruction – evaluates the reconstruction results using the Structural Similarity Index (SSIM), Peak Signal-to-Noise Ratio (PSNR), and MS-SSIM. SSIM measures the similarity between the ideal noise-free cavity radar image and the reconstructed radar image, focusing on brightness, contrast, and structural information. A higher SSIM value indicates better structural preservation in the reconstructed image, where 0 indicates the two images are independent, and 1 indicates they are identical. Unlike traditional methods based on absolute errors (such as image difference, mean squared error (MSE), and PSNR), SSIM is independent of saturation and distortion because it assumes strong interdependence between adjacent cells. MS-SSIM assesses structural similarity at multiple scales, providing a more comprehensive evaluation of reconstruction quality at different spatial resolutions.

[0185] Phase 3 – Void Condition Estimation: Accuracy and Precision are the primary evaluation metrics. Accuracy measures the proportion of correctly estimated void sizes relative to the actual ground conditions, while precision focuses on the proportion of correctly identified void sizes among all void size predictions made by the model. These metrics help determine the extent to which the model can accurately estimate void sizes, which is crucial for effective tunnel maintenance and intervention.

[0186] Results and performance evaluation:

[0187] Overall performance of CD-GPR in diagnosing voids:

[0188] Table 2 presents the cavity condition diagnosis results of the proposed method. The accuracy of size estimation improves with increasing cavity size, as smaller cavities are more difficult to detect accurately. Overall, the average accuracy and precision of cavity depth and size estimation based on the radar image constructed with the improved CUT are 93.31% and 81.16%, respectively, while the average accuracy and precision of the original radar image are 42.57% and 37.29%, respectively. Figure 11 Figures (b) and (c) show that the hole size was accurately estimated in the reconstructed radar image, while it was incorrectly estimated in the original radar image. Furthermore, the hole identification accuracy was also improved. Figure 11 As shown in Figure (a), although the radar image cross-section was classified as a defect, no cavity was identified. After reconstruction using the method provided in this embodiment, a cavity of 0.7m in size was found, demonstrating the capability of the method provided in this embodiment in detecting small defects. Furthermore, compared to the original radar image, the detection confidence of the reconstructed radar image was improved, as shown in Figure (a). Figure 12 and 13 As shown.

[0189] Table 2. Comprehensive performance in diagnosing cavities of different sizes:

[0190]

[0191] Tunnel Defect Classification Performance:

[0192] Accurate diagnosis relies on correctly identifying defective tunnel linings, making the performance of Stage 1 classification crucial. Precise classification of defective tunnel linings is essential to ensure no voids are missed and to avoid false alarms in defect-free sections. The contributions of optimizations using EMA and transfer learning were also quantified. As shown in Table 3, EfficientNetV2 achieved an accuracy of 90.8%, demonstrating its feasibility as a baseline for tasks involving heavy noise. With the proposed optimizations, the accuracy was further improved to 92.9%, and the increases in recall and F1 score indicate a better balance between classes. Figure 12 The optimized classification process converges faster (marked in red). Figure 13 Examples of classification results are provided, showing that the classification confidence is still above 0.9 despite the small voids in the radar image of the noisy tunnel lining.

[0193] Table 3, Improved EfficientNetV2 performance:

[0194]

[0195]

[0196] Ablation experiments to enhance noise suppression:

[0197] Table 4 and Figure 14 The performance of the second-stage improvement in enhancing noise suppression is presented. Table 4 illustrates the impact of various optimizations on the performance of the baseline CUT model from the perspectives of PSNR, SSIM, and MS-SSIM. The baseline model achieves modest performance, with each optimization progressively improving the results. Introducing the MS-SSIM loss function improves MS-SSIM by addressing multi-scale structural similarity, although the gains in PSNR and SSIM are small. Adding L1 loss along with MS-SSIM achieves a more balanced improvement across all metrics, particularly enhancing PSNR and SSIM. The addition of CBAM significantly increases PSNR and SSIM, but MS-SSIM remains relatively stable. The proposed method, combining CBAM, MS-SSIM, and L1 loss, yields the best overall results, showing consistent growth across all three metrics. This trend highlights the cumulative benefits of integrating feature attention and multi-scale loss functions, improving the reconstruction fidelity and structural integrity of radar images.

[0198] Table 4. Comparison of different optimization modules:

[0199]

[0200] like Figure 14 As shown in (a), (b), (c), and (d) of the diagram, in the original radar image, holes typically appear as strong reverberant reflections of varying lengths, making it difficult to accurately determine the size of the hole from these patterns alone. In the reconstructed radar image, interference noise and signals from the tunnel structure are effectively suppressed, leaving a clearer signal. Importantly, although the hole pattern is inverted compared to the original radar image, key semantic information about the hole, such as depth and location, is preserved. This pattern inversion, combined with noise removal, allows for a more precise diagnosis of hole conditions in the reconstructed radar image, enabling a more accurate assessment of hole size and other characteristics.

[0201] Performance evaluation:

[0202] The Importance of Preliminary Defect Classification in Tunnel Lining:

[0203] The improved EfficientNet V2s method utilizes an improved deep learning model to classify normal and defective tunnel linings. The effectiveness of this method is validated by comparing it with traditional methods that include non-defective samples in the reconstruction model. Figure 14The results of noise suppression on normal lining maps using the baseline CUT method are shown. To address the sample imbalance issue, the ratio of normal lining to defective lining (including all types of voids) was set to 4:5, thus using 777 normal lining sections in the training dataset to enhance their signals. In the original radar map, two layers of reinforcement are clearly visible, and there are no defects in these sections. However, after noise suppression and reconstruction, the resulting radar map exhibits artifact reflections. These artifacts can lead to false positives in void identification, which can effectively exclude normal tunnel sections to avoid unnecessary misclassification during the identification process.

[0204] Performance comparison of cavity morphology reconstruction:

[0205] As mentioned earlier, unsupervised generative models do not require paired data, but can encounter performance limitations when transforming images into drastically different styles. In contrast, supervised generative models excel at producing accurate, task-specific outputs, but their dependence on paired data makes them prone to overfitting and limits their generalization ability. Furthermore, obtaining paired data is often difficult and can only represent a limited number of scenarios. In this embodiment, the performance of the improved CUT model is compared with other commonly used deep learning models, including unsupervised CycleGAN and DCLGAN, as well as the unsupervised baseline CUT.

[0206] Table 5 presents a comparison of the results, highlighting the advantages of unsupervised models in semantic preservation (PSNR and SSIM) when working on smaller training datasets. The proposed model outperforms other models, including CycleGAN, DCLGAN, and CUT, on all metrics. It achieves the highest PSNR (19.832), indicating the best signal quality, and the highest SSIM (0.805), indicating good preservation of hole information. In MS-SSIM, the model scores 0.712, maintaining a high level of structural integrity across multiple scales. CUT performs well but is still inferior to the proposed model, while CycleGAN performs the worst across all metrics. These results demonstrate that the proposed method is highly effective in GPR (Ground Penetrating Radar) reconstruction tasks, especially when dealing with noisy and unpaired data.

[0207] Table 5 shows a performance comparison between the proposed method and state-of-the-art methods:

[0208]

[0209] Figure 15 Example results are shown for each method applied to numerically simulated radar plots. Numerical radar plots are used instead of real radar plots because the actual hole conditions are known, allowing for accurate performance evaluation.

[0210] For larger cavities located between the surrounding rock and the primary lining layer, the improved CUT method successfully reconstructed the diffraction caused by the edges of the two cavities. In contrast, the radar image reconstructed by CycleGAN showed multiple superimposed hyperbolic reflections, DCLGAN produced an oversized cavity pattern, and the cavities reconstructed by CUT were too small.

[0211] For smaller circular voids in the primary lining, hyperbolic reflections are typically produced. The proposed model accurately reconstructs the hyperbolic pattern, such as... Figure 15 As shown. Similar to the results with larger holes, CUT produces patterns that are too small, while DCLGAN produces patterns that are too large. The output of CycleGAN consists of three superimposed hyperbolic reflections.

[0212] These methods successfully suppressed rebar noise and highlighted radar reflections caused by voids, aiding in void detection in noisy environments. However, radar maps reconstructed by other DL models are distorted and do not accurately reflect actual void conditions. In contrast, the improved CUT method can more accurately capture the size and shape of voids, providing detailed support for void diagnosis.

[0213] Performance Boundaries:

[0214] To further evaluate the effectiveness of the proposed method, the model was tested by reconstructing voids of different sizes and depths. Figure 16 The results illustrate the model's ability to reconstruct cavities ranging from 0.2m to 0.8m. Overall, the reconstructed reflection pattern varies with increasing cavity size. Shallower cavities located in the primary lining layer, closer to the reinforcing steel, are subject to more interference, as shown in the figure. Figure 16 Figure (e) shows that cavities located deeper in the surrounding rock are less affected by reinforcing steel, such as... Figure 16 Figure (f) shows the model. This model successfully reconstructed primary lining cavities ranging from 0.2 m to 0.8 m. However, for cavities as small as 0.2 m in the surrounding rock, such as… Figure 16 Figure (b) was misdiagnosed as 0.3m, with a relatively low confidence score, indicating reduced accuracy in these cases. Therefore, this method demonstrates its ability to accurately reconstruct cavities with a minimum size of 0.2m or larger.

[0215] In summary, this embodiment provides a quantitative diagnostic method for GPR tunnel lining voids with noise interference suppression. When quantitatively diagnosing tunnel lining voids, target radar data is first acquired. This target radar data refers to the radar data corresponding to the tunnel cross-section of the defective portion of the target tunnel. Then, image reconstruction is performed based on the target radar data to obtain a target radar image. Finally, the target radar image is analyzed to obtain a void condition diagnostic report for the target tunnel. The noise interference suppression-based quantitative diagnostic method for GPR tunnel lining voids provided in this embodiment, by reconstructing radar data within the tunnel and analyzing the reconstructed radar image, solves the problem in existing technologies where there is no method to accurately quantify the size of voids within the tunnel based on denoised radar images. It accurately reconstructs the radar reflection pattern within the tunnel and precisely identifies the state of voids within the tunnel.

[0216] It should be understood that although the steps in the flowcharts shown in the accompanying drawings are displayed sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of the steps in this invention, and these steps can be executed in other orders. Moreover, at least a portion of the steps in this invention 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 performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0217] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0218] Example 2

[0219] Based on the noise interference suppression quantitative diagnosis method for GPR tunnel lining voids described in Embodiment 1 above, this invention also provides a terminal, the principle block diagram of which is as follows: Figure 17 As shown. The terminal includes a memory 10 and a processor 20. The memory 10 stores a quantitative diagnostic program for GPR tunnel lining voids with noise interference suppression. When the processor 10 executes the computer program, it can perform at least the following steps:

[0220] Acquire target radar data, which is the radar data corresponding to the tunnel cross-section of the damaged part of the target tunnel;

[0221] Image reconstruction is performed based on the target radar data to obtain the target radar image;

[0222] The radar image of the target is analyzed to obtain a cavity condition diagnosis report for the target tunnel.

[0223] Example 3

[0224] The present invention also provides a storage medium storing one or more programs that can be executed by one or more processors to implement the steps of the quantitative diagnosis method for noise interference suppression in GPR tunnel lining described in the above embodiments.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A quantitative diagnostic method for voids in GPR tunnel lining with noise interference suppression, characterized in that, include: Acquire target radar data, which is the radar data corresponding to the tunnel cross-section of the damaged part of the target tunnel; Image reconstruction is performed based on the target radar data to obtain the target radar image; The radar image of the target is analyzed to obtain a cavity condition diagnosis report for the target tunnel; The process of reconstructing the target radar image based on the target radar data includes: Construct a target CUT model, which includes a generator and a discriminator; Based on the target CUT model, the interference of steel bars in the target radar data is eliminated to perform image reconstruction and obtain the target radar image; The construction of the target CUT model includes: A lightweight attention mechanism, CBAM, is introduced in the decoding stage. CBAM consists of a channel attention module and a spatial attention module. Obtain the target baseline loss function and the target hybrid loss function, wherein the target hybrid loss function is a hybrid function obtained by fusing MS-SSIM and L1 loss; The target loss function is obtained by fusing the target baseline loss function and the target hybrid loss function; The target CUT model is constructed based on the CBAM and the target loss function; The generator in the target CUT model adopts an encoder-decoder architecture, consisting of 4 convolutional layers, 9 residual blocks, and 1 additional convolutional layer; The encoder in the generator is transmitted through nine sequentially connected residual blocks, each residual block containing skip connections and a batch normalization layer; The analysis of the target radar image to obtain a cavity condition diagnosis report for the target tunnel includes: Based on Fast-RCNN, the target radar image is used to perform hole state diagnosis, and multiple target suggestions are obtained; The target suggestion is projected onto the radar image of the corresponding tunnel section to obtain a cavity condition diagnosis report for the target tunnel.

2. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 1, characterized in that, The acquisition of target radar data includes: Acquire the initial radar data of the target tunnel, preprocess the initial radar data, and obtain the initial radar image; An initial classification model is constructed, with EfficientNet V2s as the base model and the EMA module as the deep feature extraction module. The initial classification model is pre-trained based on the target dataset to obtain the target classification model; The radar data of the target tunnel is classified based on the target classification model to obtain the target radar data.

3. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 1, characterized in that, The process of eliminating interference from reinforcing bars in the target radar data based on the target CUT model includes: Weighted features are obtained based on the channel attention module; Key spatial enhancement features are obtained based on the spatial attention module; The interference of reinforcing bars in the target radar data is eliminated based on the weighted features, the spatial enhancement features, and the target loss function.

4. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 3, characterized in that, The acquisition of weighted features based on the channel attention module includes: Based on the channel attention module, global average pooling and global max pooling are applied to the target radar data to obtain a first feature vector and a second feature vector. The first feature vector and the second feature vector are processed by MLP and then added together to obtain the channel attention map; A weighted output feature map is obtained based on the target radar data and the channel attention map.

5. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 3, characterized in that, The acquisition of key spatial enhancement features based on the spatial attention module includes: The target radar data is subjected to global average pooling and global max pooling based on the spatial attention module to obtain the first pooling feature and the second pooling feature. After concatenating the first pooling feature and the second pooling feature, through Convolutional layers are used to obtain the attention map in the target space. The target space attention map is normalized to obtain the key space enhancement features.

6. A terminal, characterized in that, The terminal includes: a processor and a storage medium communicatively connected to the processor. The storage medium is adapted to store multiple instructions, and the processor is adapted to call the instructions in the storage medium to execute the steps of the quantitative diagnosis method for noise interference suppression of GPR tunnel lining voids as described in any one of claims 1-5.

7. A storage medium, characterized in that, The storage medium stores one or more programs, which can be executed by one or more processors to implement the steps of the quantitative diagnosis method for noise interference suppression in GPR tunnel lining voids as described in any one of claims 1-5.

Images