GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression

Abstract

The invention discloses a noise interference suppression GPR tunnel lining cavity quantitative diagnosis method. The method comprises the following steps: acquiring target radar data, wherein the target radar data is radar data corresponding to a tunnel section of a disease part of a target tunnel; performing image reconstruction based on the target radar data to obtain a target radar image; and analyzing the target radar image to obtain a cavity condition diagnosis report of the target tunnel. According to the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the invention, the problem that a method for accurately quantifying the size of a cavity in a tunnel based on a de-noised radar image does not exist in the prior art can be solved. The radar reflection pattern in the tunnel is accurately reconstructed, and the state of the cavity in the tunnel is accurately identified.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and particularly to a method for quantitatively diagnosing voids in GPR tunnel linings with noise interference suppression. Background Art

[0002] Diseases such as cracks and voids are likely to occur in the lining of subway tunnels, posing a threat to system safety. Voids are the most risky diseases, often caused by insufficient grouting or soil and water erosion, occurring between lining layers, and different maintenance strategies are required to deal with them. Small voids can be patched 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 detection due to its advantages such as non-destructive and high resolution. The difference in dielectric properties between voids and linings causes radar wave reflection or diffraction, and the reflection signal pattern changes with the size of the void, which is the key to diagnosis. However, the interpretation of GPR data relies on professional knowledge, and the results are variable, affecting the accuracy. The expansion of the subway tunnel network makes manual interpretation time-consuming and laborious, and an efficient and reliable automatic interpretation method is needed.

[0004] In the prior art, to address the challenges, deep learning methods are usually used to suppress the interference of steel bars in GPR data. This research relies on a generative model to reconstruct the radar image to exclude steel bar signals and enhance disease reflections.

[0005] However, diagnosing voids in tunnel linings from denoised radar images still faces challenges, and it is impossible to accurately quantify the size of voids. Therefore, the prior art has limited progress in quantitatively evaluating tunnel lining diseases in GPR data.

[0006] Therefore, the prior art still needs to be improved and enhanced. Summary of the Invention

[0007] In view of the above-mentioned defects of the prior art, a method for quantitatively diagnosing voids in GPR tunnel linings with noise interference suppression is provided, aiming to solve the problem that there is no method in the prior art that can accurately quantify the size of voids in tunnels based on denoised radar images.

[0008] In the first aspect of the present invention, a method for quantitatively diagnosing voids in GPR tunnel linings with noise interference suppression is provided, including:

[0009] Obtain target radar data, where the target radar data is the radar data corresponding to the cross-section of the diseased part of the target tunnel;

[0010] Perform image reconstruction based on the target radar data to obtain a target radar image;

[0011] Analyze the target radar image to obtain a diagnosis report on the void condition of the target tunnel.

[0012] In one implementation, the obtaining of the target radar data includes:

[0013] Obtain the initial radar data of the target tunnel, preprocess the initial radar data to obtain an initial radar image;

[0014] Construct an initial classification model, the basic model of the initial classification model is EfficientNet V2s, and the deep feature extraction module is the EMA module;

[0015] Pre-train the initial classification model based on the target data set to obtain a target classification model;

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

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

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

[0019] Based on the target CUT model, eliminate the interference of steel bars in the target radar data for image reconstruction to obtain the target radar image.

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

[0021] Introduce a lightweight attention mechanism CBAM in the decoding stage, and the CBAM is composed of a channel attention module and a spatial attention module;

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

[0023] Fuse the target baseline loss function and the target hybrid loss function to obtain a target loss function;

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

[0025] Obtain weighted features based on the channel attention module;

[0026] Obtain key spatial enhancement features based on the spatial attention module;

[0027] Based on the weighted features, the spatial enhancement features and the target loss function, eliminate the interference of steel bars in the target radar data.

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

[0029] Applying global average pooling and global max pooling to the target radar data based on the channel attention module to obtain a first feature vector and a second feature vector;

[0030] Performing MLP processing on the first feature vector and the second feature vector and adding them to obtain a channel attention map;

[0031] Obtaining a weighted output feature map based on the target radar data and the channel attention map.

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

[0033] Applying global average pooling and global max pooling to the target radar data based on the spatial attention module to obtain a first pooled feature and a second pooled feature;

[0034] After concatenating the first pooled feature and the second pooled feature, passing them through a 7×7 convolutional layer to obtain a target spatial attention map;

[0035] Performing normalization processing on the target spatial attention map to obtain the key spatial enhancement features.

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

[0037] Performing cavity state diagnosis on the target radar image based on Fast-RCNN to obtain multiple target proposals;

[0038] Projecting the target proposals onto the radar image of the corresponding tunnel section to obtain the diagnosis report on the cavity condition of the target tunnel.

[0039] In a second aspect of the present invention, there is provided a terminal, including: 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 implementing the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression described in any one of the above.

[0040] In a third aspect of the present invention, there is provided a storage medium, wherein the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression described in any one of the above.

[0041] Beneficial effects: Compared with the prior art, the present invention provides a method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression. In the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention, first, target radar data is obtained, and the target radar data is the radar data corresponding to the tunnel cross-section of the damaged part 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 diagnostic report on the void condition of the target tunnel. The method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention can solve the problem in the prior art that there is no method capable of accurately quantifying the size of voids in a tunnel based on a denoised radar image by reconstructing the radar data in the tunnel and analyzing the reconstructed radar image. It can accurately reconstruct the radar reflection pattern in the tunnel and precisely identify the state of voids in the tunnel. Brief Description of the Drawings

[0042] Figure 1 It is a flowchart of an embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0043] Figure 2 It is a general implementation flowchart in the embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0044] Figure 3 It is a schematic diagram of the classification process in the embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0045] Figure 4 It is a structural diagram of the classification model in the embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0046] Figure 5 It is a schematic diagram of the EMA module in the embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0047] Figure 6 It is a schematic diagram of the CUT model structure in the embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0048] Figure 7 It is a diagram of the position and structure of the CBAM in the embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0049] Figure 8 It is a tunnel diagram in the experimental case in the embodiment of the method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by the present invention;

[0050] Figure 9 It is an example diagram of the lining cavity model in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention;

[0051] Figure 10 It is the radar chart in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention;

[0052] Figure 11 It is the experimental result in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention Figure 1 ;

[0053] Figure 12 It is the experimental result in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention Figure 2 ;

[0054] Figure 13 It is the experimental result in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention Figure 3 ;

[0055] Figure 14 It is the experimental result in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention Figure 4 ;

[0056] Figure 15 It is the experimental result in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention Figure 5 ;

[0057] Figure 16 It is the experimental result in the experimental case of the GPR tunnel lining cavity quantitative diagnosis method for noise interference suppression provided by the present invention Figure 6 ;

[0058] Figure 17 It is the structural schematic diagram of the embodiment of the terminal provided by the present invention. Detailed implementation manners

[0059] To make the objectives, technical solutions and effects of the present invention clearer and more definite, the following further describes the present invention in detail with reference to the accompanying drawings and by way of examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] A quantitative diagnosis method for GPR tunnel lining cavities with noise interference suppression provided by the present invention can be applied to a terminal with computing capabilities, and the terminal can execute the quantitative diagnosis method for GPR tunnel lining cavities with noise interference suppression provided by the present invention to diagnose the lining cavities of the tunnel.

[0061] Embodiment 1

[0062] At present, the lining of aging subway tunnels is prone to diseases such as cracks, cavities, voids, and delamination, which pose risks to system safety. Among them, cavities are the most threatening diseases, usually formed due to insufficient grouting or water and soil erosion, and usually occur between the lining layers. Since the required maintenance strategies are different, it is crucial to accurately diagnose the location and size of cavities in the tunnel lining. Small cavities can be solved by local repair (such as directional grouting), while larger cavities may require more extensive intervention, such as relining or structural reinforcement. Therefore, a reliable method for diagnosing the state of tunnel lining cavities is crucial for ensuring effective maintenance and avoiding unnecessary repairs (resulting in subway operation interruptions or neglecting dangerous diseases).

[0063] Ground Penetrating Radar (GPR) has the advantages of non-destructive, high resolution, high efficiency, non-contact measurement, etc., and has been widely used in tunnel lining detection. Compared with the tunnel lining, the air in the cavity will cause differences in dielectric properties, resulting in the reflection or diffraction of radar waves at the disease boundary. As the cavity size increases, the reflected signal presents different patterns in the radar image, which is used as a key indicator for disease diagnosis. However, the interpretation of GPR data is challenging and often depends on professional knowledge, resulting in variability in results and potentially affecting the detection accuracy. With the expansion of the subway tunnel network, manual interpretation of GPR data has become increasingly time-consuming and laborious, and there is an urgent need for an efficient and reliable automatic interpretation method.

[0064] Deep Learning (DL) has become a promising solution for large-scale GPR quantitative interpretation tasks, such as the estimation of steel bar radius, underground facility classification, pavement thickness, and void size. Recent research has developed robust models for identifying tunnel diseases, mainly based on architectures such as Region-based Convolutional Neural Network (RCNN), Residual Network (ResNet), and You Only Look Once (YOLO), which are specifically designed to manage the complex patterns and variations encountered in the tunnel environment. These models have significantly improved the accuracy and reliability of disease identification. However, compared with other aspects of GPR applications, it is more difficult to identify and diagnose targets from tunnel GPR data. Infrastructure in the tunnel, especially steel bars and utilities, will generate diffraction and superimposed reflections of GPR, complicating and distorting the reflections of target diseases, thus forming noise and blurred patterns.

[0065] To address these challenges, many scholars have been focusing on deep learning methods to suppress the interference of steel bars in GPR data. These include the Re2-AttUnet network, equipped with residual and attention modules, which removes enhanced clutter in GPR b-scan images and enhances disease signals. There is also the RCE-GAN method, which combines dilated convolution and attention modules to suppress reinforcement signals at different spacings and improves the accuracy of void detection. In the existing technology, research also relies on generative models that can reconstruct radar images to exclude steel bar signals and strengthen disease reflections. However, there are still challenges in diagnosing tunnel lining voids from denoised radar images:

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

[0067] (2) In scenarios where there are significant differences between diseased tunnels and normal tunnels, the performance of the reconstruction model will decline. Since most tunnels are disease-free, the suppression and reconstruction processes may introduce artifacts or false diseases, which may lead to misjudgment of tunnel lining voids.

[0068] (3) Although the interference suppression model improves the accuracy of void detection, these models still face difficulties in processing distorted or noisy reflection signal features.

[0069] Tunnel lining voids vary in geometry and location, so it is difficult to establish a representative pattern for differentiation. However, these models may produce similar and distorted patterns, thus affecting the accurate assessment of the void state and being unable to accurately quantify the size of the void. Therefore, the progress of the existing technology in quantitatively evaluating tunnel lining diseases using GPR data is still limited.

[0070] Based on this, in this embodiment, a method for quantitatively diagnosing GPR tunnel lining voids with noise interference suppression is provided to solve the above problems. As Figure 1 shown, the method for quantitatively diagnosing GPR tunnel lining voids with noise interference suppression provided in this embodiment includes the steps:

[0071] S100. Obtain target radar data, where the target radar data is the radar data corresponding to the cross-section of the diseased part of the target tunnel.

[0072] The obtaining of the target radar data includes:

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

[0074] Construct an initial classification model, where the base model of the initial classification model is EfficientNet V2s and the deep feature extraction module is the EMA module;

[0075] Pre-train the initial classification model based on the target dataset to obtain a target classification model;

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

[0077] Refer to Figure 2 , in this embodiment, there are a total of three stages. The first stage is to perform scene classification on the target tunnel and extract the disease cross-sections, then perform interference suppression in the second stage, and perform void diagnosis in the third stage.

[0078] Specifically, in the first stage, classify the GPR measurement results of the tunnel lining. In this stage, identify the radar map segments with diseases in the target tunnel and exclude the normal radar map segments to obtain the target radar data. In this way, the number of radar maps that need to be reconstructed can be significantly reduced, the workload can be significantly reduced, and the diagnostic efficiency can be improved.

[0079] First, obtain the initial radar data of the target tunnel and preprocess the initial radar data to obtain an initial radar image. Specifically, collect the GPR data of the target tunnel as the initial radar data, which usually exists in the form of radar images and contains different states (normal and diseased) of the lining. Then clean the collected initial radar data to remove noise and invalid data.

[0080] Then, construct a classification model. In this embodiment, the provided classification model is built on the powerful basis of EfficienetNetV2s and uses the advantages of EMA and transfer learning to improve its performance, which helps to avoid artifacts and pseudo-noises that may be introduced in normal tunnel segments. Specifically, the structure of the target classification model is as Figure 3 shown.

[0081] EfficienetNet is an efficient neural network that solves the limitations of traditional convolutional networks (such as VGG, GoogleNet, and ResNet), which usually expand along a single dimension (such as width, depth, or resolution). By adopting a compound scaling strategy, EfficientNet adjusts all three dimensions simultaneously when expanding the network, thus significantly improving the accuracy and training speed.

[0082] The EfficientNetV2 is a variant, where "s" represents a small network size. EfficientNet V2s has fewer parameters and lower computational complexity, making it suitable for lightweight tasks and applications that require faster inference speeds. Its core structure uses inverted bottleneck convolutions, incorporating techniques such as depthwise separable convolutions, expanded convolutions, and pointwise convolutions, reducing computational complexity while maintaining high accuracy. At the shallower layers, fuse-MBConv is an optimized version of MBConv, which further improves the speed and performance of EfficientNetv2 by combining 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 map scene classification stage. The structure of EfficientNet V2s, including MBConv and fuse-MBConv, is as Figure 4 shown.

[0083] Different from optical images, tunnel GPR data is often affected by significant noise, making it difficult to distinguish weak signals. Attention mechanism modules are widely used to enhance weak feature extraction. Among the 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 groups the channel dimensions into multiple sub-features. By implementing cross-dimensional interactions, it captures pixel-level relationships and promotes richer feature fusion. This enhancement improves the extraction of spatial and channel information, thereby improving the model's ability to understand complex features while also reducing parameter requirements and computational costs.

[0084] As Figure 5 shown, in this embodiment, the deep feature extraction module uses the EMA module. The structure of the EMA module operates on the given input feature map, divides the input feature X into sub-feature groups of RC (number of channels) × H (height) × W (width), and uses three parallel processing paths to extract attention weight descriptors from these grouped feature maps. Two of these paths use 1×1 convolution branches, which apply global average pooling in two spatial dimensions to encode channel information. The third path uses a 3×3 convolution branch, aiming 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 through the Sigmoid function, aiming to capture pixel-level correlations, thereby enhancing the global context information of all pixels.

[0085] In this embodiment, it is crucial to place the EMA module in deeper layers of the EfficientNetV2s because these layers capture more abstract and higher-level features that convey global information and semantic content, enabling them to perform complex classification tasks more effectively. This enhances the 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 and ignore more important information. Additionally, the EMA module improves feature representation but also increases computational cost. Placing it in deeper layers ensures that these computational resources are utilized effectively to process complex global features. Using the EMA module in shallow layers may lead to unnecessary computational work because shallow features provide limited semantic information for more complex tasks.

[0086] Furthermore, due to the lack of ground truth data and the rarity of excavation verification in tunnel inspections, in this embodiment, transfer learning is introduced to improve performance on small datasets. Transfer learning utilizes knowledge obtained from large-scale datasets to solve new tasks. This approach reduces the training time for new tasks, minimizes the need for a large amount of labeled data, improves generalization, and reduces the risk of overfitting.

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

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

[0089] S200. Perform image reconstruction based on the target radar data to obtain a target radar image.

[0090] The performing image reconstruction based on the target radar data to obtain a target radar image includes:

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

[0092] The constructed target CUT model includes:

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

[0094] S212. Fuse MS-SSIM and L1 loss to obtain a target hybrid loss function;

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

[0096] S220. Eliminate the interference of steel bars in the target radar data based on the target CUT model for image reconstruction to obtain the target radar image.

[0097] The elimination of the interference of steel 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 obtaining of weighted features based on the channel attention module includes:

[0100] Apply global average pooling and global maximum pooling to the target radar data based on the channel attention module to obtain a first feature vector and a second feature vector;

[0101] Perform MLP processing on the first feature vector and the second feature vector and add them to obtain a channel attention map;

[0102] Obtain a weighted output feature map 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 obtaining of key spatial enhancement features based on the spatial attention module includes:

[0105] Apply global average pooling and global maximum pooling to the target radar data based on the spatial attention module to obtain a first pooled feature and a second pooled feature;

[0106] After connecting the first pooled feature and the second pooled feature, pass them through a 7×7 convolutional layer to obtain a target spatial attention map;

[0107] Perform normalization processing on the target spatial attention map to obtain the key spatial enhancement features.

[0108] S223. Eliminate the interference of steel bars in the target radar data based on the weighted features, the spatial enhancement features, and the target hybrid loss function.

[0109] Specifically, when the disease tunnel lining diagram, that is, 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 bars in the interference disease section, so as to be able to reconstruct the reflection pattern related to the shape and size of the cavity.

[0110] In this stage, by improving the CUT algorithm and combining CBAM enhanced feature extraction and fusion loss function, fast and stable convergence is achieved. The overall structure of the improved CUT is as Figure 6 shown.

[0111] Specifically, CUT is a new unsupervised method that is based on the concept of translating images from one domain to another without the need for paired examples, which is different from traditional supervised methods. CUT adopts a simpler architecture that only requires one network to perform the conversion from domain A to domain B, which reduces the number of model parameters and thus makes the training converge faster. Its effectiveness has been demonstrated in tasks involving optical, thermal, x-ray, and MRI image translation. However, applying the CUT algorithm to noisy and blurred GPR data poses great challenges. The CUT algorithm lacks back projection, which limits its ability to maintain global structural consistency and makes it prone to semantic inconsistencies during the noise suppression process, further affecting the reliability of disease detection in tunnel lining analysis.

[0112] In the prior art, CUT is an unsupervised image translation method based on contrast learning. Its core framework is a generative adversarial network composed 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 the enhanced signal and the non-enhanced signal on unpaired radar images to achieve the suppression of steel bar noise.

[0113] Specifically, the generator in the target CUT model adopts an encoder-decoder architecture, which consists of 4 convolutional layers, 9 residual blocks, and 1 additional convolutional layer. The encoder is responsible for extracting high-level features from the input image and then passing them through 9 sequentially connected residual blocks. Each residual block contains a skip connection and a batch normalization layer, which helps to accelerate training and enhance stability. These residual blocks capture the non-linear relationships between domains, learn the mapping between the input image and the target domain, and at the same time alleviate the problem of gradient disappearance. Finally, the decoder gradually increases the data dimension by using transposed convolutional layers and gradually restores the feature map into the output image, thereby generating the final image in the target domain.

[0114] The discriminator in the target CUT model uses the PatchGAN architecture and consists of 5 convolutional layers. After each convolutional operation, instance normalization is applied, followed by the non-linear activation function LeakyReLU. The last 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 mean of the feature map is 0 and the variance is 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 representation ability.

[0115] Specifically, since there is no inverse transformation in CUT, image feature information is often lost during the encoding and decoding process, which limits the similarity and quality between the generated image and the original image and introduces irrelevant pseudo-signals in regions unrelated to the task. To solve this problem, in this embodiment, while maintaining the processing speed and model size, a lightweight attention mechanism CBAM is introduced in the decoding stage to preserve invariant information. CBAM consists of a channel attention module and a spatial attention module, which jointly enhance the semantic information extraction in the channel and spatial dimensions and improve the feature representation ability. Specifically, the position and structure of CBAM are as Figure 7 shown.

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

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

[0118] Applying global average pooling and global max pooling to the target radar data based on the channel attention module to obtain a first feature vector and a second feature vector;

[0119] Performing MLP processing on the first feature vector and the second feature vector and adding them to obtain a channel attention map;

[0120] Obtaining a weighted output feature map 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 to generate two feature maps, each with dimensions C×1×1. These two feature maps are then passed through a shared multi-layer perceptron (MLP), where the output channel size of the first layer is C / r and the output channel size of the second layer is C. After passing through the MLP, the two feature maps are added together, and a channel attention map is generated through the 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 feature is obtained based on the weighted feature formula, and the weighted feature formula is:

[0123]

[0124] where F is the input feature map, AvgPool is the global average pooling operation applied to the feature map, MaxPool is the global max pooling operation, MLP represents the multi-layer perceptron, σ is the activation function, W 0 and W 1 are the weights of the first and second fully connected layers respectively, represents the vector obtained after global average pooling, represents the vector obtained after global max pooling.

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

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

[0127] Applying global average pooling and global max pooling to the target radar data based on the spatial attention module to obtain a first pooled feature and a second pooled feature;

[0128] After connecting the first pooled feature and the second pooled feature, passing them through a 7×7 convolutional layer to obtain a target spatial attention map;

[0129] Normalizing the target spatial attention map to obtain the key spatial 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 to obtain two pooled feature maps. Then, these feature maps are connected and passed through a 7×7 convolutional layer to generate a spatial attention map. Finally, the attention map is normalized using the Sigmoid activation function and multiplied by the input feature map to enhance the features at key spatial positions.

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

[0132]

[0133] where f 7×7 represents performing a convolution operation using a 7×7 kernel.

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

[0135] Before that, it is necessary to obtain the target loss function, which specifically includes:

[0136] Obtain a target baseline loss function and a target hybrid loss function. 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 the discriminator. By balancing the contributions of different loss components, the target CUT model can generate more realistic images while retaining structure and details. In this embodiment, the total target baseline loss function is:

[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; is the total loss, and are the adversarial loss and the contrastive loss respectively, and λ X and λ Y are the weights in the constructive loss respectively.

[0141] Furthermore, fuse MS - SSIM and L1 loss to obtain the target hybrid loss function.

[0142] Specifically, the contrast loss in the baseline model is effective in maintaining local consistency of images. However, for more complex tasks such as style transfer, denoising, or reconstruction, relying solely on local feature matching may not produce optimal results. In tasks such as steel bar signal suppression, the contrast loss may cause the model to overly focus on local features, leading the generator to overly focus on small regions and neglect the overall structure of the image. Therefore, the generated images may perform well in terms of local details but exhibit problems in terms of global structure, such as unbalanced proportions or disordered backgrounds.

[0143] In the prior art, MS-SSIM effectively preserves high-level information during the image reconstruction process, especially preserving edges and details. However, it tends to cause changes in brightness and color. To address the potential brightness and color distortion brought by MS-SSIM, in this embodiment, the absolute error loss (L1 loss) is incorporated into the 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 the brightness of the reconstructed image, thereby improving the overall image quality.

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

[0145]

[0146] Where is the fusion loss function, α is the weight, is the Gaussian filter, is the MS-SSIM loss function, is the absolute error loss function, that is, the L1 loss function.

[0147] Specifically, in this embodiment, the MS-SSIM loss function is introduced to ensure the consistency in structure and position between the reconstructed disease image after removing the enhancement noise and the original noisy image. By measuring the image similarity at different scales, MS-SSIM effectively preserves the spatial information and detail features, generating a more realistic and higher-quality image. The MS-SSIM loss function is:

[0148]

[0149] Where μ x and μ y represent the local means of images x and y at the current scale; and represent the local variances of images x and y at the current scale; σ xy is the covariance of images x and y at the current scale; c 1 and c 2 are stability constants; β m and γm They are the weights of luminance similarity and contrast structure similarity on the m scale respectively.

[0150] The L1 loss function calculates the mean absolute difference between the predicted value and the true value, demonstrating excellent robustness and sparsity. In tasks such as image denoising and reconstruction, the L1 loss will preferentially preserve the main structure and details of the image and will not be significantly affected by single-pixel outliers, thus obtaining clearer and more realistic results. In addition, the L1 loss promotes sparsity by driving the feature weights to zero, which not only improves the interpretability of the model but also improves the 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, fusing the target hybrid loss function and the target baseline loss function yields 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 for image reconstruction, and the target radar image can be obtained.

[0157] After obtaining the target radar image, the following steps are further included:

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

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

[0160] Based on Fast-RCNN, diagnose the cavity state of the target radar image to obtain multiple target proposals;

[0161] Project the target proposals onto the radar image of the corresponding tunnel section to obtain a diagnostic report on the cavity condition of the target tunnel.

[0162] Specifically, in the research on diagnosing the cavity state using Fast-RCNN for the reconstructed radiographic images, this embodiment adopts an efficient and accurate object detection framework - Fast-RCNN. Compared with the early R-CNN model, this framework has improved both in speed and accuracy. The core advantage of Fast-RCNN is that it integrates the two steps of region proposal generation and feature extraction into a single, more simplified processing flow, thus significantly optimizing the detection efficiency.

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

[0164] Different from the traditional R-CNN method, Fast-RCNN performs both the classification task and the bounding box regression task in a single network. This design not only simplifies the processing flow but also significantly improves the detection speed. More importantly, Fast-RCNN reduces the redundancy in calculating features for overlapping regions, successfully achieving a significant improvement in detection speed while maintaining a high accuracy rate.

[0165] To ensure the accuracy of cavity state diagnosis, in this embodiment, a sufficiently reliable and well-prepared training dataset is constructed. Specifically, in this embodiment, the cavity size is regarded as discrete finite categories, and the interval for each category is set to 0.1 meters, thus providing a size diagnosis resolution of 0.1 meters, which is accurate enough to meet the requirements of actual tunnel maintenance work. And eight size categories are defined, covering a range from 0.1 meters to 0.8 meters and even larger.

[0166] To construct a robust training dataset and address the problem of data imbalance, in this embodiment, the same number of 100 samples are collected for each cavity size category. The annotation work for these samples is completed through the LabelIMG tool, and at the same time, the cavities and their sizes in each sample are accurately annotated. This method not only helps to balance the dataset but also enhances the accuracy and generalization performance of the model in the cavity size diagnosis task. With such a well-prepared dataset and the efficient Fast-RCNN framework, it is expected to achieve a more accurate and reliable cavity state diagnosis.

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

[0168] The data for this experiment was collected from the section of the 12th line of Subway A from Station B to Station C. The ALA GPR system was used, equipped with 800 and 500 MHz shielded antennas, as Figure 8 shown. The time windows for the 800 and 500 MHz shielded antennas were set to 50 and 80 ns respectively, and the GPR step size was 0.01 m. A total of 2.4 km of radar images were collected. The collected GPR data was preprocessed, which included typical signal processing techniques such as dc drift removal, gain adjustment, zero-time correction, band-pass filtering, and moving average.

[0169] Numerical simulation data: Due to the lack of ground truth data, numerical simulation has become a popular method for studying the GPR responses of cavities of different sizes. The finite-difference time-domain (FDTD) method is commonly used for numerically simulating the propagation of ground-penetrating radar waves because of its relative simplicity and directness. This method involves discretizing the electromagnetic (EM) field on a spatial grid and using a time-stepping technique. In addition, when studying the GPR response patterns of tunnel lining diseases, 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 general simulated radar images.

[0170] According to the subway tunnel lining design standard, the lining usually consists of two layers of steel mesh. The thickness of the secondary concrete lining is 0.3 - 0.4 m [3,40], the thickness of the primary lining is 0.2 m, and there is also a layer of surrounding rock. The cavities are located at the interface between the primary lining and the surrounding rock, and between the primary lining layer 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 of each size is approximately equal. In addition, for each cavity model, a corresponding non-reinforced double model is constructed to establish the cavity size model and verify the radar image reconstruction. This model is the same as the original model in terms of the cavity and lining layer structure, but does not contain steel bars. An example of the tunnel lining cavity model is Figure 9 shown.

[0171] Data annotation and enhancement: The data annotation process includes three stages: diseased tunnel segmentation, rebar suppression and cavity reconstruction, and cavity size diagnosis. Before converting to the COCO dataset format, annotation is performed using the LabelImg software following the PASCAL VOC format. For the tasks of classifying diseased tunnels and cavity reconstruction, the tunnel lining parts containing cavities need to be marked. First, the collected and simulated enhanced radar images are segmented into 2-meter sections. Subsequently, the parts containing cavities are identified, extracted, and marked as "cavity", and the remaining parts are marked as "disease-free". Cavities usually generate strong reverberation reflection signals, which are usually shown as bright areas or high-amplitude waveforms in the images, as Figure 10 shown.

[0172] The dataset is organized into four groups: Group A: 137 measured radar images of diseased tunnel linings with voids; Group B: 490 measured radar images of normal tunnel linings; Group C: 1000 simulated radar images with voids and steel bars; Group D: 1000 simulated radar images with voids but without steel bars.

[0173] Since the proposed CD-GPR method is unsupervised and does not require paired samples, Group A and Group C are used as input data for the source domain, while Group D is used as training data for the target domain. These radar images do not require any annotations except for the void size labels, which are applied to Group D to form the third part of the training dataset. Then, all four groups of data (consisting of 1137 void images and 1490 normal tunnel lining images) are subjected to data augmentation techniques such as random mirroring, Gaussian blur, brightness adjustment, and scaling transformation, generating a total of 7533 diseased images. As a result, the final dataset consists of 8370 images, including 10425 annotated voids with corresponding size labels.

[0174] Experimental settings and evaluation metrics:

[0175] (1) Experimental environment:

[0176] The experiments were conducted using a computing system equipped with an NVIDIA GeForce RTXA6000 graphics processing unit with 48GB of video random access memory. The operating system used for the experiments was Windows 10, and Python 3.8 was used as the software development framework. The experiments were carried out using the PyTorch library (version 1.1) and CUDA (version 12.2). The training parameters used for the experimental model are detailed in Table 1. These training parameters were carefully selected to optimize the 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] The first stage - Tunnel lining disease classification: For the classification of tunnel lining diseases in the first stage, 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, which is defined as the ratio of correctly identified disease sections and non-disease sections. Recall measures the ability of the model to correctly identify the diseased part and shows the degree to which the model can identify real diseases. This is particularly important in tunnel disease detection because missing diseases (false negatives) may pose safety risks. Precision refers to the proportion of correctly identified disease sections among all sections identified as diseased. This is crucial to ensure that the model does not over-predict diseases (which may lead 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 the two, 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 sections with diseases may be much smaller than the number of sections without diseases. This metric evaluates the overall performance of the classification by measuring the ratio of correctly classified sections to the total number of sections.

[0184] The second stage - Noise suppression and reconstruction: The structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and MS-SSIM are used to evaluate the reconstruction results. SSIM measures the similarity between the ideal noiseless cavity radar image and the reconstructed radar image, focusing on luminance, contrast, and structural information. The higher the SSIM value, the better the structure of the reconstructed image is retained, where 0 indicates that the two images are independent and 1 indicates that the two images are identical. Different from traditional methods based on absolute error (such as image difference, mean square error (MSE), peak signal-to-noise ratio (PSNR)), SSIM is independent of saturation and distortion because it assumes that adjacent units have strong interdependencies. MS-SSIM evaluates the structural similarity at multiple scales, providing a more comprehensive assessment of the reconstruction quality at different spatial resolutions.

[0185] The third stage - Void condition estimation: Accuracy and Precision are the main evaluation metrics. Accuracy measures the ratio of correctly estimated void sizes to the ground truth, 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 cavities:

[0188] Table 2 presents the cavity state diagnosis results of the method in this paper. The accuracy of size estimation improves with the increase in cavity size because smaller cavities are more difficult to detect accurately. Overall, the average accuracy and precision of cavity depth and size estimation based on the radar chart constructed with the improved CUT are 93.31% and 81.16% respectively, while those of the original radar chart are 42.57% and 37.29% respectively. Figure 11 Figures (b) and (c) in [reference] show that the cavity size is accurately estimated in the reconstructed radar chart, while the cavity size is misestimated in the original radar chart. In addition, the recognition accuracy of cavities has also been improved. As Figure 11 shown in Figure (a) in [reference], although the radar chart cross - section is classified as a disease, the cavity is not recognized. After reconstruction based on the method provided in this embodiment, a cavity with a size of 0.7 m is found, demonstrating the ability of the method provided in this embodiment in detecting small diseases. And compared with the original radar chart, the detection confidence of the reconstructed radar chart has been improved, as Figure 12 and 13 shown.

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

[0190]

[0191] Tunnel disease classification performance:

[0192] Accurate diagnosis depends on correctly identifying diseased tunnel linings, which makes the performance of stage 1 classification crucial. Precise classification of diseased tunnel linings is essential to ensure that cavities are not missed and false alarms in disease - free sections are avoided. The contributions of using EMA and transfer learning for optimization are also quantified. As shown in Table 3, EfficientNetV2 achieves an accuracy of 90.8%, demonstrating its feasibility as a baseline for tasks involving heavy noise. Through the proposed optimization, the accuracy is further improved to 92.9%, and the increase in recall rate and F1 - score indicates a better balance between classes. Figure 12 shows that the optimized classification process converges faster (red mark). Figure 13 provides an example of the classification results. Although the cavity is very subtle in the radar chart of the noisy tunnel lining, the classification confidence is still above 0.9.

[0193] Table 3, Performance of the improved EfficientNetV2:

[0194]

[0195]

[0196] Ablation experiment on enhanced noise suppression:

[0197] Table 4 and Figure 14 present the performance of the improved enhanced noise suppression in the second stage. 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 achieved moderate performance, and each optimization gradually improved the results. Introducing the MS-SSIM loss function improved MS-SSIM by addressing multi-scale structural similarity, although the gains in PSNR and SSIM were small. Adding the L1 loss together with MS-SSIM enabled more balanced improvements in all metrics, especially increasing PSNR and SSIM. The addition of CBAM significantly increased PSNR and SSIM, but MS-SSIM remained relatively stable. The proposed method combining CBAM, MS-SSIM, and L1 loss generally produced the best results, showing consistent growth in 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 maps.

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

[0199]

[0200] As Figure 14 shown in (a), (b), (c), and (d) therein, in the original radar map, voids usually appear as strong reverberation reflections of different lengths, making it difficult to accurately determine the size of voids solely from these patterns. In the reconstructed radar map, both the interference noise and signals from the tunnel structure are effectively suppressed, leaving a clearer signal. Importantly, although the void patterns are inverted compared to the original radar map, the key semantic information about the voids (such as depth and location) is retained. This combination of pattern inversion and noise removal enables more precise diagnosis of void conditions in the reconstructed radar map, thus allowing for more accurate assessment of void size and other features.

[0201] Performance evaluation:

[0202] Importance of preliminary disease classification for tunnel linings:

[0203] The improved EfficientNet V2s method uses an improved deep learning model to classify normal and diseased tunnel linings. The effectiveness of this method is verified by comparing it with traditional methods that include non-diseased samples in the reconstruction model. Figure 14Shows the results of noise suppression on normal lining diagrams using the baseline CUT method. To address the problem of sample imbalance, the ratio of normal linings to defective linings (including all types of voids) was set to 4:5, so that 777 normal lining sections were used in the training dataset to enhance their signals. In the original radar diagrams, two layers of steel bars can be clearly seen, and there are no defects in these sections. However, after noise suppression and reconstruction, the resulting radar diagrams exhibit artifact reflections. These artifacts may lead to false positives in void identification, which can effectively exclude normal tunnel sections to avoid unnecessary misjudgments during the identification process.

[0204] Performance comparison of void shape reconstruction:

[0205] As mentioned above, unsupervised generative models do not require paired data, but may encounter performance limitations when converting images into images with very different styles. In contrast, supervised generative models are good at producing accurate, task-specific outputs, but their dependence on paired data makes them prone to overfitting and limits their generalization ability. In addition, obtaining paired data is usually somewhat difficult, and it can only represent limited scenarios. In this embodiment, the performance of the improved CUT model is compared with that of other commonly used deep learning models, including the unsupervised CycleGAN and DCLGAN, and the unsupervised baseline CUT.

[0206] Table 5 presents the comparison of the results, highlighting the advantages of unsupervised models in terms of semantic preservation (PSNR and SSIM) when working on a small training dataset. The proposed model outperforms other models in all metrics, including CycleGAN, DCLGAN, and CUT. It has the highest PSNR (19.832), indicating the best signal quality; the highest SSIM (0.805), indicating better preservation of void information. In MS-SSIM, the model scored 0.712, maintaining a high level of structural integrity at multiple scales. CUT performed well but was still lower than the proposed model, while CycleGAN performed the worst among all metrics. These results show that the proposed method is very effective in GPR (ground penetrating radar) reconstruction tasks, especially when dealing with noisy data and unpaired data.

[0207] Table 5, Performance comparison between the proposed method and state-of-the-art methods:

[0208]

[0209] Figure 15 Shows example results of each method applied to numerical simulation radar diagrams. Numerical radar diagrams are used instead of real radar diagrams because the actual void conditions are known, allowing for accurate performance evaluation.

[0210] For the large voids 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 voids. In contrast, the radar images reconstructed by CycleGAN showed multiple superimposed hyperbolic reflections, DCLGAN produced an oversized void pattern, and the void reconstructed by CUT was too small.

[0211] For the small circular voids in the primary lining, hyperbolic reflections are usually generated, and the proposed model accurately reconstructed the hyperbolic pattern, as Figure 15 shown. Similar to the results of the large voids, the pattern produced by CUT was too small, while the pattern produced by DCLGAN was too large. The output of CycleGAN consisted of three superimposed hyperbolic reflections.

[0212] All these methods successfully suppressed the steel bar noise, highlighted the radar reflections caused by the voids, and contributed to void detection in a noisy environment. However, the radar images reconstructed by other DL models were distorted and could not accurately reflect the actual void conditions. In contrast, the improved CUT method could capture the size and shape of the voids more precisely, providing detailed support for void diagnosis.

[0213] Performance Boundary:

[0214] To further evaluate the effectiveness of the proposed method, the model was tested for reconstructing voids of different sizes and depths. Figure 16 The results shown illustrate the ability of the model to reconstruct voids in the range of 0.2 m to 0.8 m. Generally, the reflection patterns reconstructed by the model changed with the increase in void size. The shallower voids located in the primary lining layer were closer to the steel bars and were more disturbed, as shown in Figure 16 Figure (e), while the deeper voids located in the surrounding rock were less affected by the steel bars, as shown in Figure 16 Figure (f). The model successfully reconstructed the voids in the primary lining in the range of 0.2 m to 0.8 m. However, for the 0.2 m void in the surrounding rock, as shown in Figure 16 Figure (b) was misdiagnosed as 0.3 m with a relatively low confidence score, indicating a reduction in accuracy in these cases. Therefore, the method demonstrated the ability to accurately reconstruct voids with a minimum size of 0.2 m or larger.

[0215] In summary, this embodiment provides a method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression. When quantitatively diagnosing voids in a tunnel lining, first, target radar data is obtained, where the target radar data is the radar data corresponding to the cross-section of the damaged part 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 diagnostic report on the void condition of the target tunnel. The method for quantitatively diagnosing voids in a GPR tunnel lining with noise interference suppression provided by this embodiment can solve the problem in the prior art that there is no method capable of accurately quantifying the size of voids in a tunnel based on a denoised radar image by reconstructing the radar data in the tunnel and analyzing the reconstructed radar image. It can accurately reconstruct the radar reflection pattern in the tunnel and precisely identify the state of voids in the tunnel.

[0216] It should be understood that although the steps in the flowchart given in the accompanying drawings of the present invention are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of the steps in the present invention has no strict order limitation, and these steps can be executed in other orders. Moreover, at least a part of the steps of the present invention may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential either, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0217] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. 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), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0218] Embodiment 2

[0219] Based on the method for quantitatively diagnosing voids in GPR tunnel linings with noise interference suppression described in the above Embodiment 1, the present invention also provides a terminal, and its principle block diagram can be as Figure 17 shown. The terminal includes a memory 10 and a processor 20. A program for quantitatively diagnosing voids in GPR tunnel linings with noise interference suppression is stored in the memory 10. When the processor 10 executes the computer program, it can at least implement the following steps:

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

[0221] Based on the target radar data, perform image reconstruction to obtain a target radar image;

[0222] Analyze the target radar image to obtain a diagnosis report on the void condition of the target tunnel.

[0223] Embodiment 3

[0224] The present invention also provides a storage medium that stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the method for quantitatively diagnosing voids in GPR tunnel linings with noise interference suppression 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 are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. And these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression, characterized in that: include: Acquire target radar data, where the target radar data is radar data corresponding to a tunnel section of a defective portion of the target tunnel; Performing image reconstruction based on the target radar data to obtain a target radar image; The target radar image is analyzed to obtain a cavity condition diagnosis report of the target tunnel.

2. The noise interference suppressed GPR tunnel lining void quantitative diagnosis method according to claim 1, characterized in that: The acquiring of target radar data comprises: Acquiring initial radar data of the target tunnel, and preprocessing the initial radar data to obtain an initial radar image; Constructing an initial classification model, wherein the basic model of the initial classification model is EfficientNet V2s, and the deep feature extraction module is the EMA module; Pre-training the initial classification model based on the target data set to obtain a 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 noise interference suppressed GPR tunnel lining void quantitative diagnosis method according to claim 1, characterized in that: The step of reconstructing an image based on the target radar data to obtain a target radar image includes: Constructing a target CUT model, wherein the target CUT model includes a generator and a discriminator; Based on the target CUT model, interference of steel bars in the target radar data is eliminated to perform image reconstruction, thereby obtaining the target radar image.

4. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 3 is characterized in that: The constructing of the target CUT model includes: A lightweight attention mechanism CBAM is introduced in the decoding stage, which consists of a channel attention module and a spatial attention module; Obtain a target baseline loss function and a target mixed loss function, where the target mixed loss function is a mixed 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 mixed loss function; The target CUT model is constructed based on the CBAM and the target loss function.

5. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 4 is characterized in that: Eliminating the interference of steel bars in the target radar data based on the target CUT model includes: Acquire weighted features based on the channel attention module; Acquire key spatial enhancement features based on the spatial attention module; 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.

6. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 5 is characterized in that: The obtaining weighted features based on the channel attention module includes: Applying global average pooling and global maximum pooling to the target radar data based on the channel attention module to obtain a first eigenvector and a second eigenvector; Performing MLP processing on the first eigenvector and the second eigenvector and adding them to obtain a channel attention map; A weighted output feature map is obtained based on the target radar data and the channel attention map.

7. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 5, characterized in that: The obtaining of key spatial enhancement features based on the spatial attention module includes: Applying global average pooling and global maximum pooling to the target radar data based on the spatial attention module to obtain a first pooling feature and a second pooling feature; After connecting the first pooled features and the second pooled features, a 7×7 convolution layer is passed to obtain a target space attention map; The target space attention map is normalized to obtain the key space enhancement feature.

8. The method for quantitative diagnosis of GPR tunnel lining voids with noise interference suppression according to claim 1, characterized in that: The step of analyzing the target radar image to obtain a cavity condition diagnosis report of the target tunnel includes: Based on Fast-RCNN, a cavity state diagnosis is performed on the target radar image to obtain multiple target suggestions; The target suggestion is projected onto a radar image of a corresponding tunnel section to obtain a cavity condition diagnosis report of the target tunnel.

9. A terminal, characterized in that: The terminal includes: a processor, a storage medium communicatively connected to the processor, the storage medium being suitable for storing a plurality of instructions, and the processor being suitable for calling the instructions in the storage medium to execute the steps of the quantitative diagnosis method for GPR tunnel lining voids with noise interference suppression as described in any one of claims 1 to 7.

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

Images