Ground penetrating radar signal intelligent denoising and enhancement method and system

Abstract

The application discloses a kind of ground penetrating radar signal intelligent denoising and enhancement method and system, it is related to ground penetrating radar data processing technical field, including the following steps: using forward simulation software to generate synthetic radar profile, and adding simulated noise in radar profile, generate noisy profile, construct training dataset;Build and train DRSN-ATT-UNET network model, the network model is with U-Net as basic architecture, and integrated depth residual shrinkage module in encoder, integrated attention gate module in jump connection;Load actual ground penetrating radar profile data to be processed, and carry out standardization preprocessing;Profile data after preprocessing is input into the network model that has been trained, and output denoising and enhanced radar profile;After model output is carried out anti-standardization post-processing, output high signal-to-noise ratio radar image.The application can effectively suppress ground penetrating radar noise under complex environment such as mine, significantly improve data quality, and provide reliable basis for accurate geological interpretation.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar data processing technology, specifically to a method and system for intelligent denoising and enhancement of ground-penetrating radar signals. Background Technology

[0002] Ground-penetrating radar (GPR) is a highly efficient, non-destructive testing tool widely used in shallow surface exploration, engineering geology, and mine safety. However, in complex environments such as underground mine spaces, radar signals are easily contaminated by internal system noise, random environmental noise, and strong coherent noise and impulse noise caused by metal components and irregular interfaces. These noises significantly reduce the signal-to-noise ratio and resolution of radar images, masking effective weak reflection signals and posing significant challenges to geological interpretation and disaster identification.

[0003] Traditional denoising methods (such as frequency domain filtering, FK filtering, median filtering, etc.) are usually based on simple linear or statistical assumptions, which have inherent limitations when dealing with complex nonlinear and non-stationary noise: denoising is not thorough, it is difficult to effectively distinguish noise from effective signals, especially when frequency bands overlap, which can easily lead to damage to effective signals; adaptability is poor, the algorithm parameters need to be manually adjusted based on experience, and there is a lack of adaptability; the ability to recover weak signals is insufficient, and it is difficult to separate weak reflection signals from deep or small geological bodies from strong noise backgrounds.

[0004] In recent years, although some studies have attempted to apply deep learning to seismic data denoising, ground-penetrating radar (GPR) signals have higher frequencies, more complex waveform characteristics, and drastically different noise patterns, making direct application of seismic denoising models ineffective. Therefore, there is an urgent need for an intelligent processing method that can adaptively and faithfully suppress complex noise from GPR and effectively recover and enhance weak signals. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for intelligent denoising and enhancement of ground penetrating radar signals, which has better denoising effect, higher signal fidelity, and can adaptively recover weak signals, significantly improving the signal-to-noise ratio and resolution of ground penetrating radar data in complex environments such as mines.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent denoising and enhancement method for ground-penetrating radar signals, comprising the following steps:

[0008] S1: Use forward modeling software to generate synthetic radar profiles, add simulated noise to the radar profiles to generate noisy profiles, and build a training dataset.

[0009] S2: Construct the DRSN-ATT-UNET network model, which is based on U-Net and integrates a deep residual shrinkage module in the encoder and an attention gating module in the skip connections; train the DRSN-ATT-UNET network model using a composite loss function.

[0010] S3: Load the actual ground-penetrating radar profile data to be processed and perform standardized preprocessing; input the preprocessed profile data into the trained DRSN-ATT-UNET network model for forward inference, and the model output is the denoised and enhanced radar profile; perform inverse standardized postprocessing on the model output and output the final high signal-to-noise ratio radar image.

[0011] Further, step S1 specifically includes:

[0012] 5,000 synthetic radar profiles containing cavities, faults, and bedding of different sizes were generated using ground-penetrating radar forward modeling software as clean signals.

[0013] Gaussian white noise, pink noise, coherent noise, and random impulse noise are manually added to the clean signal to simulate noise interference in the complex environment of a mine, forming a noisy-clean signal pair and constructing a training set.

[0014] Furthermore, the structure of the DRSN-ATT-UNET network model includes:

[0015] The encoder consists of four downsampling units. Each downsampling unit includes a convolutional block, a depth residual shrinking module, and a max pooling layer. The depth residual shrinking module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features. The convolutional block consists of a convolutional layer, batch normalization, and a ReLU activation function.

[0016] Decoder: Composed of four upsampling units, each upsampling unit includes a transposed convolutional layer, an attention gating module and a convolutional block; the feature maps of the encoder and decoder are spliced ​​together through skip connections, and the attention gating module is embedded in the skip connections to gradually restore spatial resolution and focus on important signal features;

[0017] Bottleneck layer: Located between the encoder and decoder, it contains a non-local attention block to capture long-range dependencies across the entire profile, enhancing the model's understanding of the global structure.

[0018] Furthermore, the depth residual shrinkage module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features, specifically including:

[0019] Let the input feature map be Where H is the height, W is the width, and C is the number of channels, residual feature extraction is performed on the input feature map. The extraction process of residual feature R is expressed as follows:

[0020]

[0021] in For convolution kernel weights, Indicates the convolution operation;

[0022] Adaptive soft thresholding is performed by global average pooling on the input feature map X to obtain channel-level statistical features. :

[0023]

[0024] in c For channel indexing;

[0025] Channel attention weights are generated using two fully connected (FC) layers:

[0026]

[0027] in For the weights of the fully connected layer, For the Sigmoid function;

[0028] Calculate the soft threshold for each channel :

[0029]

[0030] Based on the soft threshold of each channel By applying soft thresholding to each element of the residual feature R, we obtain the residual shrinkage feature:

[0031]

[0032] The residual shrinkage feature and the input feature map are added element-wise, and the denoised feature map is output. .

[0033] Furthermore, the attention gating module is embedded in the skip connection to progressively restore spatial resolution and focus on important signal features, specifically including:

[0034] Feature interaction: The feature map p output by the encoder is convolved with the gate signal g of the decoder, and then added element by element to generate intermediate features;

[0035] Weight calculation: ReLU activation, convolution and Sigmoid activation steps are performed on the intermediate features to generate a spatial attention weight map A, which outputs weight values ​​of 0-1;

[0036] Feature weighting: Multiply the attention weight A with the feature map p output by the original encoder, and output p'=Ap.

[0037] Furthermore, the non-local attention block specifically includes:

[0038] Feature Mapping and Embedding: Let the input feature map of the bottleneck layer be... The query map Q, key map K, and value map V are generated through three independent 1×1 convolutional layers, respectively:

[0039]

[0040] in These are the convolution weights;

[0041] Similarity matrix calculation: Flatten the query mapping Q and key mapping K into two-dimensional matrices with dimensions respectively. and Calculate the spatial attention weight matrix S:

[0042]

[0043] in Softmax is the scaling factor, and Softmax is the normalization function;

[0044] Feature aggregation: Multiply the attention weight matrix S with the value mapping V of flattened size HW×C to obtain the weighted aggregated features.

[0045]

[0046] Then The dimensions have been resized to H×W×C;

[0047] Residual Connections and Output: The aggregated features are channel-adjusted through a 1×1 convolutional layer and then residually added to the original input features to obtain the final output of the non-local attention block.

[0048]

[0049] in This is for outputting convolution weights.

[0050] Furthermore, the composite loss function used when training the network model is:

[0051]

[0052] in, L1 loss is used to ensure that the absolute error between the denoising result and the clean signal is minimized at the pixel level; This is a multi-scale structural similarity loss, used to ensure that the denoised result has the highest structural similarity to the clean signal at multiple scales; For spectral loss, it is used in the frequency domain to force the network to recover the high-frequency components of the signal and improve resolution; , and For loss weights;

[0053] The formula for calculating the L1 loss is as follows:

[0054]

[0055] in, The pixel values ​​represent the true, clean signal. The model predicts pixel values, where N is the total number of pixels;

[0056] The formula for calculating the multi-scale structural similarity loss is as follows:

[0057]

[0058] Where M is the number of scales, SSIM m The SSIM value represents the structural similarity index at the m-th scale. Weights for each scale;

[0059] The formula for calculating the spectral loss is:

[0060]

[0061] in, For Fast Fourier Transform (FFT), This represents the number of frequency components.

[0062] Secondly, the present invention provides an intelligent denoising and enhancement system for ground-penetrating radar signals, comprising:

[0063] Training dataset construction module: used to generate synthetic radar profiles using forward modeling software, add simulated noise to the radar profiles to generate noisy profiles, and construct training datasets;

[0064] Model building and training module: Construct the DRSN-ATT-UNET network model, which is based on U-Net and integrates a deep residual shrinkage module in the encoder and an attention gating module in the skip connections; train the DRSN-ATT-UNET network model using a composite loss function;

[0065] The denoising and enhancement module is used to load the actual ground-penetrating radar profile data to be processed and perform standardized preprocessing; the preprocessed profile data is input into the pre-trained DRSN-ATT-UNET network model for forward inference, and the model output is the denoised and enhanced radar profile; the model output is de-standardized postprocessing, and the final high signal-to-noise ratio radar image is output.

[0066] Thirdly, the present invention provides an electronic device, including a processor and a memory, wherein the memory stores at least one program, which is loaded and executed by the processor to implement the intelligent denoising and enhancement method for ground penetrating radar signals as described in any one of the first aspects.

[0067] Fourthly, the present invention provides a storage medium storing at least one program, which is loaded and executed by a processor to implement the intelligent denoising and enhancement method for ground penetrating radar signals as described in any one of the first aspects.

[0068] The beneficial effects of this invention are:

[0069] 1) The innovative DRSN-ATT-UNET network of this invention achieves adaptive soft thresholding of noise through a deep residual shrinkage module, effectively suppressing complex noise without damaging the effective signal. Through attention gating and spectral loss, it significantly improves the signal fidelity and resolution, especially for weak signals and signal edges. Experiments show that the signal-to-noise ratio (PSNR) of the data processed by this method can reach over 12.3 dB, and the structural similarity index (SSIM) is greater than 0.85.

[0070] 2) Because the model is trained on synthetic data containing a large number of mine noise patterns and can be fine-tuned to adapt to specific work areas, it has strong generalization ability and robustness to various complex electromagnetic environments, overcoming the shortcomings of traditional methods in terms of difficult parameter adjustment. Once the model training is complete, its denoising processing of a single radar profile can be completed in seconds (in a GPU environment), realizing the automation of the processing, greatly reducing manual intervention and improving data processing efficiency. Attached Figure Description

[0071] Figure 1 This invention provides an overall flowchart of a method for intelligent denoising and enhancement of ground-penetrating radar signals.

[0072] Figure 2 This is a schematic diagram of the DRSN-ATT-Unet network model.

[0073] Figure 3 This is a schematic diagram of the deep residual shrinkage module;

[0074] Figure 4This is a schematic diagram of the attention gating module. Detailed Implementation

[0075] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] See Figures 1-4 The present invention provides a technical solution:

[0077] Example 1:

[0078] This invention provides an intelligent denoising and enhancement method for ground-penetrating radar signals, such as... Figure 1 As shown, it includes the following steps:

[0079] S1: Generate synthetic radar profiles using forward modeling software, add simulated noise to the radar profiles to generate noisy profiles, and construct a training dataset, specifically including:

[0080] 5,000 synthetic radar profiles containing cavities, faults, and bedding of different sizes were generated using ground-penetrating radar forward modeling software as clean signals.

[0081] Gaussian white noise, pink noise, coherent noise, and random impulse noise are manually added to the clean signal to simulate noise interference in the complex environment of a mine, forming a noisy-clean signal pair and constructing a training set.

[0082] S2: Construct the DRSN-ATT-UNET network model, which is based on U-Net and integrates a deep residual shrinkage module in the encoder and an attention gating module in the skip connections; train the DRSN-ATT-UNET network model using a composite loss function.

[0083] S3: Load the actual ground-penetrating radar profile data to be processed and perform standardized preprocessing; input the preprocessed profile data into the trained DRSN-ATT-UNET network model for forward inference, and the model output is the denoised and enhanced radar profile; perform inverse standardized postprocessing on the model output and output the final high signal-to-noise ratio radar image.

[0084] like Figure 2 As shown, the structure of the DRSN-ATT-UNET network model includes:

[0085] The encoder consists of four downsampling units. Each downsampling unit includes a convolutional block, a depth residual shrinking module, and a max pooling layer. The depth residual shrinking module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features. The convolutional block consists of a convolutional layer, batch normalization, and a ReLU activation function.

[0086] Decoder: Composed of four upsampling units, each upsampling unit includes a transposed convolutional layer, an attention gating module and a convolutional block; the feature maps of the encoder and decoder are spliced ​​together through skip connections, and the attention gating module is embedded in the skip connections to gradually restore spatial resolution and focus on important signal features;

[0087] Bottleneck layer: Located between the encoder and decoder, it contains a non-local attention block to capture long-range dependencies across the entire profile, enhancing the model's understanding of the global structure.

[0088] The structure of the deep residual shrinkage module in the DRSN-ATT-UNET network model is as follows: Figure 3 As shown, the depth residual shrinkage module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features, specifically including:

[0089] Let the input feature map be Where H is the height, W is the width, and C is the number of channels, residual feature extraction is performed on the input feature map. The extraction process of residual feature R is expressed as follows:

[0090]

[0091] in For convolution kernel weights, Indicates the convolution operation;

[0092] Adaptive soft thresholding is performed by global average pooling on the input feature map X to obtain channel-level statistical features. :

[0093]

[0094] in c For channel indexing;

[0095] Channel attention weights are generated using two fully connected (FC) layers:

[0096]

[0097] in For the weights of the fully connected layer, For the Sigmoid function;

[0098] Calculate the soft threshold for each channel :

[0099]

[0100] Based on the soft threshold of each channel By applying soft thresholding to each element of the residual feature R, we obtain the residual shrinkage feature:

[0101]

[0102] The residual shrinkage feature and the input feature map are added element-wise, and the denoised feature map is output. The subnetwork learns channel-level soft thresholds; these soft thresholds are then used to shrink the residuals of the convolutional features, setting features with absolute values ​​less than the threshold to zero or reducing them.

[0103] The structure of the attention gating module in the DRSN-ATT-UNET network model is as follows: Figure 4 As shown, the attention gating module is embedded in the skip connection to progressively restore spatial resolution and focus on important signal features, specifically including:

[0104] Feature interaction: The feature map p output by the encoder is convolved with the gate signal g of the decoder, and then added element by element to generate intermediate features;

[0105] Weight calculation: ReLU activation, convolution and Sigmoid activation steps are performed on the intermediate features to generate a spatial attention weight map A. The spatial attention weight map outputs weight values ​​of 0-1, which are used to weight the encoder features before feature fusion.

[0106] Feature weighting: Multiply the attention weight A with the feature map p output by the original encoder, and output p'=Ap.

[0107] The attention gating module uses the high-level semantic features of the decoder to weight the low-level spatial features of the encoder to generate an attention map, so as to focus on important regions during feature fusion, making the network pay more attention to the target signal region and suppressing irrelevant background.

[0108] The non-local attention block specifically includes:

[0109] Feature Mapping and Embedding: Let the input feature map of the bottleneck layer be... The query map Q, key map K, and value map V are generated through three independent 1×1 convolutional layers, respectively:

[0110]

[0111] in These are the convolution weights;

[0112] Similarity matrix calculation: Flatten the query mapping Q and key mapping K into two-dimensional matrices with dimensions respectively. and Calculate the spatial attention weight matrix S:

[0113]

[0114] in Softmax is the scaling factor, and Softmax is the normalization function;

[0115] Feature aggregation: Multiply the attention weight matrix S with the value mapping V of flattened size HW×C to obtain the weighted aggregated features.

[0116]

[0117] Then The dimensions have been resized to H×W×C;

[0118] Residual Connections and Output: The aggregated features are channel-adjusted through a 1×1 convolutional layer and then residually added to the original input features to obtain the final output of the non-local attention block.

[0119]

[0120] in This is for outputting convolution weights.

[0121] Furthermore, the composite loss function used when training the network model is:

[0122]

[0123] in, L1 loss is used to ensure that the absolute error between the denoising result and the clean signal is minimized at the pixel level; This is a multi-scale structural similarity loss, used to ensure that the denoised result has the highest structural similarity to the clean signal at multiple scales; For spectral loss, it is used in the frequency domain to force the network to recover the high-frequency components of the signal and improve resolution; , and This is the loss weight.

[0124] The formula for calculating the L1 loss is as follows:

[0125]

[0126] in, The pixel values ​​represent the true, clean signal. The model predicts pixel values, where N is the total number of pixels;

[0127] The formula for calculating the multi-scale structural similarity loss is as follows:

[0128]

[0129] Where M is the number of scales, SSIM m The SSIM value represents the structural similarity index at the m-th scale. Weights for each scale;

[0130] SSIM (Structural Similarity Index Measure) and MS-SSIM (Multi-Scale Structural Similarity Index Measure) are metrics for measuring the fidelity of structural information in an image. MS-SSIM is an improved version of SSIM.

[0131] The formula for calculating the spectral loss is:

[0132]

[0133] in, For Fast Fourier Transform (FFT), This represents the number of frequency components.

[0134] In a specific embodiment, taking the processing of 100MHz ground-penetrating radar data collected in a metal mine tunnel as an example, the specific steps include:

[0135] 5000 synthetic radar profiles containing targets of different sizes, such as cavities, faults, and bedding, were generated using gprMax forward modeling software. Subsequently, Gaussian white noise, pink noise, coherent noise, and random impulse noise were added to these "clean" data to simulate mine interference and construct a training set.

[0136] Build the DRSN-ATT-UNET network using the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -4 The batch size was 16, and training was performed on a workstation equipped with an NVIDIA RTX 3090 GPU. The weight coefficients in the composite loss function were set to... Training lasts for 200 epochs until the loss function converges. The AdamW optimizer update rule during training is as follows:

[0137]

[0138] in: and These are first-order and second-order momentum deviation correction terms, respectively; This is the learning rate.

[0139] Load a heavily jammed radar profile (1024 sampling points, 500 channels) collected in the field. Standardize the data for mean and variance. Input the standardized data into the trained model, perform forward propagation, and output the denoised profile. Destandardize the output data and render it as a grayscale image for interpretation.

[0140] Compared with traditional filtering methods (such as wavelet denoising and FK filtering), the results processed by this invention effectively suppress random noise and coherent noise, and the layered reflection phase axes that were originally submerged by noise in the deep become clearly visible. Moreover, the breakpoints are crisp and clear, without obvious blurring or trailing phenomena, which greatly improves the usability of the data and the accuracy of geological interpretation.

[0141] This invention presents an innovative DRSN-ATT-UNET network that achieves adaptive soft thresholding of noise through a deep residual shrinkage module, effectively suppressing complex noise without damaging the valid signal. Through attention gating and spectral loss, it significantly improves the signal fidelity and resolution, especially for weak signals and signal edges. Experiments show that the signal-to-noise ratio (PSNR) of the data processed by this method can reach over 12.3 dB, and the structural similarity index (SSIM) is greater than 0.85.

[0142] Because the model is trained on synthetic data containing a large number of mine noise patterns and can be fine-tuned to suit specific work areas, it has strong generalization ability and robustness to various complex electromagnetic environments, overcoming the shortcomings of traditional methods in terms of difficult parameter adjustment. Once the model training is complete, its denoising processing of a single radar profile can be completed in seconds (in a GPU environment), realizing the automation of the processing, greatly reducing manual intervention and improving data processing efficiency.

[0143] Example 2:

[0144] This invention provides an intelligent denoising and enhancement system for ground-penetrating radar signals, comprising:

[0145] Training dataset construction module: used to generate synthetic radar profiles using forward modeling software, add simulated noise to the radar profiles to generate noisy profiles, and construct training datasets;

[0146] Model building and training module: Construct the DRSN-ATT-UNET network model, which is based on U-Net and integrates a deep residual shrinkage module in the encoder and an attention gating module in the skip connections; train the DRSN-ATT-UNET network model using a composite loss function;

[0147] The denoising and enhancement module is used to load the actual ground-penetrating radar profile data to be processed and perform standardized preprocessing; the preprocessed profile data is input into the pre-trained DRSN-ATT-UNET network model for forward inference, and the model output is the denoised and enhanced radar profile; the model output is de-standardized postprocessing, and the final high signal-to-noise ratio radar image is output.

[0148] Example 3:

[0149] The present invention provides an electronic device, including a processor and a memory, wherein the memory stores at least one program, and the at least one program is loaded by the processor and executed as described in Embodiment 1 above, which is a method for intelligent denoising and enhancement of ground penetrating radar signals.

[0150] Example 4:

[0151] The present invention provides a storage medium storing at least one program, which is loaded and executed by a processor to implement the intelligent denoising and enhancement method for ground penetrating radar signals described in Embodiment 1 above.

[0152] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for intelligent de-noising and enhancement of ground penetrating radar signals, characterized in that: Includes the following steps: S1: Use forward modeling software to generate synthetic radar profiles, add simulated noise to the radar profiles to generate noisy profiles, and build a training dataset. S2: Construct the DRSN-ATT-UNET network model, which is based on U-Net and integrates a deep residual shrinkage module in the encoder and an attention gating module in the skip connections; train the DRSN-ATT-UNET network model using a composite loss function. S3: Load the actual ground-penetrating radar profile data to be processed and perform standardized preprocessing; input the preprocessed profile data into the trained DRSN-ATT-UNET network model for forward inference, and the model output is the denoised and enhanced radar profile. The model output is denormalized and then processed to output the final high signal-to-noise ratio radar image. The structure of the DRSN-ATT-UNET network model includes: The encoder consists of four downsampling units. Each downsampling unit includes a convolutional block, a depth residual shrinking module, and a max pooling layer. The depth residual shrinking module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features. The convolutional block consists of a convolutional layer, batch normalization, and a ReLU activation function. Decoder: Composed of four upsampling units, each upsampling unit includes a transposed convolutional layer, an attention gating module and a convolutional block; the feature maps of the encoder and decoder are spliced ​​together through skip connections, and the attention gating module is embedded in the skip connections to gradually restore spatial resolution and focus on important signal features; Bottleneck layer: Located between the encoder and decoder, it contains a non-local attention block to capture long-range dependencies in the entire profile and enhance the model’s understanding of the global structure. The depth residual shrinkage module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features, specifically including: Let the input feature map be where H is the height, W is the width, and C is the number of channels, the residual feature R is extracted from the input feature map, and the extraction process of the residual feature R is represented as: wherein is a convolution kernel weight, denotes a convolution operation; Adaptive soft threshold calculation, global average pooling is performed on the input feature map X to obtain channel-level statistical features : Where c is the channel index; Channel attention weights are generated using two fully connected (FC) layers: wherein is the fully connected layer weight, is the Sigmoid function; Computing soft threshold for each channel : Soft thresholding according to each channel Soft thresholding each element of the residual feature R to obtain a residual shrinked feature: The residual shrinkage feature and the input feature map are added element-wise, and the denoised feature map is output. ; The non-local attention block specifically includes: Feature Mapping and Embedding: Let the input feature map of the bottleneck layer be... The query map Q, key map K, and value map V are generated through three independent 1×1 convolutional layers, respectively: in These are the convolution weights; Similarity matrix calculation: Flatten the query mapping Q and key mapping K into two-dimensional matrices with dimensions respectively. and Calculate the spatial attention weight matrix S: in Softmax is the scaling factor, and Softmax is the normalization function; Feature aggregation: Multiply the attention weight matrix S with the value mapping V of flattened size HW×C to obtain the weighted aggregated features. Then The dimensions have been resized to H×W×C; Residual Connections and Output: The aggregated features are channel-adjusted through a 1×1 convolutional layer and then residually added to the original input features to obtain the final output of the non-local attention block. in This is for outputting convolution weights.

2. The intelligent denoising and enhancement method for ground-penetrating radar signals according to claim 1, characterized in that: Step S1 specifically includes: 5,000 synthetic radar profiles containing cavities, faults, and bedding of different sizes were generated using ground-penetrating radar forward modeling software as clean signals. Gaussian white noise, pink noise, coherent noise, and random impulse noise are manually added to the clean signal to simulate noise interference in the complex environment of a mine, forming a noisy-clean signal pair and constructing a training set.

3. The intelligent denoising and enhancement method for ground-penetrating radar signals according to claim 1, characterized in that: The attention gating module, embedded in the skip connection, is used to progressively restore spatial resolution and focus on important signal features, specifically including: Feature interaction: The feature map p output by the encoder is convolved with the gate signal g of the decoder, and then added element by element to generate intermediate features; Weight calculation: ReLU activation, convolution and Sigmoid activation steps are performed on the intermediate features to generate a spatial attention weight map A, which outputs weight values ​​of 0-1; Feature weighting: Multiply the attention weight A with the feature map p output by the original encoder, and output p'=Ap.

4. The intelligent denoising and enhancement method for ground-penetrating radar signals according to claim 1, characterized in that: The composite loss function used when training the network model is: in, L1 loss is used to ensure that the absolute error between the denoising result and the clean signal is minimized at the pixel level; This is a multi-scale structural similarity loss, used to ensure that the denoised result has the highest structural similarity to the clean signal at multiple scales; For spectral loss, it is used in the frequency domain to force the network to recover the high-frequency components of the signal and improve resolution; , and For loss weights; The formula for calculating the L1 loss is as follows: in, The pixel values ​​represent the true, clean signal. The model predicts pixel values, where N is the total number of pixels; The formula for calculating the multi-scale structural similarity loss is as follows: Where M is the number of scales, SSIM m The SSIM value represents the structural similarity index at the m-th scale. Weights for each scale; The formula for calculating the spectral loss is: in, For Fast Fourier Transform (FFT), This represents the number of frequency components.

5. A ground-penetrating radar signal intelligent denoising and enhancement system, characterized in that, include: Training dataset construction module: used to generate synthetic radar profiles using forward modeling software, add simulated noise to the radar profiles to generate noisy profiles, and construct training datasets; Model building and training module: Construct the DRSN-ATT-UNET network model, which is based on U-Net and integrates a deep residual shrinkage module in the encoder and an attention gating module in the skip connections; train the DRSN-ATT-UNET network model using a composite loss function; Denoising and Enhancement Module: This module loads the actual ground-penetrating radar profile data to be processed and performs standardized preprocessing. The preprocessed profile data is then input into the pre-trained DRSN-ATT-UNET network model for forward inference, and the model output is the denoised and enhanced radar profile. The model output is denormalized and then processed to output the final high signal-to-noise ratio radar image. The structure of the DRSN-ATT-UNET network model includes: The encoder consists of four downsampling units. Each downsampling unit includes a convolutional block, a depth residual shrinking module, and a max pooling layer. The depth residual shrinking module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features. The convolutional block consists of a convolutional layer, batch normalization, and a ReLU activation function. Decoder: Composed of four upsampling units, each upsampling unit includes a transposed convolutional layer, an attention gating module and a convolutional block; the feature maps of the encoder and decoder are spliced ​​together through skip connections, and the attention gating module is embedded in the skip connections to gradually restore spatial resolution and focus on important signal features; Bottleneck layer: Located between the encoder and decoder, it contains a non-local attention block to capture long-range dependencies in the entire profile and enhance the model’s understanding of the global structure. The depth residual shrinkage module is used to extract multi-level depth features from noisy input and adaptively suppress noise-related features, specifically including: Let the input feature map be Where H is the height, W is the width, and C is the number of channels, residual feature extraction is performed on the input feature map. The extraction process of residual feature R is expressed as follows: in For convolution kernel weights, Indicates the convolution operation; Adaptive soft thresholding is performed by global average pooling on the input feature map X to obtain channel-level statistical features. : Where c is the channel index; Channel attention weights are generated using two fully connected (FC) layers: in For the weights of the fully connected layer, For the Sigmoid function; Calculate the soft threshold for each channel : Based on the soft threshold of each channel By applying soft thresholding to each element of the residual feature R, we obtain the residual shrinkage feature: The residual shrinkage feature and the input feature map are added element-wise, and the denoised feature map is output. ; The non-local attention block specifically includes: Feature Mapping and Embedding: Let the input feature map of the bottleneck layer be... The query map Q, key map K, and value map V are generated through three independent 1×1 convolutional layers, respectively: in These are the convolution weights; Similarity matrix calculation: Flatten the query mapping Q and key mapping K into two-dimensional matrices with dimensions respectively. and Calculate the spatial attention weight matrix S: in Softmax is the scaling factor, and Softmax is the normalization function; Feature aggregation: Multiply the attention weight matrix S with the value mapping V of flattened size HW×C to obtain the weighted aggregated features. Then The dimensions have been resized to H×W×C; Residual Connections and Output: The aggregated features are channel-adjusted through a 1×1 convolutional layer and then residually added to the original input features to obtain the final output of the non-local attention block. in This is for outputting convolution weights.

6. An electronic device, characterized in that, The method includes a processor and a memory, wherein the memory stores at least one program, which is loaded and executed by the processor to implement the intelligent denoising and enhancement method for ground penetrating radar signals according to any one of claims 1 to 4.

7. A storage medium, characterized in that, The storage medium stores at least one program segment, which is loaded and executed by a processor to implement the intelligent denoising and enhancement method for ground penetrating radar signals according to any one of claims 1 to 4.

Images