Ground penetrating radar data intelligent gain method, equipment and medium

Abstract

The invention relates to the field of geophysical exploration, and discloses a ground penetrating radar data intelligent gain method and device and a medium, and the method comprises the steps: obtaining ground penetrating radar data, and carrying out the preprocessing of the ground penetrating radar data, and obtaining the preprocessed data; gaining the preprocessed data by adopting a self-adaptive segmented gain optimization algorithm based on information entropy guidance to obtain gained data; labeling the gained data by adopting a self-adaptive gain labeling method based on multi-strategy information entropy optimization, and constructing a high-quality gain training data set; designing and constructing a gain optimization network based on U-Net; training parameters are set, and the gain optimization network pair is trained by using a high-quality gain training data set so as to optimize a gain adjustment effect; the effectiveness of the network in ground penetrating radar data gain optimization is verified through experiments, and quantitative analysis is carried out on the result. The data gained by the method has the advantages of higher information amount, lower error, better signal-to-noise ratio and better structure retention capability.

Description

Technical Field

[0001] The present invention relates to the field of geophysical exploration, and particularly to an intelligent gain method, device, and medium for ground penetrating radar data. Background Art

[0002] Ground Penetrating Radar (GPR) is a non-destructive testing technology that uses high-frequency electromagnetic waves to detect underground objects and structures. It emits electromagnetic waves into the ground and receives signals reflected from underground targets, thereby obtaining physical information about underground structures, and is widely used in fields such as road construction, underground pipeline detection, dam detection, and cavity detection.

[0003] In the processing of ground penetrating radar data, gain adjustment is to increase the amplitude value of the target echo signal of the ground penetrating radar and enhance the visibility of the target signal. Since the electromagnetic wave energy of the ground penetrating radar gradually attenuates as the propagation depth increases, in order to maintain the recognizability of deep targets, it is usually necessary to apply appropriate gain compensation to the received signals. Currently, common gain processing methods include: empirical gain function setting, logarithmic gain, exponential gain, etc. These methods usually perform gain adjustment based on fixed mathematical models or empirical formulas. Although the implementation is relatively simple, it is difficult to perform adaptive optimization for different detection environments. Therefore, automatic gain processing methods have gradually attracted attention and have been widely used due to their good enhancement effect and convenience.

[0004] Currently, the automatic gain processing methods for ground penetrating radar signals mainly include technical means such as data statistics, signal analysis, and normalization processing.

[0005] The Chinese Patent Network published an automatic gain method based on data statistics (patent publication number CN117930226A). This method calculates the gain coefficient through steps such as mean removal, filtering, and data distribution correction to enhance the signal visualization effect in the disease and structure areas, and has a certain degree of self-adaptability.

[0006] Another patent (patent publication number CN117055028A) proposed an automatic gain method based on signal analysis. This method combines technologies such as wavelet transform and spectrum analysis to identify road stratification points and calculates the gain ratio based on the signal energy of each layer, making the gain more in line with the signal characteristics of different depth layers.

[0007] In addition, the open-source software GPRPy provides an automatic gain control method based on window normalization (hereinafter referred to as the window normalization method). Its core idea is to calculate the local signal energy within a set window and perform normalization processing on the signal to keep the signal amplitude of the entire radar profile balanced and enhance the visualization effect of deep targets.

[0008] At the same time, the open source software GPRLab implements an automatic gain method based on Gaussian weighted root mean square (GRMS). This method calculates the root mean square energy of the signal by setting a window, smoothes the energy distribution using Gaussian weighting, and then normalizes the signal amplitude to improve the signal contrast at different depth layers and enhance target visibility.

[0009] Although the above methods have improved the gain effect of GPR data to varying degrees, there are still problems such as limited adaptability and difficulty in taking into account complex detection environments. Therefore, how to further optimize the gain method and improve signal quality remains a key issue that needs to be urgently addressed in GPR data processing. Summary of the Invention

[0010] The purpose of the present invention is to propose a method, device and medium for intelligent gain of ground penetrating radar data to solve the technical problems of limited adaptability and difficulty in taking into account complex detection environments in existing ground penetrating radar data gain.

[0011] Specifically, the present invention provides a method, device, and medium for intelligently gaining ground penetrating radar data, the method comprising the following steps:

[0012] S1, acquiring ground penetrating radar data and performing preprocessing to obtain preprocessed data;

[0013] S2. Amplify the preprocessed data using an adaptive segmented gain optimization algorithm guided by information entropy to obtain the gained data;

[0014] S3. Annotate the gained data using an adaptive gain annotation method based on multi-strategy information entropy optimization to construct a high-quality gain training dataset;

[0015] S4. Design and build a gain optimization network based on U-Net;

[0016] S5. Setting training parameters and using a high-quality gain training data set to train the gain optimization network to optimize the gain adjustment effect;

[0017] S6. Experimentally verify the effectiveness of the U-Net network in optimizing the gain of ground penetrating radar data and quantitatively analyze the results.

[0018] A storage medium stores instructions and data for implementing a ground penetrating radar data intelligent gain method.

[0019] A ground penetrating radar data intelligent gain device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a ground penetrating radar data intelligent gain method.

[0020] The beneficial effects provided by the present invention are as follows:

[0021] Compared with traditional gain methods, the U-Net network proposed in the present invention automatically learns the gain mapping relationship of ground penetrating radar data in a data-driven manner through end-to-end training, without relying on empirical operations such as manually setting the window size and gain parameters, thus significantly improving the adaptability of the algorithm to different detection scenarios. Secondly, through multi-scale feature extraction and skip connections, U-Net enhances the deep-layer signals while retaining the details of the shallow-layer signals, avoiding the problem of unbalanced energy distribution. In addition, U-Net can accurately enhance local abnormal regions, improve the contrast of the disease regions, and make the target features clearer. Different from the step-by-step processing of traditional methods, U-Net adopts end-to-end training to complete gain optimization at one time, reducing the error propagation in the calculation process and improving gain stability. At the same time, U-Net can use GPUs for parallel computing, greatly improving the data processing efficiency and making it more suitable for large-scale GPR data analysis and real-time applications. Brief Description of the Drawings

[0022] Figure 1 is a schematic diagram of the simple process of the method of the present invention;

[0023] Figure 2 is a schematic diagram of the U-Net network structure;

[0024] Figure 3 is a schematic diagram of the change of loss in 50 training epochs;

[0025] Figure 4 is a schematic diagram of preprocessed data;

[0026] Figure 5 is a schematic diagram of label data;

[0027] Figure 6 is a schematic diagram of U-Net predicted data;

[0028] Figure 7 is a schematic diagram of the information entropy of label data and U-Net predicted data;

[0029] Figure 8 is a schematic diagram of the mean square error of label data and U-Net predicted data;

[0030] Figure 9 is a schematic diagram of the structural similarity of label data and U-Net predicted data;

[0031] Figure 10 is a schematic diagram of the working of the hardware device in the embodiment of the present invention. Detailed Embodiments

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0033] Before formally elaborating on the present invention, a general description of the solution of the present invention is first provided for easy understanding.

[0034] Please refer to Figure 1 , a ground penetrating radar data intelligent gain method provided by the present invention includes:

[0035] S1. Obtain ground penetrating radar data and perform preprocessing to obtain preprocessed data;

[0036] It should be noted that the preprocessing in step S1 includes: direct wave removal and background removal.

[0037] As an embodiment, the original data stored in the ground penetrating radar and the key information recorded in the header file are read, including parameters such as the direct wave position, the number of data bits, the number of samples per channel, and the number of acquired channels.

[0038] For the original data, first perform direct wave removal based on the direct wave position. Removing the direct wave can improve the quality of the radar image and enhance the visibility of the reflected signals of underground targets. Then perform background removal on the data after removing the direct wave. The background removal is performed by subtracting the mean value from each row, aiming to eliminate fixed interference signals, such as system noise or surface direct waves, and enhance the prominence of the reflected signals in the data.

[0039] Both of the above two data preprocessing steps are automatically calculated by the program in combination with the header file information in the read data, without manual intervention by the operator. Hereinafter, the data after removing the direct wave and background removal will be referred to as preprocessed data.

[0040] S2. Apply an adaptive segmented gain optimization algorithm guided by information entropy to the preprocessed data for gain to obtain gain data;

[0041] It should be noted that step S2 is specifically as follows:

[0042] S21. Divide the preprocessed data into n segments from top to bottom;

[0043] S22. Initialize the gain value of the first sampling point of the i-th segment of data to 0, and initialize the gain value of the last sampling point of the i-th segment of data to 1. For the intermediate sampling points of the i-th segment of data, calculate the gain values of each sampling point by linear interpolation; where i = 1, 2, 3…, n;

[0044] S23. Calculate the information entropy of the complete data using the calculated gain value of each sampling point and compare it with the initial information entropy. If the information entropy increases, record the new information entropy value as the current optimal information entropy, and continue to increase the gain value of the sampling point at the end of the i-th segment of data by 1 each time, while updating the gain value of the middle part.

[0045] Repeat the above steps and compare with the current optimal information entropy. When the information entropy begins to decrease or remains unchanged, it means that the gain effect has reached the optimal value, and the tail gain value at this time is determined as the optimal tail gain value of the i-th segment data;

[0046] S24. Perform global traversal optimization to obtain the gained data.

[0047] As an example, the present invention proposes an adaptive segmented gain optimization algorithm guided by information entropy. Compared to the limitations of traditional manual gain methods, which require manual segmentation of data and setting gain values for each segment, this method introduces information entropy theory to construct a quantitative evaluation system, achieving intelligent optimization of gain parameters. The specific implementation process of this method is as follows:

[0048] In terms of theoretical basis, information entropy is the core indicator of information theory (Shannon entropy), and its mathematical expression is:

[0049]

[0050] Where p(x i ) indicates that the gray value in the image is x i The probability of a pixel appearing. This indicator can effectively quantify the complexity and information content of the data. The higher the entropy value, the richer the effective information contained in the data.

[0051] Due to environmental interference and dielectric attenuation, raw ground-penetrating radar signals often have low contrast and resolution, resulting in low information entropy. Gain processing can enhance the local characteristics of target signals and improve the information entropy of the data.

[0052] This method is based on the data processed in step S1, which has completed the direct wave removal and background removal. The information entropy is recorded as the initial information entropy. The specific gain optimization steps are as follows:

[0053] 1. Data segmentation: The GPR data is divided into n segments from top to bottom. The gain value of each segment is continuously adjusted during the traversal process to obtain the optimal gain solution.

[0054] 2. Gain initialization. Set the gain value at the beginning of the first data segment to 0 and the gain value at the end to 1. Calculate the gain value of each sampling point in the middle of this segment through linear interpolation. For the remaining unprocessed segments, temporarily set the gain value to 0, i.e., no gain.

[0055] 3. Information entropy optimization.

[0056] After applying the calculated gain value, calculate the information entropy of the complete data and compare it with the initial information entropy. If the information entropy increases, record the new information entropy value as the current optimal information entropy, and continue to increase the gain value at the end of the first segment by 1 each time, while updating the gain value in the middle part.

[0057] Iterate the above steps repeatedly, calculate the information entropy of the processed data, and compare it with the current optimal information entropy. When the information entropy starts to decrease or remains unchanged, it indicates that the gain effect has reached the optimal, and the gain value at the end at this time is determined as the optimal gain value at the end of the first segment of data.

[0058] 4. Global traversal optimization. Traverse and optimize the remaining segments of data in the same way in turn. Through segment-by-segment optimization, finally obtain the optimal gain distribution of the overall data to maximize the information entropy. In practical applications, a segmentation interval can be set for each data, and the algorithm traverses all possible segmentation methods within this interval and selects the segmentation value that maximizes the information entropy as the finally determined optimal gain strategy.

[0059] The technical advantages of this method are reflected in: ① Establish a quantitative evaluation standard for the gain effect through information entropy, avoiding the subjectivity of manual experience; ② Use a segment-by-segment traversal algorithm to achieve automatic optimization of gain parameters, with significantly improved efficiency compared to the manual mode; ③ Combine linear interpolation technology to ensure the smooth transition of the gain curve and effectively suppress signal distortion. Experiments show that this method can increase the information entropy of ground penetrating radar data by 100 - 300%, significantly improving the reliability of subsequent inversion and interpretation.

[0060] S3. Label the data after gain using an adaptive gain annotation method based on multi-strategy information entropy optimization to construct a high-quality gain training data set;

[0061] It should be noted that the adaptive gain annotation method based on multi-strategy information entropy optimization in step S3 is specifically as follows:

[0062] S31. Process the preprocessed data using the method in step S2 to obtain the first data after gain;

[0063] S32. Process the preprocessed data using the GPRMS gain method to obtain the second data after gain;

[0064] S33. Process the preprocessed data using the window normalization method to obtain the third data after gain;

[0065] S34. Calculate the information entropy of the data after the first gain, the data after the second gain, and the data after the third gain respectively, compare them, and select the processing result with the highest information entropy as the final labeled data;

[0066] S35. After performing noise reduction processing on the final labeled data, construct a high-quality gain training data set based on this.

[0067] As an embodiment, in step S3, a ground penetrating radar gain adaptive annotation method based on multi-strategy information entropy optimization is proposed, aiming to comprehensively integrate different gain strategies, automatically select the optimal gain scheme, and provide high-quality labeled data for subsequent model training. This method combines the principle of maximizing information entropy, processes the ground penetrating radar data through multiple gain methods, and screens out the result with the highest information entropy as the final gain annotation data. The specific implementation process of this method is as follows:

[0068] 1. Adopt the information entropy-guided adaptive segmented gain optimization algorithm proposed in step S2 to optimize the gain of each ground penetrating radar data. To balance the gain effect and computational efficiency, set the minimum number of segments to 10 and the maximum number of segments to 50. The algorithm traverses all possible numbers of segments (between 10 and 50), calculates the optimal gain corresponding to each segmentation method, and records its information entropy. Select the gain scheme that maximizes the information entropy as the optimization result of this method.

[0069] 2. Use the GRMS gain method provided by the open-source ground penetrating radar data processing software GPRlab to enhance the ground penetrating radar data. This method is based on Gaussian mean square filtering, which can smooth the data, enhance the contrast, and improve the visualization effect of the radar data. Calculate the information entropy of the data after processing and record the optimization result.

[0070] 3. Use the window normalization method provided by the open-source ground penetrating radar data processing software GPRPy to process each ground penetrating radar data. This method can adaptively adjust the gain parameters according to the local characteristics of the data, making the gain more flexible and robust. Calculate the information entropy of the processed data and record the result.

[0071] Optimize by combining the above three gain methods to find a relatively optimal gain method.

[0072] Calculate the information entropy of the data processed by the three methods respectively and compare them.

[0073] Select the processing result with the highest information entropy as the final labeled data to ensure the optimal gain effect of the labeled data.

[0074] On this basis, the labeled data is uniformly denoised, enabling the U-Net network to not only learn the gain method but also master the denoising strategy during the training process, thereby generating data that enhances the signal and effectively suppresses noise, improving the comprehensive adaptability of the model. This method uses mean filtering for denoising, which can smooth the data and improve the signal-to-noise ratio, further optimizing the quality of the data after gain.

[0075] This method integrates three strategies: an adaptive segmented gain optimization algorithm guided by information entropy, GRMS of GPRlab, and the window normalization method provided by GPRPy. It uses information entropy evaluation to automatically screen the optimal gain scheme, avoiding human intervention. Its advantage lies in balancing the gain effect and computational efficiency, improving data quality, ensuring the maximization of the information volume and enhanced recognizability of ground-penetrating radar data, and providing high-quality labeled data for subsequent deep learning training.

[0076] S4. Design and construct a gain optimization network based on U-Net;

[0077] It should be noted that the U-Net-based gain optimization network in step S4 includes: an encoder, a decoder, and an output layer. Among them, the encoder consists of multiple double convolutional modules and a max pooling layer; the decoder restores the spatial dimension of the feature map through bilinear interpolation upsampling and splices the features with the corresponding layers of the encoder through skip connections; the output layer uses a 1×1 convolutional kernel.

[0078] As an embodiment, the network structure of U-Net is characterized by its encoding-decoding symmetric design, which can effectively extract and reconstruct the original data features. In the present invention, the U-Net network structure shown as Figure 2 is used as the deep neural network, and the specific process is as follows:

[0079] The U-Net network structure mainly includes an encoder (downsampling), a decoder (upsampling), and a final prediction output layer. The encoder part consists of multiple double convolutional modules (double_conv) and a max pooling layer (MaxPool2d). Among them, the double convolutional module includes two convolutional layers with 3×3 convolutional kernels and uses the ReLU activation function to enhance the feature expression ability. The downsampling part successively includes four convolutional layer groups. The input data first passes through a convolutional layer with 1 channel for feature extraction, and then the number of channels increases to 64, 128, 256, and 512 in sequence, and the spatial dimension is reduced through the max pooling layer, thereby extracting features of different scales.

[0080] The decoder part restores the spatial dimension of the feature map through bilinear interpolation upsampling (Upsample), and splices (Concat) the features of the corresponding layer of the encoder through skip connection to retain the low-level feature information. The upsampling part successively adopts three groups of double-layer convolution modules, with the number of channels reduced from 512 to 256, then to 128, and finally restored to 64. The skip connection splices the features of the third, second, and first layers in the encoder with the features of the corresponding layer in the decoder, so as to retain the key feature information while ensuring the resolution restoration.

[0081] The final output layer uses a 1×1 convolution kernel to convert the number of channels into the number of target categories and outputs the enhanced ground penetrating radar data. By combining the U-Net structure with skip connection, this method can effectively improve the gain effect of the data, make the data after the ground penetrating radar signal enhancement more visual and interpretable, and thus improve the detection accuracy and target recognition ability.

[0082] S5. Set the training parameters, and use the high-quality gain training dataset to train the gain optimization network to optimize the gain adjustment effect;

[0083] As an embodiment, please refer to Table 1, which is the U-Net training parameter table of the present invention;

[0084] Table 1 U-Net training parameter table

[0085]

[0086] In step S5, it involves setting relevant training parameters and using U-Net for training, and adopts a complete training process to ensure that the model can effectively learn and generalize. The following is the specific implementation method:

[0087] The parameter configuration of the training is shown in Table 1. The model adopts the U-Net architecture, and its core parameter configuration is: the number of input channels is 1, the convolutional channels of the encoder increase gradually in multiples, the decoder performs bilinear interpolation upsampling and has a skip connection with the encoder features, and finally outputs the prediction result through 1×1 convolution. The mean square error (MSE) is used as the loss function, and the Adam algorithm is selected as the optimizer (the learning rate is set to 1×10 -4 ), the batch size is 32, and the network parameters are updated by mini-batch stochastic gradient descent to ensure the convergence stability of the model on complex radar data.

[0088] The data division adopts a stratified sampling strategy, and the original dataset is divided into a training set (80%), a validation set (10%), and a test set (10%) according to a certain proportion. The validation set is used to monitor the generalization ability in real time, and the test set is used for the final performance evaluation. During the training process, GPU acceleration calculation is enabled, and the risk of overfitting is controlled through an early stopping mechanism (the patience parameter is set to 5 epochs): when the validation loss has not improved for 5 consecutive training cycles, the training is automatically terminated, and at the same time, the model state with the lowest validation loss is saved.

[0089] S6. Verify the effectiveness of the U-Net network in optimizing the gain of ground penetrating radar data through experiments, and conduct quantitative analysis on the results.

[0090] As an embodiment, step S6 involves verifying the effectiveness of the U-Net network in the gain processing of ground penetrating radar data. Based on an adaptive annotation method optimized by multi-strategy information entropy, a large-scale dataset covering various geological environments is constructed, and multi-dimensional quantitative indicators are used to evaluate the gain effect of the network.

[0091] The specific implementation process of this method is as follows:

[0092] 1. Dataset construction and preprocessing. In this study, 3000 ground penetrating radar data were collected. These data were collected by CO1760 and Raptor-45 of impulseradar, and the data sources cover various geological environments such as urban roads, highways, dams, and sand dunes to ensure that the algorithm has good adaptability under different geological conditions. To ensure data quality, each frame of data undergoes preprocessing such as direct wave removal and background removal to improve the usability and effectiveness of the data. Subsequently, using the ground penetrating radar gain adaptive annotation method optimized by multi-strategy information entropy (see step S3), the data is optimized for gain using three methods: the adaptive segmented gain optimization algorithm guided by information entropy, GRMS of GPRlab, and window normalization in GPRPy, respectively, and the optimal gain result is selected as the training label for the model. Finally, in the labeled dataset, the proportion of the data volume processed by different gain methods is 1352:850:798. The proportion is relatively average, which can ensure that the training data covers various gain methods, so that the U-Net can learn the advantages of different methods and achieve a more robust gain mapping.

[0093] 2. Model training and evaluation metrics. The dataset is divided into a training set, a validation set, and a test set according to a ratio of 8:1:1. The following Figure 3 is the loss change curve of the model within 50 rounds of training. From the loss curve, the model converges successfully within 50 rounds of training. The losses of the training set and the validation set decrease and tend to be stable, and the two are close, indicating that the U-Net has a good training effect, no obvious overfitting problem, and the model has good generalization ability.

[0094] From the visualization results, Figure 4 For the data without gain processing, the signal is weak and the noise is obvious, especially in the deep area where the signal attenuation is serious. Figure 5 For the labeled data, the signal is significantly enhanced and the background noise is effectively suppressed. The prediction results of U-Net are as Figure 6 , which is highly consistent with the labeled data in terms of signal enhancement and noise suppression, clearly retaining the characteristics of the main reflection layer. Especially in the deep area, the data predicted by U-Net amplifies the effective signal at the bottom and suppresses the noise, and the prediction effect is close to that of the labeled data.

[0095] Three key indicators, information entropy (IE), mean square error (MSE), and structural similarity (SSIM), are used to quantify the difference between the model-predicted data and the true labeled data. IE is used to evaluate the information content in the data. MSE reflects the overall pixel error of the image. The smaller the error, the more similar the images. SSIM can measure the overall structural similarity of the image, avoiding the over-smoothing or distortion problems that may occur with MSE. The experiment is evaluated based on 300 sets of test set data.

[0096] From Figure 7 It can be seen that the information entropy of the labeled data is relatively close to that of the predicted data, indicating that the data generated by U-Net maintains a high amount of information while enhancing the signal details. Figure 8 In, the MSE between the predicted data and the labeled data is about 0.0046, with a small error, indicating that the model can learn the gain mapping relationship well. Figure 9 It can be known that the average SSIM between the predicted data and the labeled data is about 0.83, indicating that the model can better retain the structural characteristics of the data.

[0097] To verify the computational efficiency of this method, tests are carried out in the hardware environment of AMD Ryzen9 9950X CPU, 128GB of memory, and NVIDIA GeForce RTX 4090 (24GB of video memory). Taking the ground penetrating radar data with a resolution of 256×256 as an example, the average processing time of each frame of data is shown in Table 2, where the processing method proposed in the present invention takes the shortest time. Compared with the traditional gain method, the present invention greatly improves the computational efficiency while ensuring the gain effect, meets the requirements of large-scale ground penetrating radar data processing, and has good real-time processing capabilities, and can be applied to on-line detection and automatic data enhancement.

[0098] Table 2 Comparison of average processing time of each method

[0099]

[0100] The experimental results show that the present invention is superior to traditional methods in terms of gain effect and computational efficiency, and can provide stable and reliable gain enhancement processing in different geological environments. The data after gain has higher information content, lower error, better signal-to-noise ratio and better structure preservation ability, which fully proves the innovation and practical value of the present method in the gain processing of ground penetrating radar data, and lays a solid foundation for subsequent intelligent detection and data automatic enhancement.

[0101] Please refer to Figure 10 , Figure 10 which is a schematic diagram of the operation of the hardware device according to an embodiment of the present invention. The hardware device specifically includes: an intelligent gain device 401 for ground penetrating radar data, a processor 402, and a storage medium 403.

[0102] An intelligent gain device 401 for ground penetrating radar data: The intelligent gain device 401 for ground penetrating radar data implements the intelligent gain method for ground penetrating radar data.

[0103] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the intelligent gain method for ground penetrating radar data.

[0104] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the intelligent gain method for ground penetrating radar data.

[0105] The beneficial effects of the present invention are:

[0106] Compared with traditional gain methods, the U-Net network proposed by the present invention automatically learns the gain mapping relationship of ground penetrating radar data in a data-driven manner through end-to-end training, without relying on empirical operations such as manually setting the window size and gain parameters, thus significantly improving the adaptability of the algorithm to different detection scenarios. Secondly, through multi-scale feature extraction and skip connections, U-Net enhances the deep signals while retaining the details of the shallow signals, avoiding the problem of unbalanced energy distribution. In addition, U-Net can accurately enhance local abnormal regions, improve the contrast of disease regions, and make the target features clearer. Different from the step-by-step processing of traditional methods, U-Net adopts end-to-end training to complete gain optimization at one time, reduces the error propagation in the calculation process, and improves gain stability. At the same time, U-Net can use GPU for parallel computing, greatly improving the data processing efficiency, making it more suitable for large-scale GPR data analysis and real-time applications.

[0107] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent gain method for ground penetrating radar data, characterized in that: It includes the following steps: S1. Obtain ground penetrating radar data and perform preprocessing to obtain preprocessed data; S2. Apply the adaptive segmentation gain optimization algorithm guided by information entropy to the preprocessed data for gain to obtain gain-adjusted data; S3. Apply the adaptive gain annotation method optimized based on multi-strategy information entropy to the gain-adjusted data for annotation to construct a high-quality gain training dataset; S4. Design and construct a gain optimization network based on U-Net; S5. Set training parameters and use the high-quality gain training dataset to train the gain optimization network to optimize the gain adjustment effect; S6. Verify the effectiveness of the U-Net network in the gain optimization of ground penetrating radar data through experiments and conduct quantitative analysis on the results.

2. The intelligent gain method for ground penetrating radar data according to claim 1, wherein: The preprocessing described in step S1 includes: direct wave removal and background removal.

3. The intelligent gain method for ground penetrating radar data according to claim 1, characterized in that: Step S2 is specifically as follows: S21. Divide the preprocessed data into n segments from top to bottom; S22. Initialize the gain value of the first sampling point of the i-th segment of data to 0 and the gain value of the last sampling point of the i-th segment of data to 1. For the intermediate sampling points of the i-th segment of data, calculate the gain value of each sampling point using linear interpolation; where i = 1, 2, 3…, n; S23. Calculate the information entropy of the complete data using the calculated gain value of each sampling point and compare it with the initial information entropy; if the information entropy increases, record the new information entropy value as the current optimal information entropy, and continue to increase the gain value of the last sampling point of the i-th segment of data by 1 each time, and update the gain value of the intermediate part at the same time; Iterate the above steps repeatedly and compare with the current optimal information entropy. When the information entropy starts to decrease or remains unchanged, it indicates that the gain effect has reached the optimum, and the gain value at the end at this time is determined as the optimal end gain value of the i-th segment of data; S24. Perform global traversal optimization to obtain gain-adjusted data.

4. The intelligent gain method for ground penetrating radar data according to claim 3, wherein: The adaptive gain annotation method optimized based on multi-strategy information entropy in step S3 is specifically as follows: S31. Process the preprocessed data using the method in step S2 to obtain the first gain-adjusted data; S32. Process the preprocessed data using the GPRMS gain method to obtain the second gain-adjusted data; S33. Process the preprocessed data using the window normalization method to obtain the third gain-adjusted data; S34. Calculate the information entropy of the first gain-adjusted data, the second gain-adjusted data, and the third gain-adjusted data respectively, and compare them, and select the processing result with the highest information entropy as the final annotation data; S35. After performing noise reduction processing on the final annotation data, construct a high-quality gain training dataset based on this.

5. The intelligent gain method for ground penetrating radar data according to claim 1, characterized in that: The gain optimization network based on U-Net in step S4 includes: an encoder, a decoder, and an output layer.

6. The intelligent gain method for ground penetrating radar data according to claim 5, characterized in that: The encoder consists of multiple double-layer convolutional modules and max pooling layers; the decoder restores the spatial dimension of the feature map through bilinear interpolation upsampling and splices the features with the corresponding layers of the encoder through skip connections; the output layer uses a 1×1 convolutional kernel.

7. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a method for intelligent gain of ground penetrating radar data according to any one of claims 1 to 6.

8. An intelligent gain device for ground penetrating radar data, characterized in that: It includes: a processor and a storage medium; the processor loads and executes the instructions and data in the storage medium for implementing a method for intelligent gain of ground penetrating radar data according to any one of claims 1 to 6.

Images