Ground penetrating radar data inversion method and device, computer equipment and storage medium

Abstract

The application provides a ground penetrating radar data inversion method and device, computer equipment and a storage medium, comprising: receiving ground penetrating radar data; calling a preset one-stage network model to perform inversion operation on the ground penetrating radar data to obtain a first dielectric constant model; calling a preset two-stage network model to perform inversion operation on the ground penetrating radar data and the first dielectric constant model to obtain a second dielectric constant model. The application ensures the accuracy of the first dielectric constant model, modifies the incorrect structural features in the first dielectric constant model, and obtains a more accurate second dielectric constant model, greatly improves the operation efficiency of ground penetrating radar data inversion, and reduces the consumption of computing resources.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a ground-penetrating radar data inversion method, apparatus, computer equipment and storage medium. Background Technology

[0002] Ground penetrating radar data inversion is a problem of solving the ill-conditioned problem of ground penetrating radar data and the dielectric constant model space with at least one relative permittivity. An ill-conditioned problem is one in which the output is very sensitive to the input data, that is, a small error in the input data may cause a large change in the output. The condition number is generally used to measure the ill-conditioning index of the problem. The larger the condition number, the more ill-conditioned the problem.

[0003] However, current ground-penetrating radar (GPR) data inversion typically employs the traditional full waveform inversion (FWI) method for qualitative and quantitative reconstruction of stratigraphic structures. It directly uses the entire received waveform to match simulated GPR data. Then, it reconstructs the dielectric distribution of the structure by minimizing the mismatch between the two sets of data.

[0004] However, the inventors discovered that traditional FWI typically uses an iterative approach to reduce the error between analog data and ground-penetrating radar data. Therefore, completing one FWI operation requires a large amount of computational resources. Furthermore, the FWI method has difficulty extracting eigenvalues ​​from ground-penetrating radar data that are sensitive to the dielectric constant model, resulting in low inversion accuracy of the FWI method. Summary of the Invention

[0005] This application provides a ground-penetrating radar data inversion method, apparatus, computer equipment, and storage medium to solve the problems of traditional FWI methods requiring a large amount of computing resources and having difficulty extracting feature values ​​from ground-penetrating radar data that are sensitive to the dielectric constant model, resulting in low inversion accuracy of FWI methods.

[0006] In a first aspect, this application provides a ground-penetrating radar data inversion method, including:

[0007] Receive ground-penetrating radar data; wherein, the ground-penetrating radar data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of the target area;

[0008] A pre-set one-stage network model is invoked to perform inversion calculations on the ground penetrating radar data to obtain a first dielectric constant model; wherein, the first dielectric constant model is a factor that determines the propagation speed of the electromagnetic wave in the geological structure of the target area; the first dielectric constant model has at least one relative dielectric constant; the relative dielectric constant is a physical parameter characterizing the dielectric properties of the geological structure;

[0009] A preset two-stage network model is invoked to perform inversion operations on the ground penetrating radar data and the first dielectric constant model to obtain a second dielectric constant model; wherein, the second dielectric constant model includes physical parameters after correcting the relative dielectric constant in the first dielectric constant model based on the ground penetrating radar data.

[0010] In the above scheme, before calling the preset one-stage network model to perform inversion calculations on the ground-penetrating radar data, the method further includes:

[0011] Obtain the first training sample;

[0012] The first-stage network model is obtained by training the first initial network model pre-set using the first training sample.

[0013] In the above scheme, before invoking the preset two-stage network model to perform inversion calculations on the ground-penetrating radar data and the first dielectric constant model, the method further includes:

[0014] Obtain the second training sample;

[0015] The second-stage network model is obtained by training the preset second initial network model using the second training samples and the first-stage network model.

[0016] In the above scheme, before obtaining the first training sample, the method further includes:

[0017] The geological structure is obtained, and irregular blocks are embedded in the geological structure to convert the geological structure into a training dielectric constant model. Forward modeling is then performed based on the training dielectric constant model to obtain training ground penetrating radar data.

[0018] A training dataset is formed by combining the aforementioned training dielectric constant model with the aforementioned training ground-penetrating radar data.

[0019] The training set is obtained by summing up several sets of training data.

[0020] In the above scheme, obtaining the first training sample includes:

[0021] M training data points are obtained from the training set, and the M training data points are aggregated to obtain the first training sample; where M is a positive integer, M≥1;

[0022] The process of obtaining the second training sample includes:

[0023] N training data points are obtained from the training set, and the N training data points are aggregated to obtain the second training sample; where N is a positive integer and N≥1.

[0024] In the above scheme, the first-stage network model is obtained by training a first initial network model pre-set using the first training sample, including:

[0025] The training ground-penetrating radar data in the first training sample is used as the first input information of the first initial network model, and the first initial network model is run to perform inversion operation on the first input information to obtain the first output information.

[0026] The training dielectric constant model of the training data in the first training sample is used as the first reference information of the first initial network model. A first loss value is generated based on the first output information and the first reference information using a preset first loss function. The first loss value characterizes the degree of difference between the first output information and the first reference information.

[0027] The first initial network model is iterated on by a preset optimization model based on the first loss value to adjust the weights of the hidden layers in the first initial network model, so that the first loss value between the first output information generated by the first initial network model and the first reference information is within a preset first threshold range, and the iterated first initial network model is set as the first-stage network model.

[0028] In the above scheme, the second-stage network model is obtained by training a pre-set second initial network model using the second training samples and the first-stage network model, including:

[0029] Using the training ground-penetrating radar data from the training data in the second training sample as the second input information, the first-stage network model is run to perform inversion operations on the second input information to obtain the first-stage output information, and the second initial network model is run to perform inversion operations on the first-stage output information to obtain the second output information.

[0030] The training dielectric constant model of the training data in the second training sample is used as the second reference information of the second initial network model. A second loss value is generated based on the second output information and the second reference information using a preset second loss function. The second loss value characterizes the degree of difference between the second output information and the second reference information.

[0031] The second initial network model is iterated on according to the second loss value by a preset optimization model to adjust the weights of the hidden layers in the second initial network model, so that the second loss value between the second output information generated by the second initial network model and the second reference information is within a preset second threshold range, and the iterated second initial network model is set as the two-stage network model.

[0032] Secondly, this application provides a ground-penetrating radar data inversion device, comprising:

[0033] An input module is used to receive ground-penetrating radar data; wherein, the ground-penetrating radar data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of the target area;

[0034] The first inversion module is used to call a preset one-stage network model to perform inversion calculations on the ground penetrating radar data to obtain a first dielectric constant model; wherein, the first dielectric constant model is a factor that determines the propagation speed of the electromagnetic wave in the geological structure of the target area; the first dielectric constant model has at least one relative dielectric constant; the relative dielectric constant is a physical parameter characterizing the dielectric properties of the geological structure;

[0035] The second inversion module is used to call a preset two-stage network model to perform inversion operations on the ground penetrating radar data and the first dielectric constant model to obtain a second dielectric constant model; wherein, the second dielectric constant model includes physical parameters after correcting the relative dielectric constant in the first dielectric constant model based on the ground penetrating radar data.

[0036] Thirdly, this application provides a computer device, including: a processor and a memory communicatively connected to the processor;

[0037] The memory stores computer-executed instructions;

[0038] The processor executes the computer execution instructions stored in the memory to implement the ground-penetrating radar data inversion method as described above.

[0039] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the above-described ground-penetrating radar data inversion method.

[0040] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described ground-penetrating radar data inversion method.

[0041] This application provides a ground-penetrating radar (GPR) data inversion method, apparatus, computer equipment, and storage medium. The GPR data is inverted by calling a preset one-stage network model. By extracting feature values ​​from the GPR data that have a sensitive relationship with the dielectric constant model, a first dielectric constant model is obtained based on the feature values, thereby ensuring the accuracy of the first dielectric constant model.

[0042] By calling a pre-set two-stage network model, the ground-penetrating radar data and the first dielectric constant model are inverted to modify the erroneous structural features in the first dielectric constant model and obtain a more accurate second dielectric constant model. This results in the second dielectric constant model being derived based on the first dielectric constant model.

[0043] By using a network model to perform inversion operations on the ground-penetrating radar data, the computational efficiency of ground-penetrating radar data inversion is greatly improved, and the consumption of computing resources is reduced. Attached Figure Description

[0044] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0045] Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of this application;

[0046] Figure 2 A flowchart of Embodiment 1 of a ground-penetrating radar data inversion method provided in this application;

[0047] Figure 3 An image of the second dielectric constant model in a ground-penetrating radar data inversion method provided in this application embodiment;

[0048] Figure 4 A flowchart of Embodiment 2 of a ground-penetrating radar data inversion method provided in this application;

[0049] Figure 5 In Embodiment 2 of a ground-penetrating radar data inversion method provided for the present application, the image of the training dielectric constant model is shown.

[0050] Figure 6 In Embodiment 2 of a ground-penetrating radar data inversion method provided for the present application, the image of the training ground-penetrating radar data is included;

[0051] Figure 7 A schematic diagram of the structure of the first initial network model in Embodiment 2 of a ground-penetrating radar data inversion method provided in this application;

[0052] Figure 8 This is a graph showing the relationship between the initial loss value and the number of iterations during the training process of the first initial network model.

[0053] Figure 9 A schematic diagram of the structure of the second initial network model in Embodiment 2 of a ground-penetrating radar data inversion method provided in this application;

[0054] Figure 10 This is a graph showing the relationship between the second loss value and the number of iterations during the training process of the second initial network model.

[0055] Figure 11 This is a schematic diagram of the program modules of a ground-penetrating radar data inversion device provided by the present invention;

[0056] Figure 12 This is a schematic diagram of the hardware structure of the computer device in the computer device of the present invention.

[0057] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0058] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0059] The specific application scenario for this application is as follows:

[0060] Ground penetrating radar (GPR) is a shallow geophysical exploration technique that uses electromagnetic waves for non-destructive testing. It has been widely applied in many fields, including glaciology, archaeology, and geotechnical engineering. GPR can convert electromagnetic information in the subsurface medium into information related to the characteristics of the geological medium (such as location, shape, and dielectric properties), which is very important for shallow geological exploration and analysis.

[0061] Existing GPR inversion methods focus on roughly inferring the location, size, and dimensions of the target body based on observed GPR data. Traditional full waveform inversion (FWI) of GPR is considered a solution for qualitative and quantitative reconstruction of stratigraphic images. It directly uses the entire received waveform to match simulated GPR data. Then, it reconstructs the dielectric distribution of the structure by minimizing the mismatch between the two sets of data. FWI originated in seismic exploration and was subsequently rapidly applied to processing radar data. However, because actual stratigraphic structures always have irregular geometric features and complex distribution patterns, received ground-penetrating radar data is generally interlaced and cluttered, accompanied by discontinuous and distorted echoes. Furthermore, multiple reflections caused by anomalies in the subsurface medium can also obscure the characteristics of the radar signal, often resulting in images with chaotic signal features. On the other hand, traditional FWI typically uses an iterative approach to reduce the error between simulated and ground-penetrating radar data; therefore, completing a single FWI operation requires significant computational resources. Thus, traditional FWI not only consumes substantial computational resources, but the accuracy of its inversion model also has considerable room for improvement.

[0062] In recent years, deep neural networks (DNNs) have seen rapid development in seismic denoising, signal processing, and geophysical inversion. DNNs automatically learn high-level features from training data and can then estimate the nonlinear mapping between input image data and various data domains. With the rapid development of deep learning technology, the breadth and depth of intelligent inversion work in the geosciences are constantly expanding. Convolutional neural networks (CNNs) are used to predict high-resolution impedance. Li et al. achieved super-resolution velocity image prediction using a multi-task learning approach. Two-dimensional inversion imaging of GPR and seismic data was achieved through end-to-end learning.

[0063] While some progress has been made in DNN-based GPR inversion, there is still considerable room for improvement. The challenge in reconstructing the complex electrical properties of subsurface media lies in extracting effective features from complex GPR data while preserving the spatial alignment between input and output.

[0064] This application proposes an incremental learning approach to predict underground dielectric constant models. A one-stage network extracts an initial dielectric constant model from GPR data in an end-to-end manner. Then, a dual-channel two-stage network model is built, treating the initial dielectric constant model as prior information and combining it with GPR data as input for inversion prediction. The one-stage network can extract GPR signal features from adjacent trajectories, while the two-stage network can effectively correct erroneous structural features in the dielectric constant model. This method, combining the idea of ​​incremental learning, proposes a novel algorithm for the learning mechanism in GPR inversion tasks.

[0065] Specifically, please refer to Figure 1 This application proposes a server 2 running a ground-penetrating radar (GPR) data inversion method, which receives GPR data from a GPR device 3. The GPR data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of a target area. A pre-set one-stage network model is invoked to perform inversion calculations on the GPR data to obtain a first dielectric constant model. The first dielectric constant model is a factor determining the propagation speed of electromagnetic waves in the geological structure of the target area. The first dielectric constant model has at least one relative dielectric constant; the relative dielectric constant is a physical parameter characterizing the dielectric properties of the geological structure.

[0066] The system calls a pre-set two-stage network model to perform inversion operations on the ground-penetrating radar data and the first dielectric constant model to obtain the second dielectric constant model. The second dielectric constant model includes physical parameters that are corrected based on the relative dielectric constant in the first dielectric constant model according to the ground-penetrating radar data.

[0067] Therefore, this application uses an incremental learning-based approach to achieve GPR inversion. GPR inversion is an ill-conditioned problem of solving the correspondence between ground-penetrating radar data and the dielectric constant model space. This application achieves the GPR inversion task by learning new knowledge from GPR data multiple times. The inversion method used in this application not only effectively improves the accuracy of GPR inversion but also has the ability to suppress false anomalies and restore deep stratigraphic structures.

[0068] This application achieves accurate analysis of ground-penetrating radar data through a one-stage network model and a two-stage network model, and obtains a second dielectric constant model with a relative dielectric constant that can accurately characterize the dielectric properties of the stratigraphic structure. The dielectric properties refer to the geological information of the stratigraphic structure's storage and loss of electrostatic energy under the action of an electric field.

[0069] Therefore, based on the second dielectric constant model, the characteristic values ​​of dielectric properties of ground-penetrating radar data are analyzed and extracted to obtain geological information, thereby achieving accurate analysis and extraction of characteristic values ​​of ground-penetrating radar data.

[0070] The technical solution of this application and how the technical solution of this application solves the problems of the prior art will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0071] Example 1:

[0072] Please see Figure 2 This application provides a ground-penetrating radar data inversion method, including:

[0073] S101: Receive ground-penetrating radar data; wherein, ground-penetrating radar data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of the target area.

[0074] In this step, Ground Penetrating Radar (GPR) is a geophysical method that uses antennas to transmit and receive high-frequency electromagnetic waves to detect the internal material properties and distribution patterns of a medium. GPR data refers to shallow surface geophysical exploration data, a non-destructive exploration technique that utilizes electromagnetic waves. It has been widely applied in many fields, including glaciology, archaeology, and geotechnical engineering. GPR data can convert electromagnetic information in the subsurface medium into information related to the characteristics of the geological medium (such as location, shape, and dielectric properties), which is crucial for shallow surface geological exploration and analysis.

[0075] S102: Call the preset one-stage network model to perform inversion calculation on the ground penetrating radar data to obtain the first dielectric constant model; wherein, the first dielectric constant model is the factor that determines the propagation speed of electromagnetic waves in the geological structure of the target area; the first dielectric constant model has at least one relative dielectric constant; the relative dielectric constant is a physical parameter characterizing the dielectric properties of the geological structure.

[0076] In this example, a first dielectric constant model is obtained by inverting ground-penetrating radar (GPR) data using a one-stage network model. This allows for the prediction of the stratigraphic structure of a target area based on GPR data, and the generation of the first dielectric constant model characterizing this stratigraphic structure. Specifically, the one-stage network model extracts the first dielectric constant model from the GPR data in an end-to-end manner.

[0077] The relative permittivity is a physical parameter characterizing the dielectric properties of a dielectric material. Dielectric properties refer to the ability of a material to store and lose electrostatic energy under the influence of an electric field, and are usually expressed by the permittivity and dielectric loss.

[0078] Optionally, the one-stage network model can extract the signal features of GPR from adjacent trajectories, and then obtain the first dielectric constant model through these signal features.

[0079] In this embodiment, the first-stage network model extracts feature values ​​that characterize the dielectric properties of the stratum structure from ground-penetrating radar data, performs inversion calculations based on these feature values ​​to obtain at least one relative permittivity, and generates a first permittivity model based on the at least one relative permittivity to ensure the accuracy of the first node constant model.

[0080] S103: Call the preset two-stage network model to perform inversion operation on the ground penetrating radar data and the first dielectric constant model to obtain the second dielectric constant model; wherein, the second dielectric constant model includes physical parameters after correcting the relative dielectric constant in the first dielectric constant model based on the ground penetrating radar data.

[0081] In this example, a dual-channel, two-stage network model is constructed. The first dielectric constant model is treated as prior information, and both the first dielectric constant model and ground-penetrating radar (GPR) data are used as inputs to the second-stage network model. This allows the second-stage network model to adjust the first dielectric constant model based on the GPR data (i.e., the original waveform of the shallow surface of the target area represented by GPR data) to achieve inversion prediction. The second-stage network model effectively corrects erroneous structural features in the first dielectric constant model, resulting in a more accurate second dielectric constant model. This method, combining the idea of ​​incremental learning, proposes a new algorithm for the learning mechanism in GPR data inversion tasks. The second dielectric constant model is as follows: Figure 3 As shown, where, Figure 3 The vertical axis on the left represents the depth of the geological structure, the horizontal axis represents the location of the target area, and the vertical axis on the right represents the relative permittivity of the second permittivity model.

[0082] Specifically, this method uses incremental learning to predict the subsurface dielectric constant model. A one-stage network model extracts the first dielectric constant model from GPR data in an end-to-end manner. Then, a dual-channel two-stage network model is built, treating the first dielectric constant model as prior information and combining it with GPR data as input for inversion prediction. The one-stage network model can extract GPR signal features from adjacent trajectories, while the two-stage network model can effectively correct erroneous structural features in the dielectric constant model. This method, combining the idea of ​​incremental learning, proposes a new algorithm for the learning mechanism in ground penetrating radar data inversion tasks.

[0083] This application employs an incremental learning-based approach to achieve ground-penetrating radar (GPR) data inversion. GPR data inversion is an ill-conditioned problem involving the correspondence between GPR data and the dielectric constant model space. This application achieves the GPR data inversion task by repeatedly learning new knowledge from GPR data. The inversion method used in this application not only effectively improves the accuracy of GPR data inversion but also has the ability to suppress false anomalies and recover deep stratigraphic structures.

[0084] Therefore, the prediction results of the first-stage network model are used as prior information for the second-stage network model to constrain the inversion process. Compared with the traditional FWI algorithm, the advantage of this application lies in effectively improving the inversion accuracy while ensuring appropriate computational efficiency. Compared with current deep learning-based ground-penetrating radar data inversion algorithms, the advantage of this application lies in the better agreement between the prediction results and model parameters, and its ability to accurately reflect the structural distribution of deep layers.

[0085] In this embodiment, the first-stage network model extracts feature values ​​from ground-penetrating radar data that characterize the dielectric properties of the strata structure, and creates a first dielectric constant model based on these feature values ​​to ensure the accuracy of the first node constant model.

[0086] The two-stage network-free model compares the first dielectric constant model with all the eigenvalues ​​in the ground-penetrating radar data to further modify the relative permittivity in the first dielectric constant model, resulting in a second dielectric constant model. This makes the relative permittivity in the second dielectric constant model more closely match the ground-penetrating radar data, thereby ensuring the accuracy of each relative permittivity in the second dielectric constant model.

[0087] Meanwhile, since this embodiment calls the first-stage network model and the second-stage network model to directly perform calculations on the ground penetrating radar data, it does not require multiple iterations to reduce the error between the simulated data and the ground penetrating radar data, as is the case with the existing FWI method. This avoids the situation where completing one FWI operation requires a lot of computing resources. Therefore, compared with the existing technology, this application greatly reduces the consumption of computing resources.

[0088] Example 2:

[0089] Please see Figure 4 This application provides a ground-penetrating radar data inversion method, including:

[0090] S201: Receive ground-penetrating radar data; wherein, ground-penetrating radar data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of the target area.

[0091] This step is the same as S101 in Example 1, so it will not be described again here.

[0092] S202: Obtain the stratigraphic structure, embed irregular blocks into the stratigraphic structure to convert the stratigraphic structure into a training dielectric constant model, and perform forward modeling based on the training dielectric constant model to obtain training ground penetrating radar data;

[0093] A training dielectric constant model and its training ground-penetrating radar data are combined to form a training dataset.

[0094] A training set is obtained by summing up several training data sets.

[0095] In this example, training data similar to the characteristics of actual ground-penetrating radar (GPR) data scenarios is generated to effectively improve the network's ability to solve practical problems, such as... Figure 5 As shown, the training dielectric constant model is randomly generated based on exploration conditions and mathematical methods. The training dielectric constant model includes a randomly simulated undulating formation structure, with the dielectric parameter gradually increasing from top to bottom. The random equivalent medium technique is used to characterize the inhomogeneity of the formation medium. At the same time, abrasive technology is combined to simulate irregular blocks embedded in the formation. The medium parameters in the model are randomly generated within a reasonable range.

[0096] Local random anomalies are considered. This application employs a non-uniform hybrid function-based stochastic medium modeling method to simulate stratigraphic structure, where the function is expressed by the following formula:

[0097]

[0098] Where r represents the fuzziness factor, and a, b, and c represent the autocorrelation lengths in the x, y, and z directions, respectively. By selecting the local perturbation radius (a, b, c) and the local perturbation intensity (r), various training dielectric constant models can be constructed, thereby achieving diversity in the training dielectric constant models.

[0099] After obtaining the training dielectric constant model, the forward modeling of the training ground-penetrating radar (GPR) data is performed using the FDTD algorithm, such as... Figure 6 As shown. The dielectric constant model size is 4.5m × 4m, the cell mesh size is 0.025m × 0.025m, the sampling time interval is 0.0201ns, and the main frequency of the transmitting antenna is 650MHz. The sample set used in this application contains 1000 pairs of ground penetrating radar data and dielectric constant models.

[0100] It should be noted that the stochastic equivalent medium technique simulates the ore deposit structure and embeds it into a stochastic undulating stratigraphic structure to obtain a geological model containing the ore deposit structure. Based on existing geological data, the geological model is transformed into a resistivity model, and a sample dataset is obtained based on the resistivity model. The sample dataset includes data on the stochastically simulated undulating stratigraphic structure, ore deposit structure, and corresponding resistivity data. The initial resistivity model reconstruction network is then trained based on the sample dataset to obtain the resistivity model reconstruction network. The resistivity model reconstruction network is used for deep learning on the first electromagnetic inversion data to obtain the second electromagnetic inversion data.

[0101] Abrasive technology is a computer algorithm that generates a three-dimensional array based on MATLAB or Python, and then generates irregular shapes based on that three-dimensional array.

[0102] The finite-difference time-domain (FDTD) algorithm is a commonly used method in the field of electromagnetic field calculations. The FDTD model is based on Maxwell's equations, the most fundamental equations in electrodynamics. Since its inception, the FDTD method has seen significant development with the advancement of computing technology, especially electronic computers, and has been widely applied in fields such as electromagnetism, electronics, and optics.

[0103] S203: Obtain the first training sample, and train the first initial network model preset by the first training sample to obtain the first-stage network model.

[0104] In a preferred embodiment, obtaining the first training sample includes:

[0105] Obtain M training data points from the training set, and summarize the M training data points to obtain the first training sample; where M is a positive integer, M≥1.

[0106] In this example, by using a preset quantity M, M training data points are obtained from the training set and aggregated to obtain the first training sample, ensuring the controllability of the number of training data points in the sample.

[0107] In a preferred embodiment, a one-stage network model is obtained by training a first initial network model pre-set using the first training samples, including:

[0108] The training ground-penetrating radar data in the first training sample is used as the first input information of the first initial network model, and the first output information is obtained by running the first initial network model to perform inversion operation on the first input information.

[0109] The training dielectric constant model of the training data in the first training sample is used as the first reference information of the first initial network model. A first loss value is generated based on the first output information and the first reference information through a preset first loss function. The first loss value characterizes the degree of difference between the first output information and the first reference information.

[0110] The first initial network model is iterated on by a preset optimization model based on the first loss value to adjust the weights of the hidden layers in the first initial network model, so that the first loss value between the first output information generated by the first initial network model and the first reference information is within a preset first threshold range, and the iterated first initial network model is set as a one-stage network model.

[0111] In this example, a deep neural network is used as the first initial network model, as shown in the example below. Figure 7 As shown, the first initial network model uses a downsampling encoding / decoding structure. The network performs four downsampling operations, implemented by convolutional layers with a "stride" parameter of 2. The network contains four sets of feature maps at different scales, with a size ratio of 8:4:2:1. Similar to the encoding layer, the decoding layer also consists of four repeating structures. Each repeating structure is preceded by a deconvolution, and after each deconvolution, the number of feature channels is halved, while the size of the feature map doubles. After deconvolution, the result of the deconvolution is concatenated with the feature map from the corresponding step in the encoding part. The last layer uses a 1x1 convolutional kernel, transforming the 64-channel feature map into a result with a specific number of categories. The input to the first-stage network model is a single channel (GPR data), and the input to the second-stage network model is a dual channel (the output of the first-stage network model and GPR data). The output of both networks is a dielectric constant model.

[0112] The first initial network model uses Mean Squared Error (MSE) as the loss function for the deep neural network and is trained using the Adam optimizer with a decaying learning rate. After inputting training data, the number of training iterations, the loss function value, and the learning rate are recorded, and the network parameters are stored in a specific file. The prediction model of the first-stage network model is then integrated with the GPR data and used in the training of the second-stage network model for prediction.

[0113] The expression for the first loss function is:

[0114]

[0115] Where loss_1 is the first loss value, y is the first output information, r is the first reference information, and 1 / n represents the mean. Training data is input, and the Adam optimizer is used for training, with a decaying learning rate during the training process. Figure 8As shown, the vertical axis represents the first loss value, loss_1, and the horizontal and vertical axes represent the number of iterations (Epoch) of the first initial network model. Figure 8 This includes the training set curve formed by the first loss value of the training set in the first training sample during the iteration process, and the validation set curve formed by the first loss value of the validation set in the first training sample during the iteration process.

[0116] The learning rate is characterized by three parameters: the initial learning rate η0, the decay period T, and the decay rate α. The real-time learning rate expression during training is:

[0117] η i =α i η0

[0118] Where i represents the current learning rate decay count. The training iterations, loss function values, and learning rate are recorded in a text file, and the network parameters are stored in a specific file.

[0119] It's important to note that deep neural networks are a technique within the field of machine learning (ML). The advantage of multiple layers is that complex functions can be represented with fewer parameters. In supervised learning, a previous problem with multi-layer neural networks was their tendency to get trapped in local optima. However, if the training samples sufficiently cover future samples, the learned multi-layer weights can be effectively used to predict new test samples.

[0120] S204: Obtain the second training sample; use the second training sample and the first-stage network model to train the pre-set second initial network model to obtain the second-stage network model.

[0121] In a preferred embodiment, obtaining the second training sample includes:

[0122] Obtain N training data points from the training set and summarize the N training data points to obtain the second training sample; where N is a positive integer and N≥1.

[0123] In this example, by using a preset number N, N training data points are obtained from the training set and aggregated to obtain the first training sample, ensuring the controllability of the number of training data points in the sample.

[0124] Preferably, a two-stage network model is obtained by training a pre-set second initial network model using a second training sample and a first-stage network model, including:

[0125] The training ground-penetrating radar data in the second training sample is used as the second input information. The first-stage network model is run to perform inversion operation on the second input information to obtain the first-stage output information, and the second initial network model is run to perform inversion operation on the first-stage output information to obtain the second output information.

[0126] The training dielectric constant model of the training data in the second training sample is used as the second reference information of the second initial network model. A second loss value is generated based on the second output information and the second reference information through a preset second loss function. The second loss value characterizes the degree of difference between the second output information and the second reference information.

[0127] The second initial network model is iterated based on the second loss value using a pre-set optimization model to adjust the weights of the hidden layers in the second initial network model. This ensures that the second loss value between the second output information generated by the second initial network model and the second reference information is within a pre-set second threshold range. The iterated second initial network model is then set as a two-stage network model.

[0128] In this example, a deep neural network is used as the second initial network model, as shown in the example below. Figure 9 As shown, the second initial network model uses a downsampling encoding / decoding structure. The network performs four downsampling operations, implemented by convolutional layers with a "stride" parameter of 2. The network contains four sets of feature maps at different scales, with a size ratio of 8:4:2:1. Similar to the encoding layer, the decoding layer also consists of four repeating structures. Each repeating structure is preceded by a deconvolution, and after each deconvolution, the number of feature channels is halved, while the size of the feature map doubles. After deconvolution, the result of the deconvolution is concatenated with the feature map from the corresponding step in the encoding part. The final layer uses a 1x1 convolutional kernel, transforming the 64-channel feature map into a result with a specific number of categories. The input to the first-stage network model is a single channel (GPR data), and the input to the second-stage network model is a dual channel (the output of the first-stage network model and GPR data). The output of both networks is a dielectric constant model.

[0129] The second initial network model uses Mean Squared Error (MSE) as the loss function for the deep neural network and is trained using the Adam optimizer with a decaying learning rate. After inputting training data, the number of training iterations, the loss function value, and the learning rate are recorded, and the network parameters are stored in a specific file. The prediction model of the first-stage network model is then integrated with the GPR data and used to train the second-stage network model for prediction.

[0130] The expression for the second loss function is:

[0131]

[0132] Here, loss_2 is the second loss value, x is the second output information, r is the second reference information, and 1 / n represents the mean. Training data is input, and the Adam optimizer is used for training, with a decaying learning rate during the training process. For example... Figure 10As shown, the vertical axis represents the second loss value loss_2, and the horizontal and vertical axes represent the number of iterations of the second initial network model, Epoch.

[0133] The learning rate is characterized by three parameters: the initial learning rate η0, the decay period T, and the decay rate α. The real-time learning rate expression during training is:

[0134] η i =α i η0

[0135] Where i represents the current learning rate decay count. The training iterations, loss function values, and learning rate are recorded in a text file, and the network parameters are stored in a specific file.

[0136] It's important to note that deep neural networks are a technique within the field of machine learning (ML). The advantage of multiple layers is that complex functions can be represented with fewer parameters. In supervised learning, a previous problem with multi-layer neural networks was their tendency to get trapped in local optima. However, if the training samples sufficiently cover future samples, the learned multi-layer weights can be effectively used to predict new test samples.

[0137] S205: Call the preset one-stage network model to perform inversion calculation on the ground penetrating radar data to obtain the first dielectric constant model; wherein, the first dielectric constant model is the factor that determines the propagation speed of electromagnetic waves in the geological structure of the target area; the first dielectric constant model has at least one relative dielectric constant; the relative dielectric constant is a physical parameter characterizing the dielectric properties of the geological structure.

[0138] This step is the same as S102 in Example 1, so it will not be described again here.

[0139] S206: Call the preset two-stage network model to perform inversion operation on the ground penetrating radar data and the first dielectric constant model to obtain the second dielectric constant model; wherein, the second dielectric constant model includes physical parameters after correcting the relative dielectric constant in the first dielectric constant model based on the ground penetrating radar data.

[0140] This step is the same as S103 in Example 1, so it will not be described again here.

[0141] An experimental result based on the above technical solution includes:

[0142] In evaluating the inversion algorithm used in this invention, the prediction results of the one-stage network model and the two-stage network model were compared. Compared to the prediction results of the one-stage network, the two-stage network has higher accuracy for detecting deep layers and anomalous targets. The GPR inversion method used in this invention draws on the idea of ​​incremental learning, requiring the dielectric constant predicted by the one-stage network as a constraint, which improves the prediction accuracy of the two-stage network for the dielectric model. From the indicators of MSE, PSNR, and SSIM, the two-stage network prediction model has significant improvements compared to the one-stage network.

[0143] The quantitative comparison of the test results is shown in the table below:

[0144] algorithm MSE↓ PSNR↑ SSIM↑ Phase 1 network 0.1719 55.7793 0.9957 Two-stage network 0.0859 58.7929 0.9979

[0145] MSE, or mean-square error, is a measure of the difference between the estimator and the estimated quantity. Let t be an estimator of the population parameter θ determined from the sample; the expected value of (θ-t)² is called the mean-square error of the estimator t.

[0146] PSNR: Peak signal-to-noise ratio (PSNR) is an engineering term that represents the ratio of the maximum possible power of a signal to the power of destructive noise that affects its representation accuracy.

[0147] SSIM (Structural Similarity) is a metric for measuring the similarity between two images. It was first proposed by the Laboratory for Image and Video Engineering at the University of Texas at Austin. SSIM uses two images: one is an uncompressed, distortion-free image, and the other is a distorted image.

[0148] Example 3:

[0149] Please see Figure 11 This application provides a ground-penetrating radar data inversion device 1, comprising:

[0150] Input module 11 is used to receive ground penetrating radar data; wherein, ground penetrating radar data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of the target area;

[0151] The first inversion module 15 is used to call a preset one-stage network model to perform inversion calculations on the ground penetrating radar data to obtain a first dielectric constant model. The first dielectric constant model is a factor that determines the propagation speed of electromagnetic waves in the geological structure of the target area. The first dielectric constant model has at least one relative dielectric constant. The relative dielectric constant is a physical parameter that characterizes the dielectric properties of the geological structure.

[0152] The second inversion module 16 is used to call a preset two-stage network model to perform inversion calculations on the ground penetrating radar data and the first dielectric constant model to obtain the second dielectric constant model; wherein, the second dielectric constant model includes physical parameters after correcting the relative dielectric constant in the first dielectric constant model based on the ground penetrating radar data.

[0153] Optionally, the ground-penetrating radar data inversion device 1 also includes:

[0154] The sample construction module 12 is used to obtain the stratigraphic structure, embed irregular blocks in the stratigraphic structure to convert the stratigraphic structure into a training dielectric constant model, and perform forward modeling based on the training dielectric constant model to obtain training ground penetrating radar data; a training dielectric constant model and its training ground penetrating radar data are combined to form a training data set; and several training data sets are combined to obtain a training set.

[0155] The first training module 13 is used to acquire the first training sample and train the first initial network model preset by the first training sample to obtain a first-stage network model.

[0156] The second training module 14 is used to acquire the second training samples; and to train the pre-set second initial network model using the second training samples and the first-stage network model to obtain the second-stage network model.

[0157] Example 4:

[0158] To achieve the above objectives, this application also provides a computer device 4, including: a processor 42 and a memory 41 communicatively connected to the processor 42; the memory stores computer-executed instructions;

[0159] The processor executes computer execution instructions stored in memory 41 to implement the aforementioned ground-penetrating radar data inversion method. The components of the ground-penetrating radar data inversion device can be distributed across different computer devices. Computer device 4 can be a smartphone, tablet, laptop, desktop computer, rack server, blade server, tower server, or cabinet server (including standalone servers or server clusters composed of multiple application servers), etc. The computer device in this embodiment includes, but is not limited to, memory 41 and processor 42, which can be interconnected via a system bus. Figure 12As shown. It should be noted that, Figure 12 Only computer devices with components are shown; however, it should be understood that it is not required to implement all of the shown components, and more or fewer components may be implemented instead. In this embodiment, memory 41 (i.e., readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, memory 41 may be an internal storage unit of the computer device, such as the hard disk or RAM of the computer device. In other embodiments, memory 41 may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Of course, memory 41 may also include both internal storage units and external storage devices of the computer device. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device, such as the program code of the ground-penetrating radar data inversion device in Embodiment 3. Furthermore, the memory 41 can also be used to temporarily store various types of data that have been output or will be output. The processor 42 in some embodiments may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device. In this embodiment, the processor 42 is used to run the program code stored in the memory 41 or process data, for example, to run the ground-penetrating radar data inversion device to implement the ground-penetrating radar data inversion method of the above embodiments.

[0160] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application. It should be understood that the processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in this application can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor. The memory may include high-speed RAM, and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk, or optical disk, etc.

[0161] To achieve the above objectives, this application also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, App application store, etc., which stores computer-executable instructions. When the program is executed by processor 42, it implements the corresponding function. The computer-readable storage medium of this embodiment is used to store computer-executable instructions for implementing the ground-penetrating radar data inversion method, and when executed by processor 42, it implements the ground-penetrating radar data inversion method of the above embodiment.

[0162] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0163] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.

[0164] This application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described ground-penetrating radar data inversion method.

[0165] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0166] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0167] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for inverting ground-penetrating radar data, characterized in that, include: Receive ground-penetrating radar data; wherein, the ground-penetrating radar data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of the target area; A pre-defined one-stage network model is invoked to perform inversion operations on the ground-penetrating radar data to obtain a first dielectric constant model. This first dielectric constant model is a factor determining the propagation speed of electromagnetic waves within the geological structure of the target area. The first dielectric constant model has at least one relative dielectric constant, which is a physical parameter characterizing the dielectric properties of the geological structure. The input to the one-stage network model is a single channel. The model extracts feature values ​​characterizing the dielectric properties of the geological structure from the ground-penetrating radar data and creates the first dielectric constant model based on these feature values. A pre-defined two-stage network model is invoked to perform an inversion operation on the ground-penetrating radar data and the first dielectric constant model to obtain a second dielectric constant model. The second dielectric constant model includes physical parameters after correcting the relative dielectric constant in the first dielectric constant model based on the ground-penetrating radar data. The input of the two-stage network model is dual-channel. By comparing all feature values ​​in the first dielectric constant model and the ground-penetrating radar data, the relative dielectric constant in the first dielectric constant model is modified to obtain the second dielectric constant model.

2. The ground-penetrating radar data inversion method according to claim 1, characterized in that, Before invoking a pre-set one-stage network model to perform inversion calculations on the ground-penetrating radar data, the method further includes: Obtain the first training sample; The first-stage network model is obtained by training the first initial network model pre-set using the first training sample.

3. The ground-penetrating radar data inversion method according to claim 2, characterized in that, Before invoking the preset two-stage network model to perform inversion calculations on the ground-penetrating radar data and the first dielectric constant model, the method further includes: Obtain the second training sample; The second-stage network model is obtained by training the preset second initial network model using the second training samples and the first-stage network model.

4. The ground-penetrating radar data inversion method according to claim 3, characterized in that, Before obtaining the first training sample, the method further includes: The geological structure is obtained, and irregular blocks are embedded in the geological structure to convert the geological structure into a training dielectric constant model. Forward modeling is then performed based on the training dielectric constant model to obtain training ground penetrating radar data. A training dataset is formed by combining the aforementioned training dielectric constant model with the aforementioned training ground-penetrating radar data. The training set is obtained by summing up several sets of training data.

5. The ground-penetrating radar data inversion method according to claim 4, characterized in that, The process of obtaining the first training sample includes: M training data points are obtained from the training set, and the M training data points are aggregated to obtain the first training sample; where M is a positive integer, M≥1; The process of obtaining the second training sample includes: N training data points are obtained from the training set, and the N training data points are aggregated to obtain the second training sample; where N is a positive integer and N≥1.

6. The ground-penetrating radar data inversion method according to claim 2, characterized in that, The process of training the first-stage network model using the first initial network model pre-set from the first training samples includes: The training ground-penetrating radar data in the first training sample is used as the first input information of the first initial network model, and the first initial network model is run to perform inversion operation on the first input information to obtain the first output information. The training dielectric constant model of the training data in the first training sample is used as the first reference information of the first initial network model. A first loss value is generated based on the first output information and the first reference information using a preset first loss function. The first loss value characterizes the degree of difference between the first output information and the first reference information. The first initial network model is iterated on by a preset optimization model based on the first loss value to adjust the weights of the hidden layers in the first initial network model, so that the first loss value between the first output information generated by the first initial network model and the first reference information is within a preset first threshold range, and the iterated first initial network model is set as the first-stage network model.

7. The ground-penetrating radar data inversion method according to claim 3, characterized in that, The second-stage network model is obtained by training a pre-set second initial network model using the second training samples and the first-stage network model, including: Using the training ground-penetrating radar data from the training data in the second training sample as the second input information, the first-stage network model is run to perform inversion operations on the second input information to obtain the first-stage output information, and the second initial network model is run to perform inversion operations on the first-stage output information to obtain the second output information. The training dielectric constant model of the training data in the second training sample is used as the second reference information of the second initial network model. A second loss value is generated based on the second output information and the second reference information using a preset second loss function. The second loss value characterizes the degree of difference between the second output information and the second reference information. The second initial network model is iterated on according to the second loss value by a preset optimization model to adjust the weights of the hidden layers in the second initial network model, so that the second loss value between the second output information generated by the second initial network model and the second reference information is within a preset second threshold range, and the iterated second initial network model is set as the two-stage network model.

8. A ground-penetrating radar data inversion device, characterized in that, include: An input module is used to receive ground-penetrating radar data; wherein, the ground-penetrating radar data is waveform data of electromagnetic waves used for non-destructive detection of the shallow surface of the target area; The first inversion module is used to call a preset one-stage network model to perform inversion calculations on the ground penetrating radar data to obtain a first dielectric constant model. The first dielectric constant model is a factor that determines the propagation speed of electromagnetic waves in the geological structure of the target area. The first dielectric constant model has at least one relative dielectric constant. The relative dielectric constant is a physical parameter characterizing the dielectric properties of the geological structure. The input of the one-stage network model is a single channel. It extracts feature values ​​characterizing the dielectric properties of the geological structure from the ground penetrating radar data and creates the first dielectric constant model based on these feature values. The second inversion module is used to call a preset two-stage network model to perform inversion operations on the ground-penetrating radar data and the first dielectric constant model to obtain a second dielectric constant model. The second dielectric constant model includes physical parameters after correcting the relative dielectric constant in the first dielectric constant model based on the ground-penetrating radar data. The input of the two-stage network model is dual-channel. By comparing all feature values ​​in the first dielectric constant model and the ground-penetrating radar data, the relative dielectric constant in the first dielectric constant model is modified to obtain the second dielectric constant model.

9. A computer device, characterized in that, include: A processor and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the ground-penetrating radar data inversion method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the ground-penetrating radar data inversion method as described in any one of claims 1 to 7.

Images