Ground Penetrating Radar Electric Field Prediction Method, Model, Device and Model Training Method

Abstract

The present application provides a ground penetrating radar electric field prediction method, model, device and model training method, which relates to the field of electromagnetics. The ground penetrating radar electric field prediction method includes: obtaining a preset parameter set characterizing the electrical properties of a target medium; performing feature encoding processing on the parameter set to obtain a first matrix; performing scaling processing on the parameters in the first matrix to obtain a second matrix; performing linear mapping processing on the parameters in the second matrix to obtain a third matrix with a preset dimension; and performing global average pooling processing on the parameters in the third matrix to obtain a predicted ground penetrating radar electric field vector. In this solution, obtaining the predicted ground penetrating radar electric field vector does not require multiple runs of iteration. Compared with the existing solutions for obtaining the predicted ground penetrating radar electric field vector, this solution takes less time to obtain the predicted ground penetrating radar electric field vector, and thus can improve the efficiency of subsequent quantification of the uncertainty in the simulation results based on the predicted ground penetrating radar electric field vector.

Description

Technical Field

[0001] The present application relates to the field of computers, and in particular, to a method, model, device, and model training method for predicting the ground penetrating radar electric field. Background Art

[0002] Ground penetrating radar is widely used in many fields such as soil analysis, civil engineering, and environmental engineering. It uses an antenna to emit electromagnetic pulses to a target in the underlying medium to obtain the characteristics and distribution laws of the target. Usually, the finite-difference time-domain method is used to explain the propagation and reflection process of waves in the ground penetrating radar system. However, the numerical simulation of the ground penetrating radar depends on a set of input parameters that affect its electromagnetic pulses, such as electromagnetic parameters such as the dielectric constant of the soil. However, in practical applications, these input parameters are usually unknown. Therefore, there is a lack of accurate prior knowledge of the input parameters required for numerical simulation, which leads to the uncertainty of the numerical simulation results.

[0003] Currently, in order to make the numerical simulation results conform to the actual situation, it is necessary to quantify the uncertainty in the simulation results. And quantifying the uncertainty in the simulation results requires first obtaining a predicted ground penetrating radar electric field vector, and the predicted ground penetrating radar electric field vector includes electric field components at multiple time sequence points. The prior art usually uses the Monte Carlo method to obtain the predicted ground penetrating radar electric field vector, and the Monte Carlo method requires multiple runs and iterations to converge, and it is difficult to use high-performance parallel computing devices for acceleration, resulting in a long time for obtaining the predicted ground penetrating radar electric field vector using the Monte Carlo method, and low efficiency in quantifying the uncertainty in the simulation results. Summary of the Invention

[0004] The present application provides a method, model, device, and model training method for predicting the ground penetrating radar electric field to solve the problem of the long time for obtaining the predicted ground penetrating radar electric field vector using the existing Monte Carlo method, resulting in low efficiency in quantifying the uncertainty in the simulation results.

[0005] In a first aspect, the present application provides a method for predicting the ground penetrating radar electric field, including: obtaining a preset parameter group characterizing the electrical properties of the target medium; performing feature encoding processing on the parameter group to obtain a first matrix; performing scaling processing on the parameters in the first matrix to obtain a second matrix; performing linear mapping processing on the parameters in the second matrix to obtain a third matrix with a preset dimension; performing global average pooling processing on the parameters in the third matrix to obtain a predicted ground penetrating radar electric field vector.

[0006] In the embodiments of the present application, by performing feature encoding processing on a parameter group, a first matrix is obtained, and then the parameters in the first matrix are scaled to obtain a second matrix; then, the parameters in the second matrix are linearly mapped to obtain a third matrix with a preset dimension; finally, global average pooling processing is performed on the parameters in the third matrix, and the predicted ground penetrating radar electric field vector can be obtained. In this solution, obtaining the predicted ground penetrating radar electric field vector does not require multiple running iterations. Compared with the existing solutions for obtaining the predicted ground penetrating radar electric field vector, this solution takes less time to obtain the predicted ground penetrating radar electric field vector, and thus can improve the efficiency of subsequent quantification of the uncertainty in the simulation results based on the predicted ground penetrating radar electric field vector. At the same time, compared with the prior art, the predicted ground penetrating radar electric field vector predicted by this solution is smoother and has higher accuracy.

[0007] Combined with the technical solution provided in the above first aspect, in some possible implementation manners, the scaling the parameters in the first matrix to obtain a second matrix includes: magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix to obtain the second matrix, where the first type of data represents the electric field characteristics of the ground penetrating radar, and the second type of data are the parameters in the first matrix other than the first type of data.

[0008] In the embodiments of the present application, by magnifying the parameters representing the electric field characteristics of the ground penetrating radar in the first matrix and reducing the parameters other than the first type of data in the first matrix, the obtained second matrix can more accurately represent the electric field characteristics of the ground penetrating radar.

[0009] Combined with the technical solution provided in the above first aspect, in some possible implementation manners, the magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix to obtain the second matrix includes: magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix to obtain an initial second matrix; magnifying the parameters belonging to the first type of data in the initial second matrix and reducing the parameters belonging to the second type of data in the initial second matrix to obtain the second matrix.

[0010] In the embodiments of the present application, by continuously performing two scaling processes on the first matrix, the parameters belonging to the first type of data in the finally obtained second matrix are larger, and the parameters belonging to the second type of data are smaller, so that the second matrix can more accurately represent the electric field characteristics of the ground penetrating radar.

[0011] Combined with the technical solution provided in the first aspect above, in some possible implementation manners, the step of magnifying the parameters of the first type of data in the first matrix and reducing the parameters of the second type of data in the first matrix to obtain an initial second matrix includes: performing feature scaling processing on the first matrix according to a plurality of preset weight matrix groups to obtain a first process matrix, wherein, according to the plurality of preset weight matrix groups, the parameters of the first type of data in the first matrix are magnified, and the parameters of the second type of data in the first matrix are reduced; adding the first matrix and the first process matrix, and performing matrix normalization processing on the obtained matrix after addition to obtain a second process matrix; performing feature fusion processing on the second process matrix to obtain a third process matrix; adding the second process matrix and the third process matrix, and performing normalization processing on the obtained matrix after addition to obtain the initial second matrix.

[0012] In the embodiment of the present application, by performing feature fusion processing on the second process matrix, features with different characteristics are fused. Since before that, the parameters of the first type of data in the first matrix have been magnified according to a plurality of preset weight matrix groups, and the parameters of the second type of data in the first matrix have been reduced, the second process matrix after feature fusion, that is, the third process matrix, can accurately represent the electric field characteristics of the ground penetrating radar, and further the finally obtained second matrix can more accurately represent the electric field characteristics of the ground penetrating radar.

[0013] Combined with the technical solution provided in the first aspect above, in some possible implementation manners, the step of performing feature scaling processing on the first matrix according to a plurality of preset weight matrix groups to obtain a first process matrix includes: for each weight matrix group in the plurality of preset weight matrix groups, based on the weight matrix group and the first matrix, obtaining a scoring matrix corresponding to the weight matrix group, wherein each preset weight matrix group includes a plurality of different weight matrices; processing the scoring matrix corresponding to each preset weight matrix group to obtain the first process matrix.

[0014] In the embodiment of the present application, each weight matrix group in the plurality of weight matrix groups processes the first matrix respectively, that is, each weight matrix group magnifies the parameters of the first type of data in the first matrix and reduces the parameters of the second type of data in the first matrix to obtain a scoring matrix, and then the first process matrix is obtained by using all the obtained scoring matrices. The first matrix is scaled from multiple aspects, so that the parameters of the first type of data in the first matrix are fully magnified, and the parameters of the second type of data are fully reduced, further improving the accuracy of the first process matrix in representing the electric field characteristics of the ground penetrating radar.

[0015] Combined with the technical solution provided in the first aspect above, in some possible implementation manners, each weight matrix group in the preset multiple weight matrix groups includes 3 weight matrices. For each weight matrix group in the preset multiple weight matrix groups, based on the weight matrix group and the first matrix, obtaining the scoring matrix corresponding to the weight matrix group includes: for each weight matrix group in the preset multiple weight matrix groups, multiplying the first matrix by each weight matrix in the weight matrix group respectively to obtain a first branch matrix, a second branch matrix, and a third branch matrix, where the second branch matrix and the third branch matrix are the same; obtaining a fusion matrix based on the first branch matrix and the second branch matrix; and obtaining the scoring matrix corresponding to the weight matrix group based on the fusion matrix and the third branch matrix.

[0016] In the embodiments of the present application, each group of weight matrix groups includes 3 weight matrices. By multiplying the first matrix by each weight matrix in the weight matrix group respectively, the parameters belonging to the first type of data in the first matrix are fully amplified, and the parameters belonging to the second type of data are fully reduced; furthermore, the scoring matrix obtained according to the first branch matrix, the second branch matrix, and the third branch matrix can accurately represent the electric field characteristics of the ground penetrating radar. Among them, the first branch matrix includes the information to be queried, the second branch matrix includes the query information, and the third branch matrix includes the query information. Therefore, the second branch matrix and the third branch matrix are the same. By the first branch matrix and the second branch matrix, a fusion matrix is obtained, and then the fusion matrix and the third branch matrix are multiplied. By processing the branch matrices including the query information twice in sequence, the obtained scoring matrix can more accurately represent the electric field characteristics of the ground penetrating radar.

[0017] Combined with the technical solution provided in the first aspect above, in some possible implementation manners, the parameter group includes at least one of the dielectric constant at infinite frequency, the static dielectric constant, the polar amplitude, the relaxation time, the static conductivity, the angular frequency, and the node constant of free space of the target medium.

[0018] In the embodiments of the present application, since the parameter group includes at least one of the dielectric constant at infinite frequency, the static dielectric constant, the polar amplitude, the relaxation time, the static conductivity, the angular frequency, and the node constant of free space of the target medium, all aspects of the electrical properties of the target medium are fully considered, and thus the final prediction result, that is, the predicted electric field vector of the ground penetrating radar, is more accurate and more in line with the requirements of practical applications.

[0019] Combined with the technical solution provided in the first aspect above, in some possible implementation manners, the parameter group is subjected to feature encoding processing through a preset ground penetrating radar electric field prediction model to obtain the first matrix; and / or, the parameters in the first matrix are scaled through the preset ground penetrating radar electric field prediction model to obtain the second matrix; and / or, the parameters in the second matrix are linearly mapped through the preset ground penetrating radar electric field prediction model to obtain the third matrix of a preset dimension; and / or, the parameters in the third matrix are subjected to global average pooling processing through the preset ground penetrating radar electric field prediction model to obtain the predicted ground penetrating radar electric field vector.

[0020] In the embodiments of the present application, the parameter group is processed through a preset ground penetrating radar electric field prediction model, and there is no need to perform separate processing on each step, so that the predicted ground penetrating radar electric field vector can be obtained more quickly and accurately.

[0021] Combined with the technical solution provided in the first aspect above, in some possible implementation manners, the method further includes: obtaining a training sample, where the training sample includes a parameter group representing the electrical properties of a target medium and the true ground penetrating radar electric field vector corresponding to the parameter group; and training an initial ground penetrating radar electric field prediction model by using the training sample to obtain the preset ground penetrating radar electric field prediction model.

[0022] In the embodiments of the present application, the model is trained by using a training sample including a parameter group representing the electrical properties of a target medium and the true ground penetrating radar electric field vector corresponding to the parameter group, so that the initial ground penetrating radar electric field prediction model can fully learn the relationship between the parameter group and the true ground penetrating radar electric field vector, and the result output by the finally obtained preset ground penetrating radar electric field prediction model is more accurate.

[0023] Combined with the technical solution provided in the first aspect above, in some possible implementation manners, the training of the initial ground penetrating radar electric field prediction model using the training samples to obtain the preset ground penetrating radar electric field prediction model includes: inputting the training samples into the initial ground penetrating radar electric field prediction model to obtain a predicted ground penetrating radar electric field vector; obtaining a first error, a second error, and a third error, where the first error is the sum of the squares of the differences between the electric field components corresponding to each time series point in the predicted ground penetrating radar electric field vector and the electric field components corresponding to the same time series point in the true ground penetrating radar electric field vector; the second error is the square of the difference between the average value of all the electric field components in the predicted ground penetrating radar electric field vector and the average value of all the electric field components in the true ground penetrating radar electric field vector; the third error is the square of the difference between the variance of all the electric field components in the predicted ground penetrating radar electric field vector and the variance of all the electric field components in the true ground penetrating radar electric field vector; obtaining a loss value based on the first error, the second error, and the third error; updating the model parameters of the initial ground penetrating radar electric field prediction model based on the loss value; and training the radar electric field prediction model with updated model parameters using the training samples again until a preset condition is satisfied to obtain the preset ground penetrating radar electric field prediction model.

[0024] In the embodiments of the present application, when obtaining the loss value, the variances and means of the predicted ground penetrating radar electric field vector and the true ground penetrating radar electric field vector are considered simultaneously, so that the obtained loss value can more accurately reflect the fitting degree of the initial ground penetrating radar electric field prediction model on the training samples, making the result output by the finally obtained ground penetrating radar electric field prediction model more accurate and ensuring a high consistency between the predicted ground penetrating radar electric field vector and the true ground penetrating radar electric field vector.

[0025] In a second aspect, the present application provides a model training method, including: obtaining training samples, where the training samples include a parameter set representing the electrical properties of a target medium and the true ground penetrating radar electric field vector corresponding to the parameter set; and training an initial ground penetrating radar electric field prediction model using the training samples to obtain a trained ground penetrating radar electric field prediction model.

[0026] In a third aspect, the present application provides a ground penetrating radar electric field prediction model, including: an encoding layer for performing feature encoding processing on a parameter set representing the electrical properties of a target medium to obtain a first matrix; a feature scaling layer for performing scaling processing on the first matrix to obtain a second matrix; a linear mapping layer for performing linear mapping processing on the second matrix to obtain a third matrix with a preset dimension; and a pooling layer for performing global average pooling processing on the third matrix to obtain a predicted ground penetrating radar electric field vector.

[0027] Combined with the technical solution provided in the above third aspect, in some possible implementation manners, the feature scaling layer includes a first sub-feature scaling layer and a second sub-feature scaling layer connected in sequence; wherein, the first sub-feature scaling layer is configured to amplify the parameters of the first type of data in the first matrix and shrink the parameters of the second type of data in the first matrix to obtain an initial second matrix, where the first type of data represents the ground penetrating radar electric field feature, and the second type of data represents other features except the ground penetrating radar electric field feature; the second sub-feature scaling layer is configured to amplify the parameters of the first type of data in the initial second matrix and shrink the parameters of the second type of data in the initial second matrix to obtain the second matrix.

[0028] Combined with the technical solution provided in the above third aspect, in some possible implementation manners, the first sub-feature scaling layer includes: a multi-branch attention layer configured to perform feature scaling processing on the first matrix to obtain a first process matrix, where the parameters of the first type of data in the first matrix are amplified and the parameters of the second type of data in the first matrix are shrunk according to a preset plurality of weight matrix groups; a first accumulation and normalization layer configured to perform matrix addition processing on the first matrix and the first process matrix, and perform matrix normalization processing on the first matrix and the first process matrix after the matrix addition processing to obtain a second process matrix; a multi-layer perceptron configured to perform feature fusion processing on the second process matrix to obtain a third process matrix; a second accumulation and normalization layer configured to perform matrix addition processing on the second process matrix and the third process matrix, and perform normalization processing on the second process matrix and the third process matrix after the matrix addition processing to obtain an initial second matrix.

[0029] Fourth aspect, the present application provides a ground penetrating radar electric field prediction device, including: an acquisition module configured to acquire a preset parameter group representing the electrical properties of a target medium; a processing module configured to perform feature encoding processing on the parameter group to obtain a first matrix; the processing module is further configured to perform scaling processing on the parameters in the first matrix to obtain a second matrix; the processing module is further configured to perform linear mapping processing on the parameters in the second matrix to obtain a third matrix with a preset dimension; the processing module is further configured to perform global average pooling processing on the parameters in the third matrix to obtain a predicted ground penetrating radar electric field vector.

[0030] Fifth aspect, an embodiment of the present application further provides an electronic device, including: a memory and a processor, the memory and the processor being connected; the memory for storing a program; the processor for calling the program stored in the memory to execute the method provided in the embodiment of the first aspect above and / or any possible implementation manner in combination with the embodiment of the first aspect above, or execute the method provided in the embodiment of the second aspect above.

[0031] Sixth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a computer, it executes the method provided in the embodiment of the first aspect above and / or any possible implementation manner in combination with the embodiment of the first aspect above, or executes the method provided in the embodiment of the second aspect above. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 It is a schematic flowchart of a ground penetrating radar electric field prediction method shown in an embodiment of the present application;

[0034] Figure 2 It is a schematic diagram of the principle of obtaining a first process matrix based on a first matrix shown in an embodiment of the present application;

[0035] Figure 3 It is a schematic diagram of the structure of a target medium shown in an embodiment of the present application;

[0036] Figure 4 It is a block diagram of a ground penetrating radar electric field prediction model shown in an embodiment of the present application;

[0037] Figure 5 It is a structural block diagram of a feature scaling layer shown in an embodiment of the present application;

[0038] Figure 6 It is a structural block diagram of a ground penetrating radar electric field prediction device shown in an embodiment of the present application;

[0039] Figure 7 It is a structural block diagram of an electronic device shown in an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

[0041] It should be noted that similar reference numerals and letters denote similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. At the same time, in the description of the present application, relational terms such as "first", "second", etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0042] Furthermore, the term "and / or" in the present application is only a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone.

[0043] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.

[0044] To solve the problems of the long time-consuming for obtaining the predicted ground penetrating radar electric field vector by using the Monte Carlo method in the prior art, and the low efficiency of quantifying the uncertainty in the simulation results, the present application provides a ground penetrating radar electric field prediction method, model, device and model training method, which can obtain the predicted ground penetrating radar electric field vector without multiple running iterations.

[0045] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of a ground penetrating radar electric field prediction method provided by an embodiment of the present application. The steps included therein will be described below with reference to Figure 1 .

[0046] S100: Obtain a parameter set preset to represent the electrical properties of the target medium.

[0047] The above-mentioned parameter set representing the electrical properties of the target medium can be pre-obtained and stored in a database, and can be directly called when needed. Alternatively, it can also be obtained in real time when needed.

[0048] Optionally, the parameter set characterizing the electrical properties of the target medium may include at least one of the dielectric constant at infinite frequency of the target medium, the electrostatic dielectric constant, the polarization amplitude, the relaxation time, the static conductivity, the angular frequency, and the nodal constant of free space.

[0049] For example, using ∈ ∞ to represent the dielectric constant at infinite frequency, ∈ s as the electrostatic dielectric constant, A p as the polarization amplitude, τ p as the relaxation time, σ s as the static conductivity. ω is the angular frequency, ∈ 0 is the nodal constant of free space. Then, in one implementation, the parameter set can be expressed as

[0050] wherein, the parameters in the parameter set characterizing the electrical properties of the target medium can be sampled from a preset interval by Latin hypercube sampling (LHS). The parameter can be any value within the preset interval and includes the two endpoints of the preset interval. The specific process of obtaining parameters from the preset interval using Latin hypercube sampling is well-known to those skilled in the art and will not be elaborated here for brevity.

[0051] The preset interval for each parameter can be obtained by first manually measuring the initial value and then presetting a fluctuation range to obtain the preset interval of the parameter. For example, if the dielectric constant at infinite frequency of the target medium measured manually is 1 and the preset fluctuation range is plus or minus 15%, then the interval of the dielectric constant at infinite frequency of the target medium can be set to 0.85 - 1.15. Accordingly, the dielectric constant at infinite frequency can be any value within 0.85 - 1.15, including 0.85 and 1.15. Another example, if the electrostatic dielectric constant of the target medium measured manually is 2.5 and the preset fluctuation range is plus or minus 13%, then the preset interval of the electrostatic dielectric constant of the target medium can be set to 2.175 - 2.825. Accordingly, the electrostatic dielectric constant of the target medium can be any value within 2.175 - 2.825, including 2.175 and 2.825.

[0052] The above-mentioned fluctuation range can be set according to actual needs. The examples here are only for easy understanding and should not be construed as a limitation to this application.

[0053] S200: Perform feature encoding processing on the parameter set to obtain the first matrix.

[0054] Among them, the feature embedding encoder can be used to perform feature encoding on the parameter set. By using feature embedding encoders with different numbers of neurons, first matrices with different dimensions can be obtained.

[0055] For example, the parameter group is Using a feature embedding encoder composed of a fully connected layer with 200 neurons to perform feature encoding, the first matrix X ∈ R 7×200 can be obtained. It can be understood that the specific type of the feature embedding encoder can be set according to actual needs. The example here is only for easy understanding and should not be regarded as a limitation to this application. Moreover, the process of the feature embedding encoder performing feature encoding on the input data is well-known to those skilled in the art. For the sake of brief description, it will not be elaborated here.

[0056] S300: Perform scaling processing on the parameters in the first matrix to obtain a second matrix.

[0057] In one implementation, the specific process of performing scaling processing on the parameters in the first matrix includes: magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix to obtain a second matrix. Among them, the first type of data represents the electric field characteristics of the ground penetrating radar, and the second type of data is the parameters in the first matrix except the first type of data. That is, magnify the meaningful parameters in the first matrix and reduce the unimportant parameters. So that the finally obtained predicted ground penetrating radar electric field vector is more accurate.

[0058] It can be understood that the first type of data can also be strongly correlated feature information beneficial to the prediction result, while the second type is weakly correlated or uncorrelated feature information with the prediction result.

[0059] In order to make the second matrix more accurately represent the electric field characteristics of the ground penetrating radar, in one implementation, the first matrix can be continuously scaled twice, so that the parameters belonging to the first type of data in the finally obtained second matrix are larger and the parameters belonging to the second type of data are smaller. At this time, the specific process of magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix to obtain a second matrix can be: magnify the parameters belonging to the first type of data in the first matrix and reduce the parameters belonging to the second type of data in the first matrix to obtain an initial second matrix. Then magnify the parameters belonging to the first type of data in the initial second matrix and reduce the parameters belonging to the second type of data in the initial second matrix to obtain the second matrix.

[0060] It can be understood that in order to make the second matrix more accurately represent the electric field characteristics of the ground penetrating radar, more scaling processes can also be performed. The more times of scaling processing, the higher the accuracy of the second matrix representing the electric field characteristics of the ground penetrating radar. No specific number of scaling times is limited here.

[0061] Among them, the specific process and calculation principle of magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix are the same as those of magnifying the parameters belonging to the first type of data in the initial second matrix and reducing the parameters belonging to the second type of data in the initial second matrix. To facilitate the understanding of the specific process of continuously performing two scaling processes on the first matrix, the following takes the example of magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix to obtain the initial second matrix for illustration.

[0062] Optionally, the specific process of magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix to obtain the initial second matrix can be as follows: First, perform feature scaling on the first matrix according to a preset multiple weight matrix group to obtain a first process matrix. Then, add the first matrix and the first process matrix, and perform matrix normalization on the obtained matrix after addition to obtain a second process matrix. Next, perform feature fusion on the second process matrix to obtain a third process matrix. Finally, add the second process matrix and the third process matrix, and perform normalization on the obtained matrix after addition to obtain the initial second matrix.

[0063] Among them, performing feature scaling on the first matrix according to a preset multiple weight matrix group includes: magnifying the parameters belonging to the first type of data in the first matrix and reducing the parameters belonging to the second type of data in the first matrix according to a preset multiple weight matrix group.

[0064] Taking the first matrix as a 7×200-dimensional matrix as an example, let X ∈ R 7×200 represent the first matrix. Then, first perform feature scaling on the first matrix X ∈ R 7×200 using a preset multiple weight matrix group to obtain a first process matrix H ∈ R 7×200 . Then, add the first matrix X ∈ R 7×200 and the first process matrix H ∈ R 7×200 , and perform matrix normalization on the obtained matrix after addition to obtain a second process matrix Then, perform feature fusion on the second process matrix to obtain a third process matrix M ∈ R 7×200 . Finally, add the second process matrix and the third process matrix M ∈ R 7×200 , and perform normalization on the obtained matrix after addition to obtain the initial second matrix

[0065] Among them, a multi-layer perceptron can be used to perform feature fusion processing on the second process matrix. The multi-layer perceptron can be selected according to actual needs. For example, when the second process matrix is a 7×200-dimensional matrix, a multi-layer perceptron with two hidden layers each having 200 neurons can be used to process the second process matrix. The example here is only for easy understanding and should not be construed as a limitation to this application.

[0066] Specifically, it can also be to use multiple multiple linear equations with the same number of elements as the third process matrix to perform linear combination on each channel of the second process matrix, and use a non-linear activation function to add non-linear characteristics to the result. Among them, the non-linear activation function can be Gaussian Error Linerar Units (GELUs).

[0067] In one implementation manner, the specific process of performing feature scaling processing on the first matrix according to a preset plurality of weight matrix groups to obtain the first process matrix can be: First, for each weight matrix group in the preset plurality of weight matrix groups, based on the weight matrix group and the first matrix, obtain the scoring matrix corresponding to the weight matrix group. Then process the scoring matrix corresponding to each preset weight matrix group to obtain the first process matrix.

[0068] Among them, each preset weight matrix group can include multiple different weight matrices. For example, it can be 2, 3, 4, 5, etc. The number of weight matrices in each weight matrix group can be set according to actual needs, and no specific limitation is imposed on its specific number here.

[0069] Processing the scoring matrix corresponding to each preset weight matrix group to obtain the first process matrix can be to first combine the scoring matrices corresponding to each weight matrix group, and then use a preset matrix to perform feature fusion on the combined matrix to mix the feature information in multiple scoring matrices to obtain the first process matrix.

[0070] To facilitate understanding of the specific process of, for each weight matrix group in the preset plurality of weight matrix groups, based on the weight matrix group and the first matrix, obtaining the scoring matrix corresponding to the weight matrix group, the following takes each weight matrix group in the preset plurality of weight matrix groups including 3 weight matrices as an example to illustrate the specific steps it includes.

[0071] When each of the preset multiple weight matrix groups includes 3 weight matrices, for each of the preset multiple weight matrix groups, the specific process of obtaining the scoring matrix corresponding to the weight matrix group based on the weight matrix group and the first matrix can be as follows: First, for each of the preset multiple weight matrix groups, multiply the first matrix by each weight matrix in the weight matrix group to obtain a first branch matrix, a second branch matrix, and a third branch matrix, where the second branch matrix and the third branch matrix are the same. Then, based on the first branch matrix and the second branch matrix, obtain a fusion matrix; finally, based on the fusion matrix and the third branch matrix, obtain the scoring matrix corresponding to the weight matrix group.

[0072] Among them, the process of obtaining the fusion matrix based on the first branch matrix and the second branch matrix can be that first, multiply the transposed matrix of the first branch matrix by the second branch matrix, and perform normalization processing on the multiplied matrix, and then use the softmax function to process the normalized matrix to obtain the fusion matrix.

[0073] The specific process of obtaining the scoring matrix corresponding to the weight matrix group based on the fusion matrix and the third branch matrix can be to multiply the fusion matrix by the third branch matrix to obtain the scoring matrix corresponding to the weight matrix group.

[0074] For example, if the nth weight matrix group includes W qn ∈R 200×50 、W kn ∈R 200×50 、W vn ∈R 200×50 three weight matrices, n is a positive integer greater than or equal to 2, and the first matrix is X ∈ R 7×200 When, for the nth weight matrix group, first multiply the first matrix X ∈ R 7×200 by W qn ∈R 200×50 、W kn ∈R 200×50 、W vn ∈R 200×50 respectively to obtain a first branch matrix Q n ∈R 7×50 、a second branch matrix K n ∈R 7×50 、and a third branch matrix V n ∈R 7×50 . Then multiply the transposed matrix of the first branch matrix Q n ∈R 7×50 by the second branch matrix K n ∈R 7×50Multiply them, normalize the resulting matrix, and then process the normalized matrix using the softmax function to obtain the fusion matrix, which can be expressed as Finally, the fusion matrix is multiplied by the third branch matrix V n ∈R 7×50 to obtain the scoring matrix H′ corresponding to this group of weight matrices n ∈R 7×50 .

[0075] When each group of weight matrices in the preset multiple groups of weight matrices includes 5 weight matrices, for each group of weight matrices in the preset multiple groups of weight matrices, the specific process of obtaining the scoring matrix corresponding to this group of weight matrices based on this group of weight matrices and the first matrix can be as follows: First, for each group of weight matrices in the preset multiple groups of weight matrices, multiply the first matrix by each weight matrix in this group of weight matrices respectively to obtain the fourth branch matrix, the fifth branch matrix, the sixth branch matrix, the seventh branch matrix, and the eighth branch matrix. Then, based on the fourth branch matrix and the fifth branch matrix, obtain the first fusion matrix; based on the sixth branch matrix and the seventh branch matrix, obtain the second fusion matrix; then, based on the first fusion matrix and the eighth branch matrix, obtain the first initial scoring matrix; based on the second fusion matrix and the eighth branch matrix, obtain the second initial scoring matrix. Finally, add the first initial scoring matrix and the second initial scoring matrix to obtain the scoring matrix corresponding to this group of weight matrices.

[0076] When each group of weight matrices in the preset multiple groups of weight matrices includes 2 weight matrices, for each group of weight matrices in the preset multiple groups of weight matrices, the specific process of obtaining the scoring matrix corresponding to this group of weight matrices based on this group of weight matrices and the first matrix can be as follows: First, for each group of weight matrices in the preset multiple groups of weight matrices, multiply the first matrix by each weight matrix in this group of weight matrices respectively to obtain the ninth branch matrix and the tenth branch matrix. Then, based on the ninth branch matrix and the tenth branch matrix, obtain the third fusion matrix; then, based on the third fusion matrix and the ninth branch matrix or the tenth branch matrix, obtain the scoring matrix.

[0077] When each of the preset multiple weight matrix groups includes 4 weight matrices, for each of the preset multiple weight matrix groups, the specific process of obtaining the scoring matrix corresponding to the weight matrix group based on the weight matrix group and the first matrix may be as follows: First, for each of the preset multiple weight matrix groups, multiply the first matrix by each weight matrix in the weight matrix group respectively to obtain the eleventh branch matrix, the twelfth branch matrix, the thirteenth branch matrix, and the fourteenth branch matrix. Then, based on the eleventh branch matrix and the twelfth branch matrix, obtain the fourth fusion matrix; based on the thirteenth branch matrix and the fourteenth branch matrix, obtain the fifth fusion matrix; then, based on the fourth fusion matrix and any one of the eleventh branch matrix, the twelfth branch matrix, the thirteenth branch matrix, and the fourteenth branch matrix, obtain the third initial scoring matrix; based on the fifth fusion matrix and any one of the eleventh branch matrix, the twelfth branch matrix, the thirteenth branch matrix, and the fourteenth branch matrix, obtain the fourth initial scoring matrix. Finally, add the third initial scoring matrix and the fourth initial scoring matrix to obtain the scoring matrix.

[0078] Among them, the specific process and principle of obtaining the fusion matrix based on two branch matrices are the same as those of obtaining the fusion matrix based on the first branch matrix and the second branch matrix recorded above. For example, the specific process and principle of obtaining the first fusion matrix based on the fourth branch matrix and the fifth branch matrix, and obtaining the second fusion matrix based on the sixth branch matrix and the seventh branch matrix are the same as those of obtaining the fusion matrix based on the first branch matrix and the second branch matrix recorded above. The specific process and principle of obtaining the scoring matrix or the initial scoring matrix based on the fusion matrix and any one branch matrix are the same as those of obtaining the scoring matrix based on the fusion matrix and the third branch matrix recorded above. For example, the specific process and principle of obtaining the first initial scoring matrix based on the first fusion matrix and the eighth branch matrix, and obtaining the second initial scoring matrix based on the second fusion matrix and the eighth branch matrix are the same as those of obtaining the scoring matrix based on the fusion matrix and the third branch matrix recorded above. Therefore, for the sake of brief description, it will not be elaborated here.

[0079] It can be understood that the example that each of the preset multiple weight matrix groups includes 2, 3, 4, or 5 weight matrices is only for easy understanding, and the number of weight matrices specifically included in the weight matrix group is not limited here.

[0080] To intuitively understand the above process of obtaining the first process matrix from the first matrix, taking the example that there are 4 weight matrix groups and each group of weight matrix groups includes 3 weight matrices, please refer to Figure 2 .

[0081] AsFigure 2 As shown, the dimension of the first matrix is L×4D. The first matrix X is multiplied by the weight matrices W q1 , W k1 , W v1 included in the first weight matrix group respectively to obtain the first branch matrix Q 1 , the second branch matrix K 1 , and the third branch matrix V 1 . Then, based on the first branch matrix Q 1 and the second branch matrix K 1 , a fusion matrix is obtained. The fusion matrix is a matrix with the dimension of L×L. After that, the fusion matrix and the third branch matrix V 1 are multiplied to obtain the scoring matrix H′ 1 corresponding to the first group of weight matrix groups. Similarly, based on the other 3 groups of weight matrix groups and the first matrix, the corresponding scoring matrices H′ 2 , H′ 3 , H′ 4 are obtained. Then, the scoring matrices H′ 1 , H′ 2 , H′ 3 , H′ 4 are merged to obtain the initial first process matrix. Finally, the preset matrix W H is used to perform feature fusion on the merged matrix to mix the feature information in the four scoring matrices and obtain the first process matrix.

[0082] S400: Perform a linear mapping process on the parameters in the second matrix to obtain a third matrix with a preset dimension.

[0083] Specifically, the parameters in the second matrix are mapped to a higher-dimensional space by linear mapping to obtain a third matrix with a preset dimension.

[0084] It can be understood that the preset dimension can be set according to actual needs. For example, it can be set to 7×5000, and no specific limitation is imposed on its specific dimension here.

[0085] The specific process and principle of performing linear mapping on the matrix are well known to those skilled in the art. For the sake of brief description, it will not be elaborated here.

[0086] S500: Perform global average pooling processing on the parameters in the third matrix to obtain the predicted ground penetrating radar electric field vector.

[0087] By performing global average pooling processing on the parameters in the third matrix, the feature information corresponding to different parameters is fused, so that the finally obtained predicted ground penetrating radar electric field vector is more accurate.

[0088] Among them, the specific process and principle of performing global average pooling on the matrix are well-known to those skilled in the art. For the sake of brief description, they will not be elaborated here.

[0089] In one implementation, at least one of the above S200, S300, S400, and S500 is performed through a preset ground penetrating radar electric field prediction model.

[0090] That is, the parameter group is subjected to feature encoding through a preset ground penetrating radar electric field prediction model to obtain a first matrix; and / or, the parameters in the first matrix are scaled through a preset ground penetrating radar electric field prediction model to obtain a second matrix; and / or, the parameters in the second matrix are linearly mapped through a preset ground penetrating radar electric field prediction model to obtain a third matrix of a preset dimension; and / or, the parameters in the third matrix are subjected to global average pooling through a preset ground penetrating radar electric field prediction model to obtain a predicted ground penetrating radar electric field vector.

[0091] Among them, the implementation principle of performing the above S200, S300, S400, and S500 through a preset ground penetrating radar electric field prediction model and the resulting technical effects are the same as those of the foregoing S200, S300, S400, and S500. For the sake of brief description, they will not be elaborated here.

[0092] Optionally, the above preset ground penetrating radar electric field prediction model can be directly obtained from a third party and can be directly used when needed.

[0093] Alternatively, the above preset ground penetrating radar electric field prediction model can be trained. At this time, the ground penetrating radar electric field prediction method further includes: obtaining training samples, where the training samples include a parameter group representing the electrical properties of the target medium and the corresponding true ground penetrating radar electric field vector. The initial ground penetrating radar electric field prediction model is trained using the training samples to obtain a preset ground penetrating radar electric field prediction model.

[0094] In one implementation, the specific process of training the initial ground penetrating radar electric field prediction model using the training samples to obtain a preset ground penetrating radar electric field prediction model can be as follows: First, the training samples are input into the initial ground penetrating radar electric field prediction model to obtain a predicted ground penetrating radar electric field vector. Then, based on the predicted ground penetrating radar electric field vector and the true ground penetrating radar electric field vector, a first error, a second error, and a third error are obtained. Then, based on the first error, the second error, and the third error, a loss value is obtained; finally, the model parameters of the initial ground penetrating radar electric field prediction model are updated based on the loss value. After that, the ground penetrating radar electric field prediction model with updated model parameters is trained again using the training samples until a preset condition is met to obtain a preset ground penetrating radar electric field prediction model.

[0095] The above preset condition may be that the ground penetrating radar electric field prediction model converges, that is, the loss value of the ground penetrating radar electric field prediction model no longer decreases; or, it may also be that the number of training times of the ground penetrating radar electric field prediction model reaches a preset number of times.

[0096] Among them, the first error is the sum of the squares of the differences between the electric field components corresponding to each time series point in the predicted ground penetrating radar electric field vector and the electric field components corresponding to the same time series point in the true ground penetrating radar electric field vector; the second error is the square of the difference between the average value of all electric field components in the predicted ground penetrating radar electric field vector and the average value of all electric field components in the true ground penetrating radar electric field vector; the third error is the square of the difference between the variance of all electric field components in the predicted ground penetrating radar electric field vector and the variance of all electric field components in the true ground penetrating radar electric field vector.

[0097] To facilitate the understanding of the specific process of obtaining the first error, the second error, and the third error based on the predicted ground penetrating radar electric field vector and the true ground penetrating radar electric field vector as described above. Assume that both the predicted ground penetrating radar electric field vector and the true ground penetrating radar electric field vector are vectors including 5000 time series point electric field components, that is, the U predicted ground penetrating radar electric field vectors are The U true ground penetrating radar electric field vectors are E U ∈R U×5000 , where U is the number of parameter groups input into the ground penetrating radar electric field prediction model. Then, use MSE_loss to represent the first error, Mean_loss to represent the second error, and Var_loss to represent the third error.

[0098] Then the first error The second error Among them, The third error Among them, Among them, E u ∈R 5000 is the predicted ground penetrating radar electric field vector corresponding to the u-th parameter group, is the true ground penetrating radar electric field vector corresponding to the u-th parameter group.

[0099] Correspondingly, use Total_loss to represent the loss value. Then, the specific process of obtaining the loss value based on the first error, the second error, and the third error can be: the loss value Total_loss = MSE_loss + Mean_loss + Var_loss.

[0100] In order to make the loss value more accurately reflect the fitting degree of the initial ground penetrating radar electric field prediction model on the training samples, a control third error, that is, a parameter λ for controlling the influence amplitude of the mean and variance losses on the overall loss value, can be introduced into the function for calculating the loss value. At this time, the loss value Total_loss = MSE_loss + Mean_loss + λVar_loss.

[0101] Among them, the larger the value of λ, after updating the parameters of the model using the loss value, the updated ground penetrating radar electric field prediction model will pay more attention to the regression of variance, thereby improving the accuracy of the model output result. The specific value of λ can be set according to actual needs. For example, it can be λ = 10 8 , the example here is only for easy understanding and should not be regarded as a limitation to this application.

[0102] In one implementation manner, the parameter groups in the above-mentioned training samples for training the initial ground penetrating radar electric field prediction model can be sampled from a preset interval through Latin hypercube sampling. Among them, the specific process of obtaining parameters from the preset interval using Latin hypercube sampling is the same as the process of obtaining the parameter group representing the electrical properties of the target medium described above. For the sake of brief description, it will not be elaborated here.

[0103] After obtaining the parameter groups in the training samples, the true ground penetrating radar electric field vectors corresponding to each parameter group can be obtained using the finite difference time domain (FDTD) method.

[0104] As Figure 3 shown, the soil layer is a dispersive lossy soil medium. The soil layer includes a 1m×1m×1m solid metal target, and next to the metal target is a 0.5m×0.5m×0.5m dry granite. The transmitting antenna and the receiving antenna are part of a ground penetrating radar (not shown in the figure). The transmitting antenna is used to transmit electromagnetic pulses to the soil layer, and the receiving antenna is used to receive the electric field signals reflected from the soil layer. Among them, the excitation source pulse of the electromagnetic pulse emitted by the transmitting antenna is a Blackmann - Harris pulse, whose center frequency f c = 200MHz, T s = 1.55 / f c , and the boundary condition is an absorbing boundary condition.

[0105] Use to calculate the complex relative permittivity ∈ r . Among them, ∈ ∞ , ∈ s , A p (p = 1, 2), τ p (p = 1, 2) and σ sThere are 7 uncertain input parameters, that is, a parameter group characterizing the electrical properties of the target medium, where, ∈ ∞ represents the dielectric constant at infinite frequency, ∈ s is the static dielectric constant, A p is the pole amplitude, τ p is the relaxation time, σ s is the static conductivity. ω is the angular frequency, ∈ 0 is the node constant of free space, j 2 = -1.

[0106] Then, the uniaxial perfectly matched layer (PUML) is used as the absorbing boundary condition, and the three auxiliary variables L| z of the electric field E| z (ω), D| z (ω), R p | z (ω) are respectively D| z (ω) = ∈ r (ω)E| z ; where, Z i is related to the normal plane of x, y, z, and i is any one of x, y, z,

[0107] Substitute the three auxiliary variables L| z (ω), D| z (ω), R p | z (ω) into the following formula, and the component values E| i (i = x, y) of the electric field in the x, y, z directions can be obtained, and then the true ground penetrating radar electric field vector corresponding to each parameter group can be obtained.

[0108]

[0109]

[0110]

[0111]

[0112] C = (2∈ 0 ∈ ∞ (θ) + σ s (θ)Δt)(2τ 1 (θ) + Δt)(2τ 2 (θ) + Δt) + 2∈ 0Δt(∈ s (θ)-∈ ∞ (θ))(A 1 (θ)(2τ 2 (θ)+Δt)+A 2 (θ)(2τ 1 (θ)+Δt))

[0113] C 1 =(2∈ 0 ∈ ∞ (θ)-σ s (θ)Δt)(2τ 1 (θ)+Δt)(2τ 2 (θ)+Δt)+2∈ 0 Δt(∈ s (θ)-∈ ∞ (θ))(A 1 (θ)(2τ 2 (θ)+Δt)+A 2 (θ)(2τ 1 (θ)+Δt])

[0114] where s i is the Debye model coefficient, σ i is the magnetic permeability of the medium, nx and ny are the spatial step sizes of the finite-difference time-domain method, Δt is the time step size of the finite-difference time-domain method, θ is the uncertain input parameter, k is the number of time steps, ∈_0 is the permittivity in the initial state, ∈_s is the static permittivity, A_1, A_2, τ_1, τ_2, S_i (i = x, y, z) are the coefficients of the Debye model in Formula 1, and σ_s is the conductivity of the dispersive lossy medium.

[0115] Among them, the specific process and principle of obtaining the true ground penetrating radar electric field vector corresponding to each parameter group by using the finite-difference time-domain method are well known to those skilled in the art. For the sake of brief description, it will not be elaborated here.

[0116] Based on the same inventive concept, the embodiment of the present application also provides a model training method, which includes: obtaining training samples, where the training samples include parameter groups representing the electrical properties of the target medium and the true ground penetrating radar electric field vectors corresponding to the parameter groups. Using the training samples to train the initial ground penetrating radar electric field prediction model to obtain a preset ground penetrating radar electric field prediction model.

[0117] In one implementation manner, the specific process of training the initial ground penetrating radar (GPR) electric field prediction model with training samples to obtain a preset GPR electric field prediction model may be as follows: First, input the training samples into the initial GPR electric field prediction model to obtain a predicted GPR electric field vector. Then, based on the predicted GPR electric field vector and the true GPR electric field vector, obtain the first error, the second error, and the third error. Next, based on the first error, the second error, and the third error, obtain a loss value; finally, update the model parameters of the initial GPR electric field prediction model based on the loss value. Then, use the training samples to train the GPR electric field prediction model with updated model parameters again until a preset condition is met to obtain a preset GPR electric field prediction model.

[0118] The above-mentioned preset condition may be that the GPR electric field prediction model converges, that is, the loss value of the GPR electric field prediction model no longer decreases; or, it may also be that the number of training times for the GPR electric field prediction model reaches a preset number of times.

[0119] Among them, the first error is the sum of the squares of the differences between the electric field components corresponding to each time sequence point in the predicted GPR electric field vector and the electric field components corresponding to the same time sequence point in the true GPR electric field vector; the second error is the square of the difference between the average value of all electric field components in the predicted GPR electric field vector and the average value of all electric field components in the true GPR electric field vector; the third error is the square of the difference between the variance of all electric field components in the predicted GPR electric field vector and the variance of all electric field components in the true GPR electric field vector.

[0120] The model training method provided in the embodiments of the present application has the same implementation principle and the same technical effects as the process and principle of training the GPR electric field prediction model in the foregoing embodiments of the GPR electric field prediction method. For a brief description, for the parts not mentioned in the embodiments of the model training method, reference may be made to the corresponding content in the foregoing embodiments of the GPR electric field prediction method.

[0121] Please refer to Figure 4 , Figure 4 which is a schematic structural diagram of a GPR electric field prediction model provided in an embodiment of the present application. As Figure 4 shown, the GPR electric field prediction model 100 includes an encoding layer 110, a feature scaling layer 120, a linear mapping layer 130, and a pooling layer 140.

[0122] The encoding layer 110 is used to perform feature encoding processing on a parameter group representing the electrical properties of a target medium to obtain a first matrix.

[0123] The feature scaling layer 120 is used to perform scaling processing on the first matrix to obtain a second matrix.

[0124] The linear mapping layer 130 is used to perform a linear mapping process on the second matrix to obtain a third matrix with a preset dimension.

[0125] The pooling layer 140 is used to perform global average pooling on the third matrix to obtain the predicted ground penetrating radar electric field vector.

[0126] Among them, in order to make the obtained second matrix more accurately represent the electric field characteristics of the ground penetrating radar, the feature scaling layer 120 may include a first sub-feature scaling layer and a second sub-feature scaling layer connected in sequence.

[0127] The first sub-feature scaling layer is used to amplify the parameters of the first type of data in the first matrix and reduce the parameters of the second type of data in the first matrix to obtain an initial second matrix, where the first type of data represents the electric field characteristics of the ground penetrating radar, and the second type of data is the parameters in the first matrix other than the first type of data.

[0128] The second sub-feature scaling layer is used to amplify the parameters of the first type of data in the initial second matrix and reduce the parameters of the second type of data in the initial second matrix to obtain the second matrix.

[0129] In one implementation, the first sub-feature scaling layer includes a multi-branch attention layer, a first accumulation and normalization layer, a multi-layer perceptron, and a second accumulation and normalization layer connected in sequence. For ease of understanding, please refer to Figure 5 .

[0130] The multi-branch attention layer is used to perform feature scaling on the first matrix to obtain a first process matrix, where the parameters of the first type of data in the first matrix are amplified according to a preset group of multiple weight matrices, and the parameters of the second type of data in the first matrix are reduced.

[0131] The first accumulation and normalization layer is used to perform matrix addition on the first matrix and the first process matrix, and perform matrix normalization on the first matrix and the first process matrix after matrix addition to obtain a second process matrix.

[0132] The multi-layer perceptron is used to perform feature fusion on the second process matrix to obtain a third process matrix.

[0133] The second accumulation and normalization layer is used to perform matrix addition on the second process matrix and the third process matrix, and perform normalization on the second process matrix and the third process matrix after matrix addition to obtain the initial second matrix.

[0134] It can be understood that the specific structures of the second sub-feature scaling layer and the first sub-feature scaling layer are the same. For a brief description, it will not be elaborated here.

[0135] The ground penetrating radar electric field prediction model 100 provided by the embodiments of the present application has the same implementation principle and technical effects as those of the foregoing embodiments of the ground penetrating radar electric field prediction method. For the sake of brief description, for the parts not mentioned in the embodiments of the ground penetrating radar electric field prediction model 100, reference may be made to the corresponding content in the foregoing embodiments of the ground penetrating radar electric field prediction method.

[0136] Please refer to Figure 6 , Figure 6 , which is a schematic structural diagram of a ground penetrating radar electric field prediction device 200 provided by an embodiment of the present application, including an acquisition module 210 and a processing module 220.

[0137] The acquisition module 210 is used to acquire a parameter set preset to characterize the electrical properties of a target medium.

[0138] The processing module 220 is used to perform feature encoding processing on the parameter set to obtain a first matrix.

[0139] The processing module 220 is further used to perform scaling processing on the parameters in the first matrix to obtain a second matrix.

[0140] The processing module 220 is further used to perform linear mapping processing on the parameters in the second matrix to obtain a third matrix with a preset dimension.

[0141] The processing module 220 is further used to perform global average pooling processing on the parameters in the third matrix to obtain a predicted ground penetrating radar electric field vector.

[0142] Specifically, the processing module 220 is used to perform amplification processing on the parameters belonging to the first type of data in the first matrix and perform reduction processing on the parameters belonging to the second type of data in the first matrix to obtain the second matrix, where the first type of data characterizes the electric field characteristics of the ground penetrating radar, and the second type of data is the parameters in the first matrix other than the first type of data.

[0143] Specifically, the processing module 220 is used to perform amplification processing on the parameters belonging to the first type of data in the first matrix and perform reduction processing on the parameters belonging to the second type of data in the first matrix to obtain an initial second matrix; perform amplification processing on the parameters belonging to the first type of data in the initial second matrix and perform reduction processing on the parameters belonging to the second type of data in the initial second matrix to obtain the second matrix.

[0144] The processing module 220 is specifically configured to perform feature scaling processing on the first matrix according to a plurality of preset weight matrix groups, so as to obtain a first process matrix, wherein the parameters of the first matrix belonging to the first type of data are amplified, and the parameters of the first matrix belonging to the second type of data are reduced according to the plurality of preset weight matrix groups; perform matrix addition on the first matrix and the first process matrix, and perform matrix normalization processing on the obtained matrix after addition to obtain a second process matrix; perform feature fusion processing on the second process matrix to obtain a third process matrix; perform matrix addition on the second process matrix and the third process matrix, and perform normalization processing on the obtained matrix after addition to obtain an initial second matrix.

[0145] The processing module 220 is specifically configured to, for each of the plurality of preset weight matrix groups, obtain a scoring matrix corresponding to the weight matrix group based on the weight matrix group and the first matrix, wherein each of the plurality of preset weight matrix groups includes a plurality of different weight matrices; process the scoring matrix corresponding to each of the preset weight matrix groups to obtain the first process matrix.

[0146] The processing module 220 is specifically configured to, for each of the plurality of preset weight matrix groups, multiply the first matrix by each of the weight matrices in the weight matrix group respectively to obtain a first branch matrix, a second branch matrix, and a third branch matrix, wherein the second branch matrix and the third branch matrix are the same; obtain a fusion matrix based on the first branch matrix and the second branch matrix; obtain a scoring matrix corresponding to the weight matrix group based on the fusion matrix and the third branch matrix.

[0147] In one implementation, the parameter group includes at least one of the dielectric constant at infinite frequency of the target medium, the electrostatic dielectric constant, the polarization amplitude, the relaxation time, the static conductivity, the angular frequency, and the node constant of free space.

[0148] The processing module 220 is specifically configured to perform feature encoding processing on the parameter group through a preset ground penetrating radar electric field prediction model to obtain the first matrix; and / or perform scaling processing on the parameters in the first matrix through the preset ground penetrating radar electric field prediction model to obtain the second matrix; and / or perform linear mapping processing on the parameters in the second matrix through the preset ground penetrating radar electric field prediction model to obtain the third matrix of a preset dimension; and / or perform global average pooling processing on the parameters in the third matrix through the preset ground penetrating radar electric field prediction model to obtain the predicted ground penetrating radar electric field vector.

[0149] The processing module 220 is further configured to obtain training samples, where the training samples include a parameter set representing the electrical properties of the target medium and the true ground penetrating radar (GPR) electric field vector corresponding to the parameter set; and use the training samples to train an initial GPR electric field prediction model to obtain the preset GPR electric field prediction model.

[0150] Specifically, the processing module 220 is configured to input the training samples into the initial GPR electric field prediction model to obtain a predicted GPR electric field vector; obtain a first error, a second error, and a third error, where the first error is the sum of the squares of the differences between the electric field components corresponding to each time series point in the predicted GPR electric field vector and the electric field components corresponding to the same time series point in the true GPR electric field vector; the second error is the square of the difference between the average value of all the electric field components in the predicted GPR electric field vector and the average value of all the electric field components in the true GPR electric field vector; the third error is the square of the difference between the variance of all the electric field components in the predicted GPR electric field vector and the variance of all the electric field components in the true GPR electric field vector; obtain a loss value based on the first error, the second error, and the third error; update the model parameters of the initial GPR electric field prediction model based on the loss value; and use the training samples to train the GPR electric field prediction model with updated model parameters again until a preset condition is met to obtain the preset GPR electric field prediction model.

[0151] The principle of implementation and the technical effects produced by the GPR electric field prediction device 200 provided in the embodiments of the present application are the same as those of the foregoing embodiments of the GPR electric field prediction method. For a brief description, for parts not mentioned in the embodiments of the GPR electric field prediction device 200, reference may be made to the corresponding content in the foregoing embodiments of the GPR electric field prediction method.

[0152] Please refer to Figure 7 , which is an electronic device 300 provided in the embodiments of the present application. The electronic device 300 includes a transceiver 310, a memory 320, a communication bus 330, and a processor 340.

[0153] The transceiver 310, the memory 320, and the processor 340 are electrically connected to each other directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses 330 or signal lines. Among them, the transceiver 310 is used to transmit and receive data. The memory 320 is used to store computer programs, such as storing Figure 6The software function module shown in the figure is the ground penetrating radar electric field prediction device 200. The ground penetrating radar electric field prediction device 200 includes at least one software function module that can be stored in the memory 320 in the form of software or firmware or solidified in the operating system (OS) of the electronic device 300. The processor 340 is used to execute the executable module stored in the memory 320, such as the software function module or computer program included in the ground penetrating radar electric field prediction device 200.

[0154] Processor 340 is used to obtain a preset parameter group characterizing the electrical properties of the target medium; perform feature encoding processing on the parameter group to obtain a first matrix; perform scaling processing on the parameters in the first matrix to obtain a second matrix; perform linear mapping processing on the parameters in the second matrix to obtain a third matrix of preset dimensions; perform global average pooling processing on the parameters in the third matrix to obtain a predicted ground penetrating radar electric field vector.

[0155] Among them, the memory 320 can be, but is not limited to, random access memory (Random Access Memory, RAM), read only memory (Read Only Memory, ROM), programmable read-only memory (Programmable Read-Only Memory, PROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, EPROM), electrically erasable read-only memory (Electric Erasable Programmable Read-Only Memory, EEPROM), etc.

[0156] The processor 340 may be an integrated circuit chip with signal processing capabilities. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed. A general-purpose processor may be a microprocessor or the processor 340 may also be any conventional processor, etc.

[0157] Among them, the above-mentioned electronic device 300 includes, but is not limited to, personal computers, servers, etc.

[0158] The embodiment of the present application further provides a non-volatile computer-readable storage medium (hereinafter referred to as the storage medium). A computer program is stored on the storage medium. When the computer program is run by a computer such as the above-mentioned electronic device 300, it executes the above-mentioned ground penetrating radar electric field prediction method. The computer-readable storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs.

[0159] The foregoing is only a preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the ground penetrating radar electric field, characterized in that, it includes: obtaining a preset parameter group characterizing the electrical properties of the target medium; the parameter group includes at least one of the dielectric constant at infinite frequency, static dielectric constant, polar amplitude, relaxation time, static conductivity, angular frequency, and node constant of free space of the target medium; performing feature encoding processing on the parameter group to obtain a first matrix; performing scaling processing on the parameters in the first matrix to obtain a second matrix; performing linear mapping processing on the parameters in the second matrix to obtain a third matrix with a preset dimension; performing global average pooling processing on the parameters in the third matrix to obtain a predicted ground penetrating radar electric field vector; The performing scaling processing on the parameters in the first matrix to obtain a second matrix includes: amplifying the parameters in the first matrix belonging to the first type of data and reducing the parameters in the first matrix belonging to the second type of data to obtain an initial second matrix; wherein, the first type of data characterizes the electric field characteristics of the ground penetrating radar, and the second type of data is the parameters in the first matrix other than the first type of data; amplifying the parameters in the initial second matrix belonging to the first type of data and reducing the parameters in the initial second matrix belonging to the second type of data to obtain the second matrix.

2. The method according to claim 1, characterized in that, The amplifying the parameters in the first matrix belonging to the first type of data and reducing the parameters in the first matrix belonging to the second type of data to obtain an initial second matrix includes: performing feature scaling processing on the first matrix according to a preset plurality of weight matrix groups to obtain a first process matrix, wherein, according to the preset plurality of weight matrix groups, the parameters in the first matrix belonging to the first type of data are amplified, and the parameters in the first matrix belonging to the second type of data are reduced; adding the first matrix and the first process matrix, and performing matrix normalization processing on the obtained matrix after addition to obtain a second process matrix; performing feature fusion processing on the second process matrix to obtain a third process matrix; adding the second process matrix and the third process matrix, and performing normalization processing on the obtained matrix after addition to obtain the initial second matrix.

3. The method according to claim 1, characterized in that, performing feature encoding processing on the parameter group through a preset ground penetrating radar electric field prediction model to obtain the first matrix; and / or, performing scaling processing on the parameters in the first matrix through the preset ground penetrating radar electric field prediction model to obtain the second matrix; and / or, performing linear mapping processing on the parameters in the second matrix through the preset ground penetrating radar electric field prediction model to obtain the third matrix with a preset dimension; and / or, performing global average pooling processing on the parameters in the third matrix through the preset ground penetrating radar electric field prediction model to obtain the predicted ground penetrating radar electric field vector.

4. The method according to claim 3, characterized in that, The method further includes: obtaining training samples, where the training samples include a parameter set characterizing the electrical properties of the target medium and the true ground penetrating radar (GPR) electric field vector corresponding to the parameter set; inputting the training samples into an initial GPR electric field prediction model to obtain a predicted GPR electric field vector; obtaining a first error, a second error, and a third error, where the first error is the sum of the squares of the differences between the electric field components corresponding to each time series point in the predicted GPR electric field vector and the electric field components corresponding to the same time series point in the true GPR electric field vector; the second error is the square of the difference between the average value of all electric field components in the predicted GPR electric field vector and the average value of all electric field components in the true GPR electric field vector; the third error is the square of the difference between the variance of all electric field components in the predicted GPR electric field vector and the variance of all electric field components in the true GPR electric field vector; obtaining a loss value based on the first error, the second error, and the third error; updating the model parameters of the initial GPR electric field prediction model based on the loss value; training the GPR electric field prediction model with updated model parameters using the training samples again until a preset condition is met to obtain the preset GPR electric field prediction model.

5. A model training method, characterized in that, it includes: obtaining training samples, where the training samples include a parameter set characterizing the electrical properties of the target medium and the true GPR electric field vector corresponding to the parameter set; training an initial GPR electric field prediction model using the training samples to obtain a trained GPR electric field prediction model; where the trained GPR electric field prediction model is the GPR electric field prediction model described in claim 3.

6. A GPR electric field prediction model, characterized in that, it includes: an encoding layer for performing feature encoding processing on a parameter set characterizing the electrical properties of the target medium to obtain a first matrix; the parameter set includes at least one of the dielectric constant at infinite frequency, static dielectric constant, polar amplitude, relaxation time, static conductivity, angular frequency, and node constant of free space of the target medium; a feature scaling layer for performing scaling processing on the first matrix to obtain a second matrix; a linear mapping layer for performing linear mapping processing on the second matrix to obtain a third matrix of a preset dimension; a pooling layer for performing global average pooling processing on the third matrix to obtain a predicted GPR electric field vector; the feature scaling layer is specifically configured to amplify the parameters belonging to the first type of data in the first matrix and reduce the parameters belonging to the second type of data in the first matrix to obtain an initial second matrix; where the first type of data characterizes the electric field characteristics of the GPR, and the second type of data is the parameters in the first matrix other than the first type of data; amplify the parameters belonging to the first type of data in the initial second matrix and reduce the parameters belonging to the second type of data in the initial second matrix to obtain the second matrix.

7. A GPR electric field prediction device, It is characterized in that It includes: An acquisition module, configured to acquire a parameter set preset to represent the electrical properties of a target medium; the parameter set includes at least one of the dielectric constant at infinite frequency, static dielectric constant, polar amplitude, relaxation time, static conductivity, angular frequency, and node constant of free space of the target medium; A processing module, configured to perform feature encoding processing on the parameter set to obtain a first matrix; The processing module is further configured to perform scaling processing on the parameters in the first matrix to obtain a second matrix; The processing module is further configured to perform linear mapping processing on the parameters in the second matrix to obtain a third matrix with a preset dimension; The processing module is further configured to perform global average pooling processing on the parameters in the third matrix to obtain a predicted ground penetrating radar electric field vector; Specifically, the processing module is configured to perform amplification processing on the parameters in the first matrix that belong to the first type of data, and perform reduction processing on the parameters in the first matrix that belong to the second type of data to obtain an initial second matrix; wherein, the first type of data represents the electric field characteristics of the ground penetrating radar, and the second type of data is the parameters in the first matrix other than the first type of data; perform amplification processing on the parameters in the initial second matrix that belong to the first type of data, and perform reduction processing on the parameters in the initial second matrix that belong to the second type of data to obtain the second matrix.

8. An electronic device It is characterized in that It includes: A memory and a processor, the memory is connected to the processor; The memory is used for storing programs; The processor is configured to call the program stored in the memory to execute the method according to any one of claims 1-4, or execute the method according to claim 5.

Images