Ground impact load prediction method, system and device based on generative neural network

Abstract

The application provides a ground impact load prediction method, system and equipment based on a generative neural network, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring explosion condition target parameters corresponding to a target area; inputting the explosion condition target parameters into a ground impact load prediction model to obtain a ground impact load time course prediction curve of the target area output by the ground impact load prediction model, wherein the ground impact load prediction model is obtained by training a conditional generative adversarial network based on explosion condition sample parameters and ground impact load time course sample curves. The application improves the prediction accuracy and efficiency of ground impact load information.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, system and device for predicting ground impact loads based on generative neural networks. Background Technology

[0002] Because of their strong shock energy absorption characteristics, rock masses in underground structures have long served as a crucial protective barrier for critical infrastructure. However, with the significant increase in the destructive power of advanced drilling equipment, the impact loads generated by underground explosions have become a serious threat to underground facilities. Therefore, it is necessary to quickly and accurately obtain information on the blast shock wave loads to analyze the structural dynamic response and assess damage.

[0003] Currently, information on ground impact loads is mainly obtained through three methods: First, field tests can realistically reflect the explosion effect, but the testing cost is extremely high, and large-scale implementation is difficult due to limitations in safety conditions, test scale, and explosion yield. Second, scaled-down model tests use similarity theory to simulate the explosion process, but errors caused by the scaling effect exist, making it difficult to fully reflect the prototype conditions. Third, numerical simulation methods based on finite element method, smoothed particle hydrodynamics (SPH) methods, or a combination of both, can simulate the impact effect under different explosion yields, burial depths, and geological conditions, but they consume huge computational resources, and modeling and post-processing are complex, making rapid prediction difficult. In addition, existing research usually establishes empirical formulas based on experimental or simulation data through curve fitting, but such formulas are mostly for single geological conditions, have limited applicability, and cannot fully describe complex geological structures and nonlinear impact propagation characteristics, resulting in insufficient prediction accuracy, especially significant deviations under unconventional conditions, making it difficult to meet the needs of engineering protection design.

[0004] Therefore, there is an urgent need for a method, system, and device for predicting ground impact loads based on generative neural networks to solve the above problems. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method, system, and device for predicting ground impact loads based on generative neural networks.

[0006] This invention provides a method for predicting ground impact loads based on generative neural networks, comprising:

[0007] Obtain the target parameters for the explosion conditions corresponding to the target area;

[0008] The target parameters of the explosion conditions are input into the ground shock load prediction model to obtain the ground shock load time history prediction curve corresponding to the target area output by the ground shock load prediction model. The ground shock load prediction model is obtained by training a conditional generative adversarial network based on the explosion condition sample parameters and the ground shock load time history sample curve.

[0009] According to the present invention, a ground impact load prediction method based on a generative neural network is provided, wherein the ground impact load prediction model is trained through the following steps:

[0010] The explosion condition sample parameters are constructed based on geological sample parameters and explosion sample parameters;

[0011] Obtain the time history sample curve of the ground impact load corresponding to the explosion condition sample parameters;

[0012] The explosion condition sample parameters and the ground impact load time history sample curve are input into the conditional generative adversarial network (GAN), and the preset iterative steps are repeatedly executed to update the parameters of the GAN until the iterative training results meet the preset training conditions. The GAN obtained in the last iteration is used as the ground impact load prediction model.

[0013] According to the present invention, a method for predicting ground impact loads based on a generative neural network, the iterative steps specifically include:

[0014] The explosion condition sample parameters of the current round are input into the generator in the conditional generative adversarial network of the current round to obtain the ground impact load time history generation curve of the current round output by the generator of the current round.

[0015] The explosion condition sample parameters of the current wheel, the ground impact load time history generation curve of the current wheel, and the ground impact load time history sample curve corresponding to the explosion condition sample parameters of the current wheel are input into the discriminator in the conditional generative adversarial network of the current wheel to obtain the ground impact load time history prediction and discrimination result of the current wheel output by the discriminator of the current wheel.

[0016] Based on the time history prediction and discrimination results of the ground impact load of the current wheel, the parameters of the generator in the conditional generative adversarial network of the current wheel are updated.

[0017] According to the present invention, a method for predicting ground shock loads based on generative neural networks includes obtaining the time history sample curve of the ground shock load corresponding to the explosion condition sample parameters, comprising:

[0018] A two-dimensional axisymmetric model for free-field confined explosion is constructed based on the fluid-structure interaction algorithm.

[0019] Based on the two-dimensional axisymmetric model and the explosion condition sample parameters, the ground impact load time history simulation is performed to obtain the ground impact load time history sample curve corresponding to the explosion condition sample parameters.

[0020] According to the present invention, a method for predicting ground impact loads based on generative neural networks is provided, wherein constructing the explosion condition sample parameters based on geological sample parameters and explosion sample parameters includes:

[0021] Based on the Latin hypercube sampling method, the parameters of multiple geological samples and multiple explosion samples are combined to obtain the explosion condition sample parameters under different parameter combinations.

[0022] According to the present invention, a ground impact load prediction method based on a generative neural network is provided, wherein the loss function of the conditional generative adversarial network is composed of mean square error and binary cross-entropy, wherein the weight coefficients corresponding to the mean square error and the binary cross-entropy are determined based on an adaptive weighting function; the adaptive weighting function is constructed based on the amplitude difference between near-field data and far-field data during the propagation of the explosion shock wave.

[0023] The present invention also provides a ground impact load prediction system based on a generative neural network, comprising:

[0024] The condition parameter input module is used to obtain the explosion condition target parameters corresponding to the target area;

[0025] The ground impact load prediction module is used to input the explosion condition target parameters into the ground impact load prediction model to obtain the ground impact load time history prediction curve corresponding to the target area output by the ground impact load prediction model. The ground impact load prediction model is obtained by training a conditional generative adversarial network based on the explosion condition sample parameters and the ground impact load time history sample curve.

[0026] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the ground impact load prediction method based on generative neural networks as described above.

[0027] The present invention provides a ground impact load prediction method, system and device based on generative neural networks. By acquiring the target parameters of the explosion conditions in the target area, and then inputting them into the ground impact load prediction model constructed by a conditional generative adversarial network trained based on the explosion condition sample parameters and the ground impact load time history sample curve, the ground impact load time history prediction curve of the target area is obtained, thereby improving the prediction accuracy and efficiency of ground impact load information. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating the ground impact load prediction method based on generative neural networks provided by the present invention.

[0030] Figure 2 A schematic diagram illustrating the network design and training strategy provided by this invention;

[0031] Figure 3 This is a schematic diagram of the structure of the ground impact load prediction system based on generative neural networks provided by the present invention;

[0032] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0034] Figure 1 This is a flowchart illustrating the ground impact load prediction method based on generative neural networks provided by the present invention, as shown below. Figure 1 As shown, this invention provides a method for predicting ground impact loads based on generative neural networks, comprising:

[0035] Step 101: Obtain the target parameters for the explosion conditions corresponding to the target area.

[0036] In this invention, it is necessary to collect various parameter information related to the explosion in the target area. The target parameters for the explosion conditions include explosion parameters and geological parameters. For example, explosion parameters include the amount of explosive. Different amounts of explosive release different amounts of energy during the explosion, and their effects on the ground impact load are also different. Geological parameters involve various properties of the rocks in the target area. Taking limestone as an example, its porosity, Poisson's ratio, and other rock mass structure parameters must be considered. Different combinations of geological parameters will significantly change the propagation and effect of the ground impact load.

[0037] To obtain accurate and comprehensive parameters, various methods can be employed. For example, on-site geological surveys can be conducted using specialized geological detection equipment and techniques to precisely measure the geological parameters of the target area. For explosion parameters, specific values ​​such as the amount of explosive can be determined based on the actual explosion plan or requirements. Preferably, during the parameter acquisition process, to ensure the uniformity and representativeness of the sample space, the Latin hypercube sampling method can be used to select combinations of geological parameters. Reasonable parameter selection ensures that the obtained explosion condition target parameters accurately reflect the actual situation of the target area, providing a reliable basis for subsequent predictions.

[0038] Step 102: Input the target parameters of the explosion conditions into the ground shock load prediction model to obtain the ground shock load time history prediction curve corresponding to the target area output by the ground shock load prediction model. The ground shock load prediction model is obtained by training a conditional generative adversarial network based on the explosion condition sample parameters and the ground shock load time history sample curve.

[0039] In this invention, the ground impact load prediction model is constructed based on a Conditional Generative Adversarial Network (CGAN). The CGAN model consists of a generator and a discriminator, both of which take conditional information (including explosion parameters and geological parameters) as input. During training, a large number of explosion condition sample parameters and ground impact load time history sample curves are required. Optionally, this invention uses LS-DYNA finite element software to establish a two-dimensional axisymmetric model of free-field confined explosion based on a fluid-structure interaction algorithm. For example, using limestone as the research object, considering different explosive charges and different combinations of geological parameters, a total of 150 sets of numerical simulations were conducted, extracting 13,500 particle velocity time history curves as a dataset. These data constitute the sample basis required for training the model.

[0040] Preferably, when training the CGAN model, to address the instability and convergence difficulties that are prone to occur during the training of existing CGAN models, a combined loss function of Mean Square Error (MSE) and Binary Cross Entropy (BCE) is introduced; and an adaptive weighting function is designed to balance the difference in amplitude between near-field and far-field data, thereby ensuring the robustness of the model in prediction across the entire range.

[0041] In this invention, the generator uses a fully connected neural network. The input layer contains explosion parameters and geological parameters. Through the calculation and transformation of the neural network, the output layer generates a complete velocity time history curve, i.e., the ground impact load time history curve. The discriminator is also a fully connected neural network. The input is conditional parameters and the generated curve or the real curve. The true and false judgment results are output through the Sigmoid function. In this way, the parameters of the generator and the discriminator are continuously adjusted to improve the prediction ability of the CGAN model.

[0042] After training, the CGAN model needs to be validated. This invention uses the coefficient of determination R... 2 The performance of the CGAN model on the training and test sets was evaluated using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Squared Error (MSE). The results show that the CGAN model has higher prediction accuracy than traditional Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs). The test set R... 2 The value reached 0.9473. Furthermore, comparison with finite element simulation results showed that the velocity field distribution and time history curves generated by the CGAN model were highly consistent with the numerical results, and the prediction time was less than 10 seconds, significantly improving computational efficiency compared to the approximately 5 hours required for traditional simulations. In addition, feature contribution analysis using the SHAP (SHapley Additive exPlanations) method verified the model's good interpretability, identifying that explosive charge, distance from detonation, and time contributed most to the prediction results, while rock mass structural parameters such as porosity and Poisson's ratio had a significant impact.

[0043] In this invention, after obtaining the target parameters of the explosion conditions in the target area, they are input into a trained ground shock load prediction model. Based on the input parameters, the generator utilizes its learned data distribution patterns and generation capabilities to output the time-history prediction curve of the ground shock load corresponding to the target area, providing crucial data for subsequent analysis and evaluation of the ground shock load in the target area.

[0044] The ground impact load prediction method based on generative neural networks provided by this invention obtains the target parameters of the explosion conditions in the target area, and then inputs them into the ground impact load prediction model constructed by a conditional generative adversarial network trained based on the explosion condition sample parameters and the ground impact load time history sample curve, thereby obtaining the ground impact load time history prediction curve of the target area, thus improving the prediction accuracy and efficiency of ground impact load information.

[0045] Based on the above embodiments, the ground impact load prediction model is trained through the following steps:

[0046] The explosion condition sample parameters are constructed based on geological sample parameters and explosion sample parameters;

[0047] Obtain the time history sample curve of the ground impact load corresponding to the explosion condition sample parameters;

[0048] The explosion condition sample parameters and the ground impact load time history sample curve are input into the conditional generative adversarial network (GAN), and the preset iterative steps are repeatedly executed to update the parameters of the GAN until the iterative training results meet the preset training conditions. The GAN obtained in the last iteration is used as the ground impact load prediction model.

[0049] In this invention, geological sample parameters encompass a variety of geologically related characteristic parameters. For example, rock porosity reflects the proportion of pore volume to total volume; rocks with different porosities have varying abilities to absorb and conduct explosive energy. Poisson's ratio describes the ratio of the absolute values ​​of transverse normal strain to axial normal strain when a material is under uniaxial tension or compression, reflecting the rock's deformation characteristics. Other parameters include rock density and elastic modulus. These parameters can be obtained through field geological surveys and laboratory rock mechanics tests. For instance, during field surveys, rock samples are obtained using geological drilling equipment and brought back to the laboratory for parameter determination.

[0050] The parameters of the explosion samples mainly involve key factors related to the explosion, such as the amount of explosive charge. Different amounts of explosive release vastly different amounts of energy during an explosion, directly affecting the intensity and range of the ground impact load. The detonation method, whether instantaneous or delayed, also affects the energy release process, thus influencing the time-history characteristics of the ground impact load. The burial depth, i.e., the depth to which the explosive is embedded in the ground or medium, also alters the propagation path and effect of the ground impact load. These parameters can be determined based on the actual explosion plan, experimental design, or simulation requirements.

[0051] Furthermore, the collected geological sample parameters and the determined explosion sample parameters are integrated to form a complete set of explosion condition sample parameters. For example, in a simulated explosion experiment, different explosive charges (such as 1 kg, 2 kg, 3 kg, etc.) are set, while considering different geological conditions, such as rocks with different porosities (such as 5%, 10%, 15%, etc.) and different Poisson's ratios (such as 0.2, 0.25, 0.3, etc.). By combining these explosive charges with geological parameters, a series of different explosion condition sample parameters can be constructed for subsequent research and model training.

[0052] In this invention, specialized finite element software, such as LS-DYNA, can be used to establish a two-dimensional axisymmetric model (or other suitable model) of a free-field confined explosion based on a fluid-structure interaction algorithm. Using the pre-constructed explosion condition sample parameters as a foundation, corresponding explosion parameters (such as explosive charge, detonation method, etc.) and geological parameters (such as rock porosity, Poisson's ratio, etc.) are set in the model. Then, a simulation program is run to simulate the explosion process, recording the changes in particle velocity over time in the target area or monitoring point under the explosion, thereby obtaining the ground impact load time history sample curve. For example, different combinations of explosive charge and geological parameters can be set in the simulation to conduct multiple sets of simulation experiments, each yielding multiple corresponding ground impact load time history sample curves.

[0053] Figure 2 This is a schematic diagram of the network design and training strategy provided by the present invention, such as... Figure 2 As shown, the CGAN model consists of two parts: a generator and a discriminator. The generator generates a ground impact load time history curve based on the input conditional information (i.e., explosion condition sample parameters, including explosion parameters and geological parameters). The discriminator determines whether the input curve is generated by the generator or is a real ground impact load time history sample curve. Both the generator and the discriminator use the explosion condition sample parameters as input. The generator outputs the generated curve, and the discriminator outputs the judgment result on the authenticity of the curve.

[0054] Furthermore, the constructed explosion condition sample parameters and the obtained ground impact load time history sample curves are input into the CGAN model. After receiving the explosion condition sample parameters, the generator generates a curve similar to the real ground impact load time history curve based on its internal network structure and parameters; the discriminator simultaneously receives the curve generated by the generator and the real ground impact load time history sample curve, distinguishes between them, and outputs the discrimination result (such as the probability of being real or false).

[0055] In each iteration, based on the discriminator's judgment results, an appropriate loss function, such as a combination of mean squared error (MSE) and binary cross-entropy (BCE) loss function, is used to calculate the errors between the generator and the discriminator. Then, the model parameters are updated using the backpropagation algorithm based on these errors, enabling the generator to produce results closer to the true curve, while simultaneously improving the discriminator's discrimination ability. For example, if the discriminator misclassifies the generator-generated curve as the true curve with a high probability, it indicates that the generator-generated curve is of good quality but may require further optimization; if the discriminator can accurately distinguish between the generated curve and the true curve, it indicates that the generator-generated curve differs significantly from the true curve, requiring substantial adjustments to the generator's parameters.

[0056] In this invention, preset training conditions are established, such as the loss function reaching a small threshold, the discriminator achieving a certain level of accuracy in distinguishing between the generated and real curves, and the number of iterations reaching a preset maximum. After each iteration, it is checked whether these preset training conditions are met. If they are met, the iterative training stops; otherwise, the next round of iterations continues.

[0057] When the iterative training results meet the preset training conditions, the conditional generative adversarial network obtained in the last iteration is used as the ground impact load prediction model. This model has been trained with a large amount of sample data and can generate a relatively accurate ground impact load time history prediction curve when new explosion condition parameters are input, so it can be used for actual ground impact load prediction work.

[0058] Based on the above embodiments, the iterative steps specifically include:

[0059] The explosion condition sample parameters of the current round are input into the generator in the conditional generative adversarial network of the current round to obtain the ground impact load time history generation curve of the current round output by the generator of the current round.

[0060] The explosion condition sample parameters of the current wheel, the ground impact load time history generation curve of the current wheel, and the ground impact load time history sample curve corresponding to the explosion condition sample parameters of the current wheel are input into the discriminator in the conditional generative adversarial network of the current wheel to obtain the ground impact load time history prediction and discrimination result of the current wheel output by the discriminator of the current wheel.

[0061] Based on the time history prediction and discrimination results of the ground impact load of the current wheel, the parameters of the generator in the conditional generative adversarial network of the current wheel are updated.

[0062] In this invention, during each round of training of the conditional generative adversarial network, the explosion condition sample parameters used in the current round are input into the generator corresponding to that round. The explosion condition sample parameters contain various information related to the explosion, such as the explosive charge, detonation method, and blast depth. The generator is a neural network model that, after receiving these parameters, performs calculations based on its existing network structure and parameters.

[0063] After internal calculations, the generator outputs the ground impact load time history generation curve for the current wheel. This generation curve is a curve of ground impact load changing with time, simulated by the generator based on the input explosion condition sample parameters. It represents the generator's prediction result of the ground impact load time history under the current parameter settings.

[0064] Furthermore, the explosion condition sample parameters of the current wheel, the ground impact load time history generated curve of the current wheel output by the generator, and the actual impact load time history sample curve corresponding to the explosion condition sample parameters of the current wheel are input into the discriminator in the current wheel conditional generative adversarial network. The actual impact load time history sample curve is obtained through actual experimental measurement or high-precision numerical simulation and serves as the standard for judging the quality of the generated curve.

[0065] The discriminator is also a neural network model that comprehensively analyzes the three parts of input data mentioned above. On the one hand, the discriminator compares the generated curve with the real sample curve in terms of differences in shape, trend, and numerical magnitude; on the other hand, it combines the parameters of the explosion condition sample to determine whether the generated curve conforms to the characteristics of the ground impact load under a given explosion condition.

[0066] After analysis and calculation, the discriminator outputs the prediction result of the ground impact load time history for the current wheel. This result can be a probability value or a classification result, indicating the likelihood that the discriminator considers the input ground impact load time history generated curve to be real data. For example, the result might be a value between 0 and 1, with the closer to 1 indicating that the discriminator considers the generated curve to be real, and the closer to 0 indicating that it considers it to be false.

[0067] In this invention, the generator parameters are updated and adjusted based on the discriminator's output of the ground impact load time history prediction discrimination result. The aim is to enable the generator to generate data that more closely approximates the real ground impact load time history curve during subsequent training, thereby improving the generator's performance and prediction accuracy. If the discrimination result shows that the generated curve has a high probability of being classified as false data, it indicates a significant difference between the generated curve and the real data. In this case, the backpropagation algorithm is used to calculate the generator's error based on the discrimination result, and the generator's network parameters, such as the connection weights between neurons, are adjusted in the direction of decreasing error. By continuously repeating this process, the generator gradually learns how to generate more realistic ground impact load time history curves based on the explosion condition sample parameters, making it increasingly difficult for the discriminator to distinguish between the generated curve and the real sample curve.

[0068] Based on the above embodiments, obtaining the time history sample curve of the ground impact load corresponding to the explosion condition sample parameters includes:

[0069] A two-dimensional axisymmetric model for free-field confined explosion is constructed based on the fluid-structure interaction algorithm.

[0070] Based on the two-dimensional axisymmetric model and the explosion condition sample parameters, the ground impact load time history simulation is performed to obtain the ground impact load time history sample curve corresponding to the explosion condition sample parameters.

[0071] Fluid-structure interaction (FSI) algorithms are analytical methods used to study the interaction between fluids and solids. In the context of a free-field confined explosion, the gas produced by the explosion is a fluid, while the surrounding rock and other media are solids. This invention employs a FSI algorithm to consider the mutual influence between the explosive fluid (such as high-temperature, high-pressure explosive gas) and the surrounding solid rock. For example, the impact of the explosive fluid on the rock causes deformation and damage, while the deformation of the rock, in turn, affects the flow and pressure distribution of the explosive fluid. This algorithm allows for a more realistic simulation of the complex dynamic behavior between fluids and solids during an explosion.

[0072] In this invention, the LS-DYNA finite element software is used to construct the model. The two-dimensional axisymmetric model is established based on the axisymmetric geometric properties. For a free-field confined explosion scenario, it is assumed that the explosion source and the surrounding environment have the same properties when rotating around a certain axis. Taking limestone as the research object, the geometry, size, and physical and mechanical parameters (such as density, elastic modulus, Poisson's ratio, etc.) of the limestone are accurately set in the two-dimensional axisymmetric model. At the same time, the position and shape of the explosion source are reasonably defined. Considering that the explosion occurs in a confined free-field environment, corresponding boundary conditions are set to simulate this confined situation, such as setting reflection boundaries or transmission boundaries, to construct a two-dimensional axisymmetric model that conforms to the actual situation for free-field confined explosions.

[0073] The explosion condition sample parameters encompass a variety of explosion-related factors, such as different explosive charges, the amount of which directly determines the energy released by the explosion and has a significant impact on ground impact loads. It also includes other parameter combinations that may affect the explosion effect and ground impact loads, such as the initiation method and burial depth. During the simulation, these explosion condition sample parameters are set according to different combinations.

[0074] Furthermore, the pre-defined explosion condition sample parameters are input into the pre-constructed two-dimensional axisymmetric model, and numerical simulation is performed using the computational capabilities of LS-DYNA software. During the simulation, the software calculates the motion and stress state of particles in the surrounding rock medium under given explosion conditions, based on the fluid-structure interaction algorithm and model settings. Specifically, it calculates physical quantities such as particle velocity, acceleration, and stress at different times and locations.

[0075] Through simulation calculations, the change in particle velocity over time is extracted. In this invention, for each set of defined explosion condition sample parameters, corresponding simulation calculations are performed, and the corresponding particle velocity time history data are extracted. Because multiple sets of simulations with different explosion condition sample parameters are performed (e.g., a total of 150 numerical simulations), a large amount of particle velocity time history data is obtained. Each particle velocity time history curve corresponds to a specific set of explosion condition sample parameters. These curves together constitute the ground impact load time history sample curve dataset, providing a data foundation for subsequent ground impact load prediction model training.

[0076] Based on the above embodiments, the construction of the explosion condition sample parameters based on geological sample parameters and explosion sample parameters includes:

[0077] Based on the Latin hypercube sampling method, the parameters of multiple geological samples and multiple explosion samples are combined to obtain the explosion condition sample parameters under different parameter combinations.

[0078] Latin hypercube sampling is a statistical method for multivariate sampling. Its core objective is to generate representative sample points in a multidimensional parameter space, ensuring that the samples are evenly distributed across all dimensions of the parameter space. This allows for the coverage of a large parameter range with a smaller number of samples, thereby improving sampling efficiency and quality.

[0079] In this invention, it is first necessary to obtain multiple geological sample parameters and multiple explosion sample parameters. Geological sample parameters may include parameters reflecting geological characteristics such as rock porosity, Poisson's ratio, density, and elastic modulus; explosion sample parameters cover explosion-related parameters such as explosive charge, detonation method, and explosion depth. These parameters can be obtained through various methods such as field surveys, experimental measurements, and literature research.

[0080] Then, the collected geological sample parameters and explosion sample parameters are combined to form a multidimensional parameter space. Each parameter corresponds to one dimension of the parameter space. For example, if there are 3 types of geological sample parameters and 2 types of explosion sample parameters, then a 5-dimensional parameter space is formed.

[0081] Furthermore, based on the Latin hypercube sampling method, sampling is performed in this multidimensional parameter space, dividing the value range of each parameter into several equally probable intervals. For example, for a certain geological parameter, its value range is [a, b], which is divided into N equally probable sub-intervals. Then, a value is randomly selected from each sub-interval of each parameter. In this way, it is ensured that the sample value is evenly distributed across the entire value range of the parameter in each parameter dimension.

[0082] After selecting values ​​from each sub-interval of each parameter, these different parameter values ​​are then combined. Each different combination of parameter values ​​forms a set of explosion condition sample parameters. For example, geological parameter values ​​are selected from each sub-interval of geological parameters, and explosion parameter values ​​are selected from each sub-interval of explosion parameters. Combining these geological and explosion parameter values ​​yields explosion condition sample parameters under different parameter combinations. In this way, multiple sets of representative explosion condition sample parameters can be obtained. These parameter combinations can comprehensively cover the possible value ranges of geological and explosion parameters, ensuring the homogeneity of the sample space and providing a reliable data foundation for subsequent model training and simulation analysis.

[0083] Based on the above embodiments, the loss function of the conditional generative adversarial network is composed of mean squared error and binary cross-entropy, wherein the weight coefficients corresponding to the mean squared error and the binary cross-entropy are determined based on an adaptive weighting function; the adaptive weighting function is constructed based on the amplitude difference between near-field data and far-field data during the propagation of the explosion shock wave.

[0084] In this invention, the loss function of the conditional generative adversarial network consists of two parts: mean squared error (MSE) and binary cross-entropy (BCE).

[0085] MSE (Mean Squared Error) is a metric used to measure the difference between generated and real data. In the CGAN model, MSE is used to calculate the average of the squared differences between the generated ground impact load time history curve and the real ground impact load time history sample curve at each time point. By minimizing MSE, the curve generated by the generator can be made as close as possible to the real curve numerically, thus ensuring the accuracy of the generated data at the numerical level.

[0086] BCE (Browser Count Elimination) is primarily used to measure the discriminator's ability to distinguish between real and generated data. The discriminator's task is to determine whether the input data comes from a real dataset or is fake data generated by the generator. BCE guides the training of both the discriminator and the generator by calculating the difference between the discriminator's output probability and the true label (1 for real data, 0 for generated data). A smaller BCE indicates that the discriminator can accurately distinguish between real and generated data; conversely, a larger BCE indicates that the discriminator misclassifies, prompting the network to adjust its parameters to improve the quality of discrimination and generation.

[0087] In this invention, the weight coefficients corresponding to MSE and BCE are determined based on an adaptive weighting function to balance the impact of the amplitude difference between near-field and far-field data on model training. During the propagation of an explosion shock wave, near-field and far-field data have different characteristics. In the near-field region, the energy of the explosion shock wave is concentrated, resulting in large and rapidly changing data amplitudes; in the far-field region, the shock wave energy decays, leading to relatively smaller and more gradual data amplitude changes. If fixed weight coefficients are used, it may not be able to adapt well to the characteristics of data in different regions, resulting in poor model performance in full-range prediction.

[0088] In this invention, the adaptive weighting function is constructed based on the amplitude difference between near-field and far-field data during the propagation of the explosion shock wave. Specifically, it analyzes factors such as the proportional relationship and changing trends of the near-field and far-field data amplitudes. For example, when the difference between the near-field and far-field data amplitudes is large, the adaptive weighting function may adjust the weight coefficients of MSE and BCE, so that more attention is paid to the regions with large amplitude differences during training, ensuring that the model can also accurately predict in these regions. For example, when the near-field data amplitude is much larger than the far-field data amplitude, the weight of the loss term related to the near-field data (involving the part of MSE or BCE associated with the near-field data) may be appropriately increased to ensure that the model can better learn the features of the near-field data, while not neglecting the training of the far-field data.

[0089] This invention introduces a loss function composed of MSE and BCE, and utilizes an adaptive weighting function to dynamically determine the weight coefficients, ensuring the robustness of the model in full-range prediction. The adaptive weighting function can rationally allocate the weights of MSE and BCE in the loss function based on the amplitude differences between near-field and far-field data. This allows the model to fully learn the characteristics of data from different regions during training, achieving a good fit whether it's high-amplitude, rapidly changing data in the near field or low-amplitude, gradually changing data in the far field. This dynamic weight adjustment helps solve the instability and convergence difficulties commonly encountered in CGAN model training. It avoids the large fluctuations and convergence problems caused by fixed weights failing to adapt to data changes, enabling the model to train more stably. This improves the model's prediction accuracy and reliability under various explosion and geological conditions, thus enhancing its robustness in full-range prediction.

[0090] The ground impact load prediction model constructed in this invention achieves rapid and accurate prediction of the time history curves of explosive ground impact velocities under different geological conditions. This model innovatively introduces an adaptive weighting function (AWF) and a multi-level verification mechanism, significantly improving prediction accuracy and model robustness. It also possesses excellent engineering applicability and interpretability, constructing an end-to-end modeling system from "data generation" to "intelligent prediction," deeply integrating physical mechanisms and data-driven methods, thus opening a new paradigm for the rapid assessment of explosive impact loads. Specifically, this is reflected in the following aspects:

[0091] I. By utilizing Conditional Generative Adversarial Networks (CGAN), we can deeply explore the advantages of deep learning in modeling nonlinear relationships, accurately capture the complex coupling relationship between explosion parameters, geological parameters and time history curves of ground impact loads, and thus achieve high-precision prediction results.

[0092] Second, compared with the inefficient performance of traditional finite element method (FEM) which takes about 5 hours for a single simulation, the present invention can generate the ground impact load time history curve within seconds after the model training is completed, which significantly reduces the calculation time and fully meets the urgent needs of rapid response and real-time decision-making in wartime.

[0093] Third, through learning from massive amounts of simulation data, this invention demonstrates a powerful generalization ability, capable of adapting to prediction tasks under different explosive yields, distances from detonation, and complex geological conditions, effectively overcoming the limitation of existing empirical formulas being applicable only to single working conditions.

[0094] Fourth, this invention can not only serve as an efficient alternative to traditional finite element simulation for rapid prediction, but also be combined with finite element simulation to enhance data, further expanding the applicability of the model and providing a powerful auxiliary tool for the design of underground engineering and underground protective structures.

[0095] The ground impact load prediction system based on generative neural networks provided by this invention will be described below. The ground impact load prediction system based on generative neural networks described below can be referred to in correspondence with the ground impact load prediction method based on generative neural networks described above.

[0096] Figure 3 This is a schematic diagram of the structure of the ground impact load prediction system based on generative neural networks provided by the present invention, as shown below. Figure 3As shown, this invention provides a ground impact load prediction system based on a generative neural network, including a conditional parameter input module 301 and a ground impact load prediction module 302. The conditional parameter input module 301 is used to obtain the explosion condition target parameters corresponding to the target area. The ground impact load prediction module 302 is used to input the explosion condition target parameters into the ground impact load prediction model to obtain the ground impact load time history prediction curve corresponding to the target area output by the ground impact load prediction model. The ground impact load prediction model is obtained by training a conditional generative adversarial network based on the explosion condition sample parameters and the ground impact load time history sample curve.

[0097] The ground impact load prediction system based on generative neural networks provided by this invention obtains the target parameters of the explosion conditions in the target area, and then inputs them into the ground impact load prediction model constructed by a conditional generative adversarial network trained based on the explosion condition sample parameters and the ground impact load time history sample curve, thereby obtaining the ground impact load time history prediction curve of the target area, thus improving the prediction accuracy and efficiency of ground impact load information.

[0098] The system provided in this embodiment of the invention is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0099] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include: a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. The processor 401 can call logical instructions in the memory 403 to execute a ground impact load prediction method based on a generative neural network. The method includes: obtaining explosion condition target parameters corresponding to the target area; inputting the explosion condition target parameters into a ground impact load prediction model to obtain a ground impact load time history prediction curve corresponding to the target area output by the ground impact load prediction model, wherein the ground impact load prediction model is obtained by training a conditional generative adversarial network based on explosion condition sample parameters and ground impact load time history sample curves.

[0100] Furthermore, the logical instructions in the aforementioned memory 403 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the ground shock load prediction method based on a generative neural network provided by the above methods, the method comprising: obtaining explosion condition target parameters corresponding to a target area; inputting the explosion condition target parameters into a ground shock load prediction model to obtain a ground shock load time history prediction curve corresponding to the target area output by the ground shock load prediction model, wherein the ground shock load prediction model is obtained by training a conditional generative adversarial network based on explosion condition sample parameters and ground shock load time history sample curves.

[0102] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the ground impact load prediction method based on a generative neural network provided in the above embodiments. The method includes: obtaining explosion condition target parameters corresponding to a target area; inputting the explosion condition target parameters into a ground impact load prediction model to obtain a ground impact load time history prediction curve corresponding to the target area output by the ground impact load prediction model, wherein the ground impact load prediction model is obtained by training a conditional generative adversarial network based on explosion condition sample parameters and ground impact load time history sample curves.

[0103] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting ground impact loads based on generative neural networks, characterized in that, include: Obtain the target parameters for the explosion conditions corresponding to the target area; The target parameters of the explosion conditions are input into the ground shock load prediction model to obtain the ground shock load time history prediction curve corresponding to the target area output by the ground shock load prediction model. The ground shock load prediction model is obtained by training a conditional generative adversarial network based on the explosion condition sample parameters and the ground shock load time history sample curve. The ground impact load prediction model is trained through the following steps: The explosion condition sample parameters are constructed based on geological sample parameters and explosion sample parameters; Obtain the time history sample curve of the ground impact load corresponding to the explosion condition sample parameters; The explosion condition sample parameters and the ground impact load time history sample curve are input into the conditional generative adversarial network, and the preset iterative steps are repeatedly executed to update the parameters of the conditional generative adversarial network until the iterative training results meet the preset training conditions. The conditional generative adversarial network obtained in the last iteration is used as the ground impact load prediction model. The iterative steps specifically include: The explosion condition sample parameters of the current round are input into the generator in the conditional generative adversarial network of the current round to obtain the ground impact load time history generation curve of the current round output by the generator of the current round. The explosion condition sample parameters of the current wheel, the ground impact load time history generation curve of the current wheel, and the ground impact load time history sample curve corresponding to the explosion condition sample parameters of the current wheel are input into the discriminator in the conditional generative adversarial network of the current wheel to obtain the ground impact load time history prediction and discrimination result of the current wheel output by the discriminator of the current wheel. Based on the time history prediction and discrimination results of the ground impact load of the current wheel, the parameters of the generator in the conditional generative adversarial network of the current wheel are updated.

2. The method for predicting ground impact loads based on generative neural networks according to claim 1, characterized in that, The process of obtaining the time history sample curve of the ground impact load corresponding to the explosion condition sample parameters includes: A two-dimensional axisymmetric model for free-field confined explosion is constructed based on the fluid-structure interaction algorithm. Based on the two-dimensional axisymmetric model and the explosion condition sample parameters, the ground impact load time history simulation is performed to obtain the ground impact load time history sample curve corresponding to the explosion condition sample parameters.

3. The method for predicting ground impact loads based on generative neural networks according to claim 1, characterized in that, The construction of the explosion condition sample parameters based on geological sample parameters and explosion sample parameters includes: Based on the Latin hypercube sampling method, the parameters of multiple geological samples and multiple explosion samples are combined to obtain the explosion condition sample parameters under different parameter combinations.

4. The method for predicting ground impact loads based on generative neural networks according to any one of claims 1 to 3, characterized in that, The loss function of the conditional generative adversarial network consists of mean squared error and binary cross-entropy, wherein the weight coefficients corresponding to the mean squared error and the binary cross-entropy are determined based on an adaptive weighting function; the adaptive weighting function is constructed based on the amplitude difference between near-field data and far-field data during the propagation of the explosion shock wave.

5. A ground impact load prediction system based on a generative neural network, characterized in that, The system is used to implement the ground impact load prediction method based on generative neural networks as described in any one of claims 1 to 4.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the ground impact load prediction method based on a generative neural network as described in any one of claims 1 to 4.

Images