CN-121706607-B - Method for predicting transient electromagnetic response of core wall dam

CN121706607BCN 121706607 BCN121706607 BCN 121706607BCN-121706607-B

Abstract

The invention discloses a core wall dam transient electromagnetic response prediction method which comprises the steps of constructing a core wall dam boundary parameter space containing geometric dimension characteristics and dielectric electrical characteristics of a core wall dam, generating a response constraint sample set for constraining a dam body structure and electromagnetic field diffusion relation according to the core wall dam boundary parameter space, constructing an electromagnetic response mapping network for executing nonlinear transformation from the parameter space to the electromagnetic response space, iteratively updating network parameters by minimizing a loss function containing physical field constraint or data fitting errors to obtain a converged electromagnetic response mapping network, and inputting a target boundary parameter vector to be predicted into the converged electromagnetic response mapping network to output a transient electromagnetic vertical magnetic field component attenuation curve corresponding to a time domain when the core wall dam transient electromagnetic response prediction method works. The invention solves the problem of network convergence in a strong heterogeneous medium through the joint constraint of the partition interfaces, ensures the physical consistency of derivatives by utilizing the constraint of a sensibility equation, and realizes the high-precision and real-time intelligent prediction of the heart wall dam diseases.

Inventors

ZHANG SHENGXING
TANG LEI
LIU YANZE
FENG YAN
Shi Lanxing
LI PO
GUAN FUHAI
Luan Pinzhi
LIANG JIAHUI

Assignees

水利部交通运输部国家能源局南京水利科学研究院

Dates

Publication Date: 20260508
Application Date: 20260211

Claims (9)

1. The method for predicting the transient electromagnetic response of the core wall dam is characterized by comprising the following steps of: constructing a boundary parameter space of the core wall dam, wherein the boundary parameter space comprises geometric dimension characteristics and dielectric electrical characteristics of the core wall dam; generating a response constraint sample set for constraining the diffusion relation between the dam body structure and the electromagnetic field based on the boundary parameter space of the core wall dam; constructing an electromagnetic response mapping network for performing a nonlinear transformation from a parameter space to an electromagnetic response space; Training the electromagnetic response mapping network by using a response constraint sample set, and iteratively updating network parameters by minimizing a loss function containing physical field constraint or data fitting errors to obtain a converged electromagnetic response mapping network; when the method works, a target boundary parameter vector to be predicted is input into a converged electromagnetic response mapping network, and a transient electromagnetic vertical magnetic field component attenuation curve corresponding to a time domain is output; The electromagnetic response mapping network is a physical information neural network; The response constraint sample set comprises sub-domain internal sampling points distributed in the calculation domain and interface sampling points distributed at the junctions of different media; training the electromagnetic response mapping network with the response constraint sample set, comprising: Constructing subdomain physical residual loss based on a transient electromagnetic control equation by utilizing sampling points in the subdomain; constructing an interface continuity residual loss based on electromagnetic field boundary conditions by using the interface sampling points; and constructing a loss function by combining the subdomain physical residual loss, the interface continuity residual loss and the data fitting loss calculated based on prestored observation or simulation data.
2. The method of claim 1, wherein the physical subzones for which subzone physical residual losses are targeted include an upstream water subzone, a core wall subzone, a dam shell and transition zone subzone, and a bedrock subzone; Constructing an interface continuity residual loss, comprising: Calling an interface continuity residual operator, and calculating electromagnetic field components at two sides of an interface sampling point; Based on electromagnetic field components, the normal current density or tangential electric field between adjacent physical subdomains is constrained by using an interface continuity residual operator to keep continuous, so that interface continuity residual loss is obtained.
3. The method of claim 1, wherein training the electromagnetic response mapping network with the response constraint sample set further comprises: in the training iteration process, respectively counting the mean value, variance or gradient norm of the subdomain physical residual error loss, the interface continuity residual error loss and the data fitting loss to generate residual error statistical characteristics; processing residual statistics characteristics according to an adaptive weight scheduling strategy, and calculating dynamic weights for balancing contribution degrees of all loss items; And carrying out weighted summation on the subdomain physical residual loss, the interface continuity residual loss and the data fitting loss by using dynamic weights, and updating a loss function.
4. The method of claim 1, wherein the electromagnetic response mapping network is configured with an automatic differential interface; the method further comprises: while outputting the transient electromagnetic vertical magnetic field component attenuation curve, calculating the partial derivative of the output of the electromagnetic response mapping network relative to the target boundary parameter vector by using an automatic differentiation technology; From the partial derivatives, parameter sensitivity information is generated that characterizes the response as a function of the rate of change of the parameter.
5. The method according to claim 4, further comprising constructing sensitive physical consistency constraints, in particular: Performing symbol derivation on a target boundary parameter vector based on a transient electromagnetic control equation, and constructing a sensitivity equation residual operator for describing a physical dependency relationship between an electromagnetic response change rate and a parameter change rate; substituting the partial derivative into a sensitivity equation residual operator, and calculating sensitivity equation residual loss; and adding the sensitivity equation residual loss into a loss function, and performing response and sensitivity joint training on the electromagnetic response mapping network to restrict the partial derivative to meet the physical condition defined by the sensitivity equation residual operator.
6. The method of claim 1, wherein the electromagnetic response mapping network is a feed-forward neural network; constructing an electromagnetic response mapping network, comprising: Constructing a network topology comprising at least two hidden layers, wherein each hidden layer is configured with a predetermined number of neurons and employs tansig functions or logsig functions as activation functions; the Levenberg-Marquardt algorithm or the Bayesian regularization algorithm is adopted as a training algorithm, and the weight parameters of the feedforward neural network are updated by minimizing the data fitting error.
7. The method of claim 6, wherein generating a response constraint sample set comprises: setting a value range for each parameter in the boundary parameter space of the core wall dam, and carrying out parameter combination by adopting an orthogonal test design method to generate a representative working condition set; establishing a parameterized core dam finite element model aiming at each group of working conditions in the representative working condition set; And solving a preconfigured time domain Maxwell equation set by using a finite element model, calculating transient electromagnetic response in an observation time window, and combining the transient electromagnetic response with corresponding working condition parameters to form a response constraint sample set.
8. The method of claim 7, wherein the core dam boundary parameter space comprises a length-like parameter and a resistivity-like parameter; The method further comprises the step of normalizing parameters in a boundary parameter space of the core wall dam, and specifically comprises the following steps: mapping the length parameters to a preset numerical interval by adopting a linear normalization method; And carrying out logarithmic transformation on the resistivity parameters, and mapping the numerical values after logarithmic transformation to a preset numerical value interval by adopting a linear normalization method to obtain normalized boundary parameters.
9. The method of claim 8, wherein the core dam boundary parameter space comprises parameters of at least 8 dimensions; training the electromagnetic response mapping network with the response constraint sample set, comprising: calculating corresponding normalized logarithmic time for each time point in the observation time window; Splicing the normalized boundary parameters and the normalized logarithmic time to construct an input feature vector; And inputting the input feature vector into an electromagnetic response mapping network, and taking the logarithmic transformation value of the transient electromagnetic response corresponding to the time point as a training target to supervise and learn the electromagnetic response mapping network.

Description

Method for predicting transient electromagnetic response of core wall dam Technical Field The invention belongs to the field of geophysical exploration, and particularly relates to a method for predicting transient electromagnetic response of a core wall dam. Background The core wall dam is used as a core barrier of a water conservancy junction, nondestructive detection of internal leakage and crack hidden danger is important to ensure the safety of the dam, and Transient Electromagnetic Method (TEM) is a key means for detecting the hidden danger because of being sensitive to low-resistance bodies and large in detection depth. However, core dams have multiple structures of water-core-dam shell-bedrock and complex trapezoid boundaries, the diffusion rule of electromagnetic fields in such strong non-uniform media is extremely complex, and traditional detection interpretation is severely dependent on high-precision forward modeling. Currently, core dam TEM forward modeling mainly depends on numerical simulation methods such as Finite Element (FEM) or finite difference (FDTD). Although these methods can theoretically simulate complex structures, the computation process requires fine grid segmentation and large-scale matrix solution, and a single forward modeling tends to take hours. Although data-driven neural network methods have begun to be applied to geophysical inversion in recent years, most of them are limited to simple lamellar models or homogeneous half-space models, and the physical structural features of dams cannot be deeply fused by fitting input-output relationships only with simple black box networks. In the prior art, when a composite structure such as a core wall dam is processed, three deep technical bottlenecks of low calculation efficiency and unreliable interface physical discontinuity and sensitivity analysis exist. The method is characterized in that the high time cost of the traditional numerical method cannot meet the requirements of real-time inversion and mass working condition analysis of an engineering site, when a conventional Physical Information Neural Network (PINN) is introduced, the conductivity mutation crossing over a plurality of orders of magnitude exists between a core wall and a dam shell as well as between a water body and a bedrock, so that the network faces serious gradient pathological problems during training, convergence at a strong intermittent interface is not possible, continuity of physical quantities such as normal current density and the like is difficult to ensure, the conventional parameter sensitivity analysis is usually used as an independent post-processing step (such as a difference method), calculated sensitivity information often cannot meet the differential constraint of a control equation, and the gradient information lacking physical consistency is extremely easy to cause subsequent inversion to fall into a local minimum value, so that the accuracy of disease diagnosis is reduced. Disclosure of Invention The invention aims to provide a method for predicting transient electromagnetic response of a core wall dam, which aims to solve the problems in the prior art. The technical scheme is that the method for predicting the transient electromagnetic response of the core wall dam comprises the following steps: constructing a boundary parameter space of the core wall dam, wherein the boundary parameter space comprises geometric dimension characteristics and dielectric electrical characteristics of the core wall dam; generating a response constraint sample set for constraining the diffusion relation between the dam body structure and the electromagnetic field based on the boundary parameter space of the core wall dam; constructing an electromagnetic response mapping network for performing a nonlinear transformation from a parameter space to an electromagnetic response space; Training the electromagnetic response mapping network by using a response constraint sample set, and iteratively updating network parameters by minimizing a loss function containing physical field constraint or data fitting errors to obtain a converged electromagnetic response mapping network; When the method works, the target boundary parameter vector to be predicted is input into a converged electromagnetic response mapping network, and a transient electromagnetic vertical magnetic field component attenuation curve corresponding to a time domain is output. The method has the beneficial effects that the problem of network convergence in a strong heterogeneous medium is solved through partition interface joint constraint, the physical consistency of the derivative is ensured by using sensitivity equation constraint, and the high-precision and real-time intelligent prediction of the heart wall dam disease is realized. Drawings Fig. 1 is a flowchart of steps of a method for predicting a transient electromagnetic response of a core dam according to an embodiment of the present application.