CN-121477311-B - Seismic wave propagation characteristic analysis and prediction system

CN121477311BCN 121477311 BCN121477311 BCN 121477311BCN-121477311-B

Abstract

The invention relates to the technical field of machine learning, in particular to a seismic wave propagation characteristic analysis and prediction system. According to the invention, the intrinsic response characteristics of the characterization medium are separated from the actually-collected multi-channel seismic data channel set, a series of possible velocity model samples are constructed probabilistically by utilizing a condition generation model based on the characteristics, and then forward modeling is carried out on all the velocity model samples, so that a wave field state snapshot sequence set containing multiple possibilities is obtained, expected propagation amplitude can be obtained and the uncertainty of prediction can be quantized through statistical calculation on the set, and finally, a comprehensive map of expected wave field propagation and uncertainty analysis is generated.

Inventors

WANG SHAOPENG
GUO MEI
Chi Peihong
WANG CHANGYA
SUN HAIYANG
LI WUQUAN
YAO YAO
HE QIANGLONG
YANG HAOZHI
WU XUELI
LU YUXIA
MEI DONGLIN
CHEN YEQIN
LIU KUN
Di Guorong
ZHANG QIANG
LEI ZHENGCHAO
HAN XIN

Assignees

甘肃省地震局（中国地震局兰州地震研究所）

Dates

Publication Date: 20260512
Application Date: 20251117

Claims (10)

1. A seismic wave propagation characterization and prediction system, the system comprising: the characteristic decoupling coding module is used for acquiring a plurality of seismic data gathers, and separating medium eigenvectors from the plurality of seismic data gathers by utilizing a gradient inversion layer arranged in the decoupling coder; The condition model generation module is used for generating a basic priori information vector each time based on the medium eigenvalue response feature vector, calculating a speed model sample for the basic priori information vector by using a condition normalization flow model, and collecting all the speed model samples as a speed model sample set; The forward propagation prediction module is used for traversing the velocity model sample set to execute forward propagation simulation calculation, recording wave field state snapshots to generate wave field state snapshot sequences, and combining the wave field state snapshot sequences into a wave field state snapshot sequence set; and the characteristic quantization analysis module is used for aligning time steps of the wave field state snapshot sequence set, traversing all coordinate points, extracting wave field amplitude values of the coordinate points on each time step, calculating the mean value of the wave field amplitude values to obtain expected propagation amplitude, calculating the standard deviation of the wave field amplitude values to obtain prediction uncertainty, and generating an expected wave field propagation sequence and prediction uncertainty map.
2. The seismic wave propagation characterization and prediction system of claim 1, wherein the medium eigenvalue response eigenvector comprises a medium velocity structure, a medium density distribution, a medium attenuation property, the velocity model sample specifically being a P-wave velocity field, an S-wave velocity field, a medium density field, the wave field state snapshot sequence comprising a simulated time span, a coordinate point, an amplitude time sequence, the expected wave field propagation sequence comprising an expected wave front position, an average energy distribution, a propagation phase, the prediction uncertainty map comprising a velocity model divergence, a prediction result reliability, a survey target area.
3. The seismic wave propagation signature analysis and prediction system of claim 1, wherein the specific functions of the signature decoupling encoding module are implemented as: the gather preprocessing sub-module is used for acquiring a plurality of seismic data gathers, performing time domain gain compensation and band-pass filtering processing on the plurality of seismic data gathers, eliminating random noise and power frequency interference mixed in the acquisition process, and then performing energy normalization operation on all the gathers by gather to ensure that the energy scales among the different gathers are consistent and generate a standardized gather; The mixed characteristic coding sub-module is used for acquiring the standardized gather, utilizing a built-in one-dimensional convolutional neural network as a shared encoder, carrying out depth characteristic extraction on the standardized gather along the time dimension, capturing space-time coupling information of seismic waves in the propagation process, and generating a mixed characteristic map; The gradient inversion sub-module inputs the mixed characteristic spectrum into the gradient inversion layer, and the gradient inversion layer is utilized to invert a gradient signal and then transmit the gradient signal to a parallel domain discriminator network, the domain discriminator network calculates domain classification loss of the mixed characteristic spectrum, forces the shared encoder to learn and extract characteristics irrelevant to an acquisition domain in an countermeasure training mode, and generates domain irrelevant characteristics; The eigenvalue response decoding submodule inputs the domain independent features into a special eigenvalue response decoder, and the eigenvalue response decoder reconstructs and separates medium eigenvalue response eigenvectors only related to physical properties of the underground medium from the domain independent features through deconvolution and nonlinear activation operation.
4. The seismic wave propagation signature analysis and prediction system of claim 1, wherein the specific functions of the condition model generation module are implemented as: The prior information construction submodule takes the medium eigenvalue response characteristic vector as a condition input, captures the internal time sequence dependency of the medium eigenvalue response characteristic vector through a cyclic neural network structure, and generates a basic prior information vector on each time step of the cyclic neural network; A conditional flow model calculation sub-module, in each cycle, inputting the basic prior information vector into the conditional normalized flow model, wherein the conditional normalized flow model uses a reversible transformation function to nonlinearly map and transform a random noise vector sampled from standard Gaussian distribution into a velocity model sample under the constraint of the basic prior information vector; And the sample collection sub-module is used for setting a preset total number of cycle generation, repeatedly executing the prior information construction and the conditional flow model calculation until a speed model sample reaching the total number of cycle generation is generated, collecting and storing all the generated speed model samples, and generating a speed model sample set.
5. The seismic wave propagation characterization and prediction system of claim 1, wherein the specific functions of the forward propagation prediction module are implemented as: The sample traversing simulation sub-module is used for acquiring the velocity model sample set, setting seismic source parameters and boundary absorption conditions, traversing each velocity model sample in the velocity model sample set, constructing a corresponding medium model based on the velocity model sample, solving a sound wave or elastic wave equation by using a finite difference method, and executing forward propagation simulation calculation; The wave field state recording submodule monitors and records wave field amplitude states in the whole calculation grid area on each time step of forward propagation simulation calculation, stores the wave field amplitude states as wave field state snapshots, and combines the wave field state snapshots on all the time steps in time sequence to generate a wave field state snapshot sequence; And the sequence set constructing sub-module repeatedly executes the wave field state record until forward propagation simulation calculation of all the velocity model samples in the velocity model sample set is completed, uniformly storing and indexing all the generated wave field state snapshot sequences, and combining to generate a wave field state snapshot sequence set.
6. The seismic wave propagation signature analysis and prediction system of claim 1, wherein the specific functions of the signature quantification analysis module are implemented as: The time-space alignment sub-module is used for acquiring the wave field state snapshot sequence set, aligning time steps of all wave field state snapshot sequences by taking common starting time of all wave field state snapshot sequences in the wave field state snapshot sequence set as a reference, and ensuring that all wave field state snapshots on the same time step correspond to the same simulation time; the grid amplitude extraction submodule traverses all coordinate points in the wave field state snapshot, extracts all wave field amplitude values corresponding to the same coordinate points in the wave field state snapshot sequence set on each time step, and forms an amplitude distribution set of the coordinate points on the time step; A statistical quantization calculation sub-module traversing all the time steps and all coordinate points, performing statistical analysis on all the wave field amplitude values in the amplitude distribution set, calculating a mathematical mean of the wave field amplitude values to obtain the expected propagation amplitude of the coordinate points at the time steps, and calculating a standard deviation of the wave field amplitude values to obtain the prediction uncertainty of the coordinate points at the time steps; a result map generation sub-module that combines the expected propagation amplitudes over all the time steps in a time-space order to generate the expected wave field propagation sequence; And performing spatial interpolation and visual rendering on the prediction uncertainty on all time steps to generate the prediction uncertainty map.
7. A seismic wave propagation signature analysis and prediction system as claimed in claim 3, wherein the gradient inversion pair sub-module, when calculating the domain classification loss, comprises: The domain discriminator network receives the mixed characteristic spectrum transmitted by the gradient inversion layer, classifies the mixed characteristic spectrum by utilizing a multi-layer perceptron, judges the acquisition domain from which target the mixed characteristic spectrum originates, and calculates the cross entropy loss between the mixed characteristic spectrum and a real acquisition domain label as the domain classification loss; the domain classification loss is automatically multiplied by a negative scalar by the gradient inversion layer when back propagated to the shared encoder, and the shared encoder maximizes the domain classification loss while minimizing the self reconstruction loss; Extracting the domain independent features using the shared encoder.
8. The seismic wave propagation signature analysis and prediction system according to claim 4, wherein the conditional normalized flow model, when mapping and transforming the random noise vector, comprises: Dividing the random noise vector into a first vector portion and a second vector portion; Driving a scaling function using the basic prior information vector as a condition Translation function Calculating the first vector part to obtain a scaling parameter and a translation parameter; performing an affine transformation on the second vector portion using the scaling parameters and the translation parameters: ; And will be transformed And unchanged Splicing the output vectors; Realizing high-dimensional complex transformation by stacking a plurality of affine coupling layers, and generating the velocity model sample; Wherein, the A first vector portion representing the random noise vector, A second vector portion representing the random noise vector, Representing the vector of basic prior information, Representing a first portion of the output vector of the coupling layer, A second portion representing an output vector of the coupling layer, Representing the function of said scaling in question, Representing the function of the translation in question, Representing element-wise multiplication.
9. The system of claim 5, wherein the finite difference method for solving acoustic or elastic wave equations comprises: Initializing an interlaced grid based on a P-wave velocity field, an S-wave velocity field and a medium density field in the velocity model sample; Setting seismic source parameters on the staggered grid, including source type, center frequency, excitation time and spatial position; Setting a convolution perfect matching layer at the boundary of the calculation grid area as a boundary absorption condition; And adopting a differential format of a second order in time and a higher order in space, iteratively solving a wave equation on a time axis, and calculating the displacement or stress state of each grid point at the next moment.
10. The seismic wave propagation signature analysis and prediction system according to claim 6, wherein the predicting uncertainty further relates to a quantized prediction result reliability in the prediction uncertainty map comprising: Setting an uncertainty threshold, and marking a region with the prediction uncertainty lower than the uncertainty threshold as a high-reliability region in the prediction uncertainty map; the regions where the prediction uncertainty is above the uncertainty threshold are marked as low reliability regions.

Description

Seismic wave propagation characteristic analysis and prediction system Technical Field The invention relates to the technical field of machine learning, in particular to a seismic wave propagation characteristic analysis and prediction system. Background The field of machine learning technology is the core branch of artificial intelligence that utilizes algorithms to enable computer systems to learn and refine from input data without explicit programming. The field covers various methods such as supervised learning, unsupervised learning and reinforcement learning, and is widely applied to data mining, pattern recognition and predictive modeling of complex systems. The traditional seismic wave propagation characteristic analysis and prediction system refers to a computing device which depends on a geophysical model and numerical simulation. They typically employ time-domain finite difference methods or ray tracing methods to calculate the response of seismic waves excited by the source in the subsurface medium. The wave field state is gradually deduced on the gridding model by the wave equation solver through inputting known geological structure parameters and seismic source parameters, so that the waveform data of a specific position is obtained. The traditional seismic wave propagation characteristic analysis and prediction system is highly dependent on preset, single determined geologic structure parameters for numerical simulation, however, the complexity of underground media and the incompleteness of exploration data make the construction of a unique and accurate model extremely difficult, and the dependence on a deterministic model leads forward modeling calculation to only output one possible wave field propagation result, and cannot reflect the wave field response change range caused by geologic structure uncertainty, so that the prediction result lacks quantitative evaluation of uncertainty, is difficult to truly reflect various possibilities of wave field propagation, and influences the reliability of subsequent analysis decisions. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a seismic wave propagation characteristic analysis and prediction system. In order to achieve the above purpose, the invention adopts the following technical scheme that the seismic wave propagation characteristic analysis and prediction system comprises: the characteristic decoupling coding module is used for acquiring a plurality of seismic data gathers, and separating medium eigenvectors from the plurality of seismic data gathers by utilizing a gradient inversion layer arranged in the decoupling coder; The condition model generation module is used for generating a basic priori information vector each time based on the medium eigenvalue response feature vector, calculating a speed model sample for the basic priori information vector by using a condition normalization flow model, and collecting all the speed model samples as a speed model sample set; The forward propagation prediction module is used for traversing the velocity model sample set to execute forward propagation simulation calculation, recording wave field state snapshots to generate wave field state snapshot sequences, and combining the wave field state snapshot sequences into a wave field state snapshot sequence set; and the characteristic quantization analysis module is used for aligning time steps of the wave field state snapshot sequence set, traversing all coordinate points, extracting wave field amplitude values of the coordinate points on each time step, calculating the mean value of the wave field amplitude values to obtain expected propagation amplitude, calculating the standard deviation of the wave field amplitude values to obtain prediction uncertainty, and generating an expected wave field propagation sequence and prediction uncertainty map. As a further aspect of the present invention, the medium eigenvalue response feature vector includes a medium velocity structure, a medium density distribution, and a medium attenuation attribute, the velocity model sample specifically refers to a P-wave velocity field, an S-wave velocity field, and a medium density field, the wave field state snapshot sequence includes a simulation time span, a coordinate point, and an amplitude time sequence, the expected wave field propagation sequence includes an expected wave front position, an average energy distribution, and a propagation phase, and the prediction uncertainty map includes a velocity model divergence, a prediction result reliability, and an exploration target area. As a further aspect of the present invention, the specific function of the feature decoupling encoding module is implemented as follows: the gather preprocessing sub-module is used for acquiring a plurality of seismic data gathers, performing time domain gain compensation and band-pass filtering processing on the plurality of seismic data gathers