CN111666721B

CN111666721B - Full-waveform inversion method and device and electronic equipment

Info

Publication number: CN111666721B
Application number: CN202010571206.5A
Authority: CN
Inventors: 王彦飞; 何清龙
Original assignee: Institute of Geology and Geophysics of CAS
Current assignee: Institute of Geology and Geophysics of CAS
Priority date: 2020-06-19
Filing date: 2020-06-19
Publication date: 2021-05-25
Anticipated expiration: 2040-06-19
Also published as: JP2022001861A; CN111666721A; JP7142968B2

Abstract

The invention provides a full waveform inversion method, a full waveform inversion device and electronic equipment, wherein the full waveform inversion method comprises the following steps: acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the underground structure; carrying out parameterization processing on the initial physical parameters to obtain physical parameters, and inputting the physical parameters into a machine learning model trained in advance so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relation; and reconstructing a velocity structure map of the subsurface structure from the full waveform inversion results. According to the method and the device, the initial physical parameters are parameterized through the weight coefficients of the machine learning model, so that the full waveform inversion problem is converted into the reconstruction problem of the network parameters of the machine learning model, the problems that the existing full waveform inversion cannot solve the nonlinear inverse problem and a large amount of calculation exists are solved, the calculation efficiency is improved, and the beneficial effect of inverting the complex underground structure with high precision is achieved.

Description

Full-waveform inversion method and device and electronic equipment

Technical Field

The invention relates to the technical field of exploration geophysical, in particular to a full waveform inversion method, a full waveform inversion device and electronic equipment.

Background

FWI (Full Waveform Inversion) is an Inversion method capable of reconstructing subsurface structures at high resolution, and its purpose is to find an optimal model that minimizes the objective function of observed data and synthesized data, usually by using local or global optimization strategies. From a computational point of view, the method of local optimization is superior to the global optimization method because the global strategy requires a large amount of computation.

In practical applications, it is still challenging to apply the local convergence algorithm to the conventional FWI method, because the FWI often has multiple local minima due to the high nonlinearity and the unsuitability of the objective function of the conventional FWI method, so that the local convergence algorithm is hindered by the multiple local minima, and the large-scale calculation is caused by the dimension, and therefore, the existing FWI method is not suitable for the nonlinear inverse problem, and has a large calculation amount, so that the inverted underground structure diagram is not ideal.

Disclosure of Invention

In view of the above, the present invention aims to provide a full waveform inversion method, a full waveform inversion device, and an electronic apparatus, so as to alleviate the above problems, and parameterize physical parameters through a weight coefficient of a machine learning model, so as to transform a full waveform inversion problem into a reconstruction problem of network parameters of the machine learning model, improve calculation efficiency, and achieve the beneficial effects of inverting a complex underground structure with high precision.

In a first aspect, an embodiment of the present invention provides a full waveform inversion method, which is applied to a server, and the method includes:

acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the underground structure;

carrying out parameterization processing on the initial physical parameters to obtain physical parameters;

inputting the physical parameters into a machine learning model trained in advance, so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relation;

and reconstructing a velocity structure chart of the underground structure according to the full waveform inversion result.

With reference to the first aspect, an embodiment of the present invention provides a first possible implementation manner of the first aspect, where the reconstructing a velocity structure map of the subsurface structure according to the full waveform inversion result includes:

forward modeling is carried out on the full waveform inversion result to obtain a full waveform inversion matrix;

and reconstructing to obtain a speed structure chart of the underground structure based on the full waveform inversion matrix.

With reference to the first possible implementation manner of the first aspect, an embodiment of the present invention provides a second possible implementation manner of the first aspect, where a full-waveform inversion result is a feature inversion result, and the reconstructing a velocity structure diagram of the subsurface structure according to the full-waveform inversion result includes:

forward modeling is carried out on the feature inversion result to obtain a feature inversion matrix;

and reconstructing to obtain a characteristic diagram of the underground structure based on the characteristic inversion.

With reference to the first aspect, an embodiment of the present invention provides a third possible implementation manner of the first aspect, where the step of performing parameterization on the initial physical parameter to obtain the physical parameter includes: carrying out parameterization processing on the initial physical parameters according to the following formula to obtain the physical parameters:

m(w)＝m＝G(w)

wherein m represents an initial physical parameter, g (w) represents a parameterized polynomial, w represents a weight coefficient, and m (w) represents a physical parameter.

With reference to the first aspect, an embodiment of the present invention provides a fourth possible implementation manner of the first aspect, where the machine learning model is a model trained based on a deep neural network, and the method further includes:

acquiring an original training physical parameter set; wherein the original training set of physical parameters comprises velocity parameters of a plurality of subsurface media in the subsurface structure;

and inputting the original training physical parameter set into a deep neural network for training to obtain a machine learning model.

With reference to the fourth possible implementation manner of the first aspect, an embodiment of the present invention provides a fifth possible implementation manner of the first aspect, where the step of inputting the original training physical parameter set to the deep neural network for training to obtain the machine learning model includes:

inputting the original training physical parameter set into a deep neural network to obtain a full waveform inversion result of the original training physical parameter set;

calculating a function value of an objective function of the deep neural network based on a full waveform inversion result of the original training physical parameter set;

and adjusting the parameters of the deep neural network through the function value of the objective function to obtain a machine learning model.

In a second aspect, an embodiment of the present invention further provides a full waveform inversion apparatus, applied to a server, where the apparatus includes:

the acquisition module is used for acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the underground structure;

the parameterization module is used for carrying out parameterization processing on the initial physical parameters to obtain physical parameters;

the input and output module is used for inputting the physical parameters into a machine learning model trained in advance so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relation;

and the reconstruction module is used for reconstructing a speed structure chart of the underground structure according to the full waveform inversion result.

With reference to the second aspect, an embodiment of the present invention provides a first possible implementation manner of the second aspect, where the machine learning model is a model obtained by training based on a deep neural network, and the apparatus further includes: acquiring an original training physical parameter set; wherein the original training set of physical parameters comprises velocity parameters of a plurality of subsurface media in the subsurface structure; and inputting the original training physical parameter set into a deep neural network for training to obtain a machine learning model.

In a third aspect, an embodiment of the present invention further provides an electronic device, including a processor and a memory, where the memory stores computer-executable instructions capable of being executed by the processor, and the processor executes the computer-executable instructions to implement the full waveform inversion method of the first aspect.

In a fourth aspect, embodiments of the present invention also provide a computer-readable storage medium having stored thereon a computer-executable program, which, when invoked and executed by a processor, causes the processor to implement the full waveform inversion method of the first aspect.

The embodiment of the invention has the following beneficial effects:

the embodiment of the invention provides a full waveform inversion method, a full waveform inversion device and electronic equipment, wherein the full waveform inversion method comprises the following steps: acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the underground structure; carrying out parameterization processing on the initial physical parameters to obtain physical parameters, and inputting the physical parameters into a machine learning model trained in advance so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relation; and reconstructing a velocity structure map of the subsurface structure from the full waveform inversion results. According to the method and the device, the initial physical parameters of the underground structure are parameterized through the weight coefficients of the machine learning model, so that the full waveform inversion problem is converted into the reconstruction problem of the network parameters of the machine learning model, the problems that the existing full waveform inversion cannot solve the nonlinear inverse problem and a large amount of calculation exists are solved, the calculation efficiency is improved, and the beneficial effect of inverting the complex underground structure with high precision is achieved.

Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and drawings.

In order to make the aforementioned and other objects, features and advantages of the present invention comprehensible, preferred embodiments accompanied with figures are described in detail below.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.

Fig. 1 is a flowchart of a full waveform inversion method according to an embodiment of the present invention;

FIG. 2 is a flow chart of another method for full waveform inversion according to an embodiment of the present invention;

FIG. 3 is a flowchart of a training method of a machine learning model according to an embodiment of the present invention;

FIG. 4 is a diagram illustrating a machine learning model according to an embodiment of the present disclosure;

fig. 5 is a schematic diagram of a true model of a Marmousi2 model according to an embodiment of the present invention;

fig. 6 is a schematic diagram of a smooth model of a Marmousi2 model according to an embodiment of the present invention;

fig. 7 is a schematic diagram of an output model of a Marmousi2 model according to an embodiment of the present invention;

fig. 8-a is a vertical cross-sectional view of a Marmousi2 model at 3.0km, according to an embodiment of the present invention;

fig. 8-b is a vertical cross-sectional view of a Marmousi2 model at 5.0km, according to an embodiment of the present invention;

fig. 8-c is a vertical cross-section of a Marmousi2 model at 6.0km according to an embodiment of the present invention;

fig. 9 is a diagram of a reconstruction result of a Marmousi2 model according to an embodiment of the present invention;

FIG. 10 is a schematic diagram of a real model of a mineral model according to an embodiment of the present invention;

FIG. 11 is a schematic illustration of a smoothed model of a mineral model according to an embodiment of the present invention;

FIG. 12 is a schematic diagram of a reconstruction result of a mineral model according to an embodiment of the present invention;

FIG. 13 is a schematic diagram of a reconstruction result of another mineral model according to an embodiment of the present invention;

FIG. 14-a is a vertical cross-section of a mineralogical model at 1.2km according to an embodiment of the present invention;

FIG. 14-b is a vertical cross-section of a mineral model at 2.8km according to an embodiment of the present invention;

FIG. 14-c is a vertical cross-section of a mineral model at 3.8km according to an embodiment of the present invention;

FIG. 15 is a schematic illustration of a smoothed model of another mineral model provided by an embodiment of the invention;

FIG. 16 is a schematic illustration of a seismic record of a mineralogical model provided in accordance with an embodiment of the present invention;

FIG. 17 is a schematic representation of a seismic recording of another mineralogical model provided in accordance with an embodiment of the present invention;

FIG. 18 is a schematic diagram of a reconstruction result of another mineral model according to an embodiment of the present invention;

FIG. 19 is a schematic diagram of a reconstruction result of another mineral model according to an embodiment of the present invention;

FIG. 20 is a comparison graph of computational performance provided by embodiments of the present invention;

FIG. 21-a is a vertical cross-section at 1.2km of another mineralogical model provided in accordance with an embodiment of the present invention;

FIG. 21-b is a vertical cross-section at 2.8km of another mineralogical model provided in accordance with an embodiment of the present invention;

FIG. 21-c is a vertical cross-section at 3.8km of another mineralogical model provided in accordance with an embodiment of the present invention;

FIG. 22 is a schematic diagram of a full waveform inversion apparatus according to an embodiment of the present invention;

fig. 23 is a schematic view of an electronic device according to an embodiment of the present invention.

Detailed Description

To make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

FWI is a high-resolution seismic inversion method that aims to find an optimal model that minimizes the objective function of observed and synthetic data, usually by using local or global optimization strategies. In practice, from a computational point of view, the local optimization method is superior to the global optimization method because the global strategy requires a large amount of computation. Over the past few decades, many local optimization methods have been introduced into FWI problems, such as the steepest descent method and its accelerated version, the nonlinear conjugate gradient method, the quasi-newton method, and the truncated newton method.

Therefore, it is still challenging to apply the local convergence algorithm to the conventional FWI method, which tends to have multiple local minima due to the high non-linearity and non-qualification of the objective function of the conventional FWI method, thus causing the local convergence algorithm to be hindered by the multiple local minima, and the large-scale computation caused by the dimension. This means that the conventional FWI method requires low frequency information or a good initial model when using a locally convergent algorithm due to the limitation of the quadratically integrable objective function in terms of seismic wavefield phase matching capability. In order to solve the above-mentioned multiple minima problems, various elaborate methods are proposed in succession, including regularization methods, Wasserstein measurement-based methods, wavefield reconstruction inversion, adaptive waveform inversion, and extended modeling methods, etc. Aiming at the large-scale calculation of the FWI, on one hand, a fast parallel forward modeling problem solving method can be adopted to solve a wave field model, the wave field model comprises a direct solver and an iterative solver, and on the other hand, an effective method can be developed to estimate the gradient of an objective function so as to reduce the times of wave field simulation.

At present, DNN (Deep Neural Networks) is a powerful and widely applied technology. Since the renaissance of DNN in the machine learning community in 2006, with the increasing modern computing power and the efficient implementation of back-propagation methods, DNN has been successfully applied in many fields such as computer vision and speech recognition. In practical applications, the success of DNN can be attributed to the universal approximation theorem, and as of now, DNN and other data science techniques have attracted attention in solving the inverse problem. In geophysical applications, DNN has been used for fault detection, low frequency reconstruction, and velocity model construction. However, in practical applications, a large number of high-quality labeled training data sets are required, and a reasonable optimization algorithm is correctly executed, so that the DNN can generate appropriate results for fault detection, low-frequency reconstruction, and velocity model construction, such as: the system comprises a GAN-FWI based on GAN (generic adaptive Networks), an RNN-FWI based on RNN (Recurrent Neural Networks) and a CNN-FWI based on Convolutional Neural Networks, wherein the RNN-FWI and the CNN-FWI are efficient compared with the GAN-FWI and do not need a large number of mark data sets in the training process, so that the RNN-FWI and the CNN-FWI have the capability of solving the inverse problem of the large-scale FWI.

Based on the general approximation theorem of GAN-FWI, RNN-FWI and CNN-FWI, the application provides a general inversion method, namely DNN-FWI, under the DNN framework, and for the inversion method, the physical parameters are parameterized by using the weight coefficient of DNN; therefore, the full waveform inversion problem is converted into the reconstruction problem of the machine learning model network parameters, and therefore, the weight coefficient of the neural network can be effectively determined by using a random optimization method widely used in the DNN field; and, DNN-FWI can also be used as an iterative regularization method by introducing neural network layer constraints with specific functions, e.g., convolutional layers can extract boundary feature information, and deep feed-forward neural networks can be considered as an iterative regularization method. Furthermore, unlike other methods, the DNN-FWI iterative framework performs repeated wavefield forward modeling, thereby improving data consistency and improving resolution of the inversion results. Furthermore, DNN-FWI can also parameterize physical parameters with any network architecture and introduce specific network layers for regularization to capture the characteristics of physical parameters, such as ability of convolutional layers to adapt to sharp boundaries, etc., as compared to CNN-FWI.

Therefore, according to the full waveform inversion method, the full waveform inversion device and the electronic equipment provided by the embodiment of the invention, the physical parameters are firstly obtained; the physical parameters are parameterized through the weight coefficients of the machine learning model, so that the full waveform inversion problem is converted into the reconstruction problem of the network parameters of the machine learning model, the problems that the existing full waveform inversion cannot solve the nonlinear inverse problem and a large amount of calculation exists are solved, the calculation efficiency is improved, and the beneficial effect of inverting the complex underground structure with high precision is achieved.

To facilitate understanding of the present embodiment, a detailed description of a full waveform inversion method provided by the present embodiment is first provided below.

Example one

The embodiment of the invention provides a full waveform inversion method, wherein an execution main body is a server, and as shown in fig. 1, the method comprises the following steps:

step S102, acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the subsurface structure.

In practical application, specific initial physical parameters of the underground structure are unknown and cannot be observed directly by an observer, wherein the initial physical parameters mainly refer to a velocity structure of the underground structure, that is, a velocity structure composed of velocity parameters of each underground medium in the underground structure, and therefore, the existing method mainly obtains a velocity structure diagram of the underground structure by inversion of a full-waveform inversion method.

In the conventional full waveform inversion method, the target function of the FWI method is set to f (m), which is used to measure the difference between observed data and simulated data, and for the sake of simplifying notation, it is assumed here that there is only one seismic source, and for the problem of multiple seismic sources, the multiple seismic sources are simply added according to formula (1), wherein the target function f (m) can be calculated according to the following formula:

wherein R represents the projection operator, d_obsFor the observation data, u (m) represents simulation data, i.e. data obtained by simulation according to initial physical parameters, u is a wave field propagated by seismic waves, s represents a seismic source function, m is an initial physical parameter, i.e. a velocity structure of the underground medium, and F represents a differential operator.

At this time, in consideration of the calculation amount problem, the constraint minimization problem in the above equation (1) is generally solved by a gradient type or newton type method. In the gradient-type or newton-type method, the physical parameter m of the objective function is updated by an iterative method. Wherein, for the (k +1) th iteration, the initial physical parameter m is updated by:

m_k+1＝m_k+α_kp_k (2)

wherein m is_k+1Denotes the initial physical parameter, m, of the (k +1) th iteration_kRepresenting the initial physical parameter, α, of the kth iteration_kDenotes the step size, k denotes the number of iteration steps, p_kIndicating the direction of the step decrease. For this equation (2), one can choose

Thereby obtaining the steepest descent method; or, select

To obtain a Newtonian type of process, in which B_kIs a Hessian matrix H_kA positive definite approximation of the symmetry.

Due to the high nonlinearity and the ill-qualification of the objective function f (m) of the FWI method, the FWI tends to have a plurality of local minima, which causes the local convergence algorithm to be hindered by the local minima, and the FWI also requires intensive forward modeling of the wavefield to estimate the gradient and step size, resulting in large-scale computation, which leads to non-ideal subsurface structure for the conventional FWI inversion.

And step S104, carrying out parameterization processing on the initial physical parameters to obtain the physical parameters.

In practical applications, DNN is widely used for machine learning tasks including signal processing, image recognition and machine translation, and its basic principle is the universal approximation theorem. To simplify the network architecture and simplify the training process, most DNNs consist of piecewise linear layers, and because piecewise linear polynomials are dense in a function space composed of all balier measurable functions, piecewise polynomial approximation properties have wide application in mathematics, such as finite element methods, which employ piecewise polynomial approximation functions on each grid cell. From the perspective of big data science, the effectiveness of the universal approximation theorem is verified by concentrating high-dimensional data near a low-dimensional manifold.

Wherein for any Boyle measurable function g (x), there is a piecewise linear polynomial p (x) that is approximated to g (x) with any desired precision according to the following equation:

wherein g (x) represents the Boyle's measurable function, p (x) represents a piecewise linear polynomial, φ (x) approximating g (x)_n(x) Representing piecewise linear sub-polynomials, e_nRepresenting corresponding piecewise linear sub-polynomials phi_n(x) N is a positive integer, N is 0,1, 2.

At this time, there is a neural network h (w) satisfying the following equation:

p(x)＝H(w)(x):＝H_L(w_L-1,L,H₁(w₁))(x) (4)

wherein H₁And H_LRespectively representing the first and Lth layers of the neural network, L representing the depth of the deep neural network, w₁And w_L-1Weight coefficients representing the first and L-1 layers of the deep neural network, and x represents an argument, i.e., an input quantity of the deep neural network.

At this time, an activation function σ can be defined according to the following formula_kSimple layer (c):

H_k(W_k)(x)＝σ_k(W_kx+b_k) (5)

wherein H_kLayer k, W, representing a deep neural network_kRepresenting an affine mapping, b_kRepresenting the offset term, x representing the argument, i.e. the input to the deep neural network, σ_kRepresenting an activation function. Here, W is_kAnd b_kConstituting a learnable parameter, randomly initialized in the k-th layer of the deep neural network, and a nonlinear activation function sigma_kAre continuous and component-wise operations, such as the rectifying Linear units ReLU max (0, x) and its variants, leakage ReLU (Linear rectifying function) and prilu (Parametric rectifying Linear function).

The above mathematical approximation theorem provides only the presence of piecewise polynomials for a given function, but does not provide how to construct piecewise polynomials. The DNN framework may provide a systematic and efficient way to construct optimal piecewise polynomials to adapt to solve the ill-posed problem.

On the basis of the mathematical approximation theorem, a parameterized polynomial g (w) may also be learned for the initial physical parameter m of the FWI problem, where the parameterized polynomial g (w) approximates the initial physical parameter m with a given precision, and specifically, the initial physical parameter m may be parameterized according to the following formula to obtain the physical parameter:

m(w)＝m＝G(w) (6)

Step S106, inputting the physical parameters into a machine learning model trained in advance, so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; and the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relationship.

Specifically, a deep neural network can be obtained by learning according to the parameterized polynomial g (w), and a machine learning model can be obtained by training the deep neural network, wherein the machine learning model can output a full-waveform inversion result according to the input physical parameters m (w), and the full-waveform inversion result and the weight coefficient of the machine learning model satisfy a preset relationship, that is, the full-waveform inversion result is a weight coefficient related to the physical parameters, so as to obtain the physical parameters by calculation according to the full-waveform inversion result. At the moment, physical parameters are input into a machine learning model trained in advance, so that the machine learning model outputs a full-waveform inversion result according to the physical parameters, namely the DNN-FWI method, the machine learning model is obtained by training in a DNN framework, the machine learning model can parameterize initial physical parameters through weight coefficients, and therefore a traditional full-waveform inversion problem is converted into a reconstruction problem of machine learning model network parameters.

At this time, for the objective function of the existing FWI, it can be converted into an objective function regarding the weight coefficient of the machine learning model according to the following formula:

wherein w represents a weight coefficient, R represents a projection operator, d_obsFor observation data, u is a wave field propagated by seismic waves, m (w) represents physical parameters, and G (w) represents a parameterized polynomial, namely a deep neural network obtained by learning according to the parameterized polynomial.

Therefore, the above equation (7) transforms the inverse problem of the conventional FWI into the reconstruction problem of the weight coefficients w of the deep neural network, thereby providing a sparse representation of the neural network by introducing a special network layer (e.g., convolutional layer), and alleviating the problem of the local extremum of the conventional FWI in a regularized manner. In addition, by using a piecewise polynomial to represent initial physical parameters as prior information, certain specific features can be extracted by constructing a special layer of a deep neural network, namely, full waveform inversion under a DNN framework can be regarded as an implicit regularization method, which is important for solving the problem of unsuitability of FWI; and, the training process of DNN can be accelerated by sophisticated and efficient neural network libraries, such as TensorFlow (https:// tensegrow. google. cn) or PyTorch (https:// PyTorch. org), using a GPU (Graphics Processing Unit), and thus, the full waveform inversion method is called DNN-FWI.

At this time, in order to minimize the objective function in the formula (7) using the gradient type method, the gradient of the objective function with respect to the weight coefficient w needs to be calculated. The objective function of the above equation (7) is differentiated and the chain rule is applied according to the following equation:

wherein w represents a weight coefficient, m (w) represents a physical parameter, G (w) represents a parameterized polynomial, i.e., a deep neural network learned from the parameterized polynomial, m represents an initial physical parameter,

a Jacobian matrix representing a deep neural network can be effectively estimated using a back propagation algorithm; factor(s)

Representing the gradient of the objective function of a conventional FWI with respect to the initial physical parameter m, this factor can also be efficiently calculated by the adjoint method.

Furthermore, before carrying out the DNN-FWI training, the initial value m of a given physical parameter must be parameterized₀To enable the deep neural network to learn the a priori information, this parameterization process is called "pre-training". In order to parameterize or learn the initial model, the minimum needs to be solved according to the following formula:

wherein w represents a weight coefficient, J_pre(w) represents a pre-trained objective function, G (w) represents a parameterized polynomial, i.e., a deep neural network learned from the parameterized polynomial, m₀Indicating an initial value, w, of a physical parameter^*Represents the minimum value of pre-training, | | | | non-conducting phosphor₁Is represented by₁And (4) norm.

The above formula (9) measures G (w) and m₀And, adopt l₁The norm can capture the main features of the initial model, and use l₂Norm is stronger than outlier value.

After the "pre-training" is completed, the DNN-FWI training process is started. Wherein the main difference between the conventional training and the re-parameterized FWI training is the gradient calculation with respect to the weight m. For conventional training, only the differences between the network output and the given label data are propagated back from the output layer to the input layer to compute the gradient; while for DNN-FWI, a backward propagation gradient is used

To calculate the gradient with respect to the weight coefficient w. Therefore, DNN-FWI ensures data consistency by synthesizing recorded data, using a PDE (Partial Differential Equation) constrained forward model, or by pairing CPU (Central Processing Unit) and GPU (GPU) respectively

And the partial derivatives of the deep neural network g (w) proceed and accelerate. In addition, algorithm 1 in table 1 details the framework of the re-parameterized FWI algorithm.

TABLE 1

It should be noted that the deep neural network g (w) may also use a generation part of GAN, where GAN is a special deep network architecture, and is composed of a generator network and a discriminator network, and GAN has achieved remarkable success in image restoration, super resolution, 3D vision, face editing, and FWI.

When the DNN-FWI training is finished, the deep neural network is a machine learning model, and when the machine learning model inputs physical parameters m (w), the machine learning model performs inversion processing on the physical parameters m (w) to output a full waveform inversion result, and the gradient only needs first-order derivative information, so that the machine learning model has higher computational efficiency compared with the existing FWI.

And S108, reconstructing a speed structure chart of the underground structure according to the full waveform inversion result.

At this time, according to the full waveform inversion result output by the machine learning model, the speed structure diagram of the underground structure can be reconstructed through a plurality of GPUs and CPUs arranged in the server, and the reconstructed speed structure diagram has high resolution, so that the precision of inverting the complex underground structure is improved.

According to the full-waveform inversion method provided by the embodiment of the invention, initial physical parameters are firstly obtained; wherein the initial physical parameter is a velocity parameter of the underground structure; then, carrying out parameterization processing on the initial physical parameters to obtain physical parameters, and inputting the physical parameters into a machine learning model trained in advance so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relation; and finally, reconstructing a speed structure chart of the underground structure according to the full waveform inversion result. According to the method and the device, the initial physical parameters are parameterized through the weight coefficients of the machine learning model, so that the full waveform inversion problem is converted into the reconstruction problem of the network parameters of the machine learning model, the problems that the existing full waveform inversion cannot solve the nonlinear inverse problem and a large amount of calculation exists are solved, the calculation efficiency is improved, and the beneficial effect of inverting the complex underground structure with high precision is achieved.

On the basis of fig. 1, another full-waveform inversion method is also provided in the embodiments of the present invention, which focuses on the process of reconstructing a velocity structure diagram of a subsurface structure from full-waveform inversion results. As shown in fig. 2, the method comprises the steps of:

step S202, acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the subsurface structure.

And step S204, carrying out parameterization processing on the initial physical parameters to obtain the physical parameters.

Step S206, inputting the physical parameters into a machine learning model trained in advance, so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; and the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relationship.

The above steps S202 to S206 can refer to the above steps S102 to S106, and the embodiments of the present invention are not described in detail herein.

And step S208, carrying out forward modeling on the full waveform inversion result to obtain a full waveform inversion matrix.

Specifically, the initial physical parameters are velocity parameters of each underground medium in the underground structure, so that the machine learning model outputs full waveform inversion results according to the physical parameters, the full waveform inversion results are weighted values meeting preset relations with the physical parameters, the physical parameters are obtained through inverse calculation according to the full waveform inversion results, at the moment, the full waveform inversion results are subjected to wave field forward simulation, the physical parameters obtained through the inverse calculation are subjected to wave field forward simulation, inversion matrix values of a plurality of physical parameters are obtained, and a full waveform inversion matrix is obtained according to the inversion matrix values of the physical parameters.

And S210, reconstructing to obtain a velocity structure diagram of the underground structure based on the full waveform inversion matrix.

And based on the full-waveform inversion matrix, the execution main body server is also provided with a plurality of GPUs and CPUs, and at the moment, the GPUs and the CPUs are reconstructed according to the full-waveform inversion matrix to obtain a speed structure diagram of the underground structure. The GPU may be drawing software, for example: matlab, Wigb, etc. are displayed through a display module (such as a display screen) connected with the CPU, so that a user can simply and visually observe the speed structure diagram of the complex underground structure, and the user can conveniently master the complex underground structure.

In one possible implementation, the full waveform inversion result may also be a feature inversion result, where the feature inversion result is a feature weight value that satisfies a preset relationship with the feature physical parameter, and at this time, the feature inversion result is subjected to a forward modeling of a wavefield to obtain a feature inversion matrix; and reconstructing to obtain a characteristic diagram of the underground structure based on the characteristic inversion. The calculation process of the characteristic inversion matrix is the same as that of the full waveform inversion matrix solving method, and the reconstruction method of the characteristic diagram is the same as that of the velocity structure diagram, so that the method is used for simply and visually mastering the specific situation of the complex underground structure according to the characteristic diagram, and has important significance for geophysical inversion research.

On the basis of the above embodiment, the embodiment of the invention also provides a training method of the machine learning model, which mainly introduces the process of obtaining the machine learning model by training based on the deep neural network. As shown in fig. 3, the method comprises the steps of:

step S302, obtaining an original training physical parameter set; wherein the original training set of physical parameters comprises velocity parameters of a plurality of subsurface media in the subsurface structure;

step S304, inputting the original training physical parameter set to a deep neural network for training to obtain a machine learning model.

Specifically, the original training physical parameter set is input to a deep neural network to obtain a full waveform inversion result of the original training physical parameter set; calculating a function value of an objective function of the deep neural network based on a full waveform inversion result of the original training physical parameter set, and specifically referring to a formula (7); the method comprises the steps of adjusting parameters of a deep neural network through a function value of an objective function to obtain a machine learning model, outputting a full waveform inversion result which meets a preset relation (an approximation relation) with the physical parameters according to the input physical parameters by the machine learning model, namely outputting a weighted value corresponding to the physical parameters, so that a velocity structure diagram of the underground structure can be reconstructed by a GPU and a CPU according to the full waveform inversion result, converting the traditional full waveform inversion problem into the reconstruction problem of network parameters of the machine learning model, solving the problems that the traditional full waveform inversion method needs intensive wave field simulation to estimate a large amount of calculated quantity caused by gradient and step length and is easy to fall into multiple local extrema, improving the calculation efficiency and the accuracy of underground structure inversion.

The DNN-FWI based on the machine learning model described above is also effective for the inverse problem of acoustic waves, elasticity, and viscoelastic media in the time domain and frequency domain. For ease of understanding, the sound waves are described here as an example. For example, for the FWI problem of the two-dimensional frequency domain acoustic wave equation, assuming that the density is constant, the pressure wave velocity is input into a trained machine learning model, at which time the acoustic wave equation can be described using Helmholtz's equation according to the following formula:

where ω denotes the angular frequency, v_pRepresenting acoustic velocity, s representing seismic sourceAnd the function u represents a seismic wave field propagated by the seismic waves, namely u is (x, y, z), x, y and z are spatial coordinate points, and delta u represents the disturbance of the seismic wave field to a background field.

Wherein m ═ v_pAnd A (m) u: ═ Δ u + (ω/v)_p)²u, then the above equation (10) can be rewritten as the following equation:

A(m)u＝s (11)

for equation (11) above, to numerically solve the positive problem, equation (11) is discretized here with a finite difference template of fourth order precision, and a perfect matching layer is employed to reduce the artificial reflections from the bounded computational domain boundaries. After discretization, a large sparse linear system can be obtained, the solution of which is a numerical solution of formula (11), or the large linear system can be effectively solved by using a HSS (hierarchical Semi-separable) structure direct solving algorithm. Compared with the standard direct solving method, the solving method of the HSS structure realizes high performance and good scalability in the aspects of computing cost and storage. The solving method is realized by adopting a distributed memory MPI and a shared memory OpenMP parallel programming framework. The DNN-FWI method and the network update are implemented using the high efficiency C language under the MPI framework and the Python language under the PyTorch platform framework, respectively.

The universal approximation theorem ensures that a polynomial approximating the function with any precision exists for each continuous function, and a DNN can be obtained by learning according to the polynomial. However, from a deep learning geometry perspective, for any DNN with a fixed architecture, there is a manifold that cannot be learned over the network. Therefore, for a particular application problem, neural networks face challenges, namely how many cells and how many layers should be employed by the deep network and how these cells should be connected to each other. Although deep networks may face vanishing and explosive gradient computation problems, deeper models may greatly reduce the number of cells required and the generalization error, which is a rule of thumb in deep neural network design. We have also observed that a fixed network may well represent multiple networks with different physical models, although it may not be the best network for these models.

The application provides a deep learning model for numerical experiments, and the network mainly comprises 32 layers of convolution (Conv) and deconvolution (Deconv) operations, and takes a hyperbolic tangent (Tanh) function as an output layer. The Conv operation is used to capture abstract features at a higher level, while the Deconv layer doubles the resolution of the Conv layer of the previous layer. Each hidden layer adds a nonlinear activation ReLU to increase the nonlinearity of the network, so that the network has the capability of approximating a nonlinear function. A batch normalization (BatchNorm) layer was introduced for each Conv and DeConv layer to remove the mean and normalize the variance. The BatchNorm layer can reduce the dependence of the gradient on the proportion of the parameters in each hidden layer or their initial values, thereby significantly speeding up the learning process. A specific network as shown in fig. 4, in which a super layer CellLayer composed of one DeConv layer and four Conv layers is introduced for simplification, and parameters I and O of the CellLayer represent the number of input and output channels, respectively, where a represents an input tensor of a machine learning model, FWI-Net, and m1 represents an output of the machine learning model, FWI-Net, i.e., a full waveform inversion result.

In another possible embodiment, the Marmousi2 model is used as an example for illustration. Marmousi2 is seismic data created from complex geological structures that require advanced processing techniques to correctly invert it, and thus the Marmousi2 model is one of the reference models for the test seismic inversion method. The original Marmousi2 model was created based on a real stratigraphic structure with a width of 16km and a depth of 3.5km with a 400m water layer on top of the model. A real model of the Marmousi2 model, shown in fig. 5, with a size of 384 × 128, with a spatial step size of 20m, was used to test the performance of the DNN-FWI method. A smoothed model of the Marmousi2 model shown in fig. 6, which is obtained by smoothing the true model using a gaussian filter.

In the inversion process described above, with the addition of 10 grid PML boundary conditions on each side of the rectangular domain, the surface acquisition system consists of 116 explosive sources evenly distributed at 60m intervals on top of the model (40m depth), each source recorded by 350 receivers 40m below the top, where the receivers are evenly distributed at 20m intervals.

At this point, the key to performing the inversion under the DNN framework is pre-training. In the pre-training stage, the above-mentioned smooth model is used as the learning model of the deep neural network, in which the input tensor a of the deep neural network is fixed to [ -0.5, 1, 0.5] in fig. 4, it should be noted that the input tensor a may be any actual tensor whose components are located at [ -1, 1], but it should be ensured that the pre-training and the inversion stages must be the same. The pre-training is carried out on a GPU, an ADAM (Adaptive motion Estimation) optimization method with the learning rate of 0.001 is used for accelerating the convergence of the pre-training, and after 30000 iterations, an output model of a pre-training network is an accurate approximate value of a smooth model; and calculating the travel time through the smoothed model and the output model, the travel time having an RMSE (Root Mean Squared Error) of about 0.94 ms. Further, since the output of the Tanh layer is [ -1, 1], the smoothing model should be converted so that the value of each element is between [ -1, 1 ].

In addition, a composite data set was calculated using the Dirac source wavelet function and twenty frequencies. The discrete frequency ranges from 3.0 to 12Hz, the constant sampling interval is 0.45Hz, the DNN-FWI method is carried out from low frequency to high frequency, namely, data is inverted one frequency at a time, after 50 iterations, the inversion is moved to the next frequency, and the process is carried out until the last frequency. This inversion strategy can mitigate the non-linearity and ill-qualification of the inversion problem and has been widely used in the conventional FWI framework. When the learning rate of this inversion stage is 0.001. After 1000 iterations, the final output model of the network is shown in fig. 7, the vertical section at the level of 3.0km is shown in fig. 8-a, the vertical section at 5.0km is shown in fig. 8-b, and the vertical section at 6.0km is shown in fig. 8-c, where curve 1 represents the vertical section corresponding to the true model of TM, i.e., the Marmousi2 model, curve 2 represents the vertical section corresponding to the smooth model of SM, i.e., the Marmousi2 model, curve 3 represents the inversion result of DM, i.e., DNN-FWI, and curve 4 represents the inversion result corresponding to FM, i.e., stagei.

In order to further improve the resolution of the reconstructed model, a new inversion is performed under the conventional FWI framework, i.e. the reconstructed model of DNN-FWI is used as the initial model, and the process is called Stage II. To speed up this Stage, an efficient truncated confidence domain method is used, unlike the DNN-FWI inversion, in Stage II only 3.0, 4.35, 5.7, 7.05, 8.4, 9.75 and 12Hz frequencies need to be selected, and the maximum number of iterations for each frequency is 20, as shown in the final reconstruction result in fig. 9.

In conclusion, the machine learning model, namely the DNN-FWI method provided by the present application can successfully and accurately reconstruct the complex Marmousi2 model, and the DNN-FWI method provided by the present application is effective and has high resolution because the vertical wheel tangents are well matched. However, as shown in fig. 8-c, Stage II improvement is limited and even decreases with increasing depth, such as the elliptical labeled portion, which may be attributed to multiple local extremes of the objective function and insufficient illumination.

In another possible embodiment, a mineral model is taken as an example for illustration. To study the DNN-FWI performance in feature extraction and high resolution imaging, a model was constructed to simulate the subsurface mineral distribution with a velocity distribution between 5000m/s and 7500m/s, where different values represent different minerals. As shown in fig. 10, the size of the real model used in the test is 512 × 256, and the space division distance is 10m, and the micro-scale structure of the real model provides a challenge to the inversion method. Smoothing the real model by using a gaussian filter to obtain an initial model for pre-training, i.e. a smoothed model, as shown in fig. 11. The surface acquisition system consists of 119 seismic sources, evenly distributed 50m below the top surface, at intervals of 40m, each seismic source being recorded by 472 receivers located 50m below the top surface, the receivers being evenly distributed at intervals of 10 m.

For pre-training, the input tensor a is [ -0.5, 0.5, -0.5, 0.5; -0.5, -0.5, 0.5 ]. It should be noted that the elements of the input tensor a here can be any real number as long as they are in the interval [ -1, 1], and are consistent during the pre-training and DNN-FWI phases. Where the maximum iteration of the pre-training is 30000, the learning rate is 0.001, and the RMSE of the travel time of the pre-trained model is about 0.14ms, indicating that the network can approximate the initial model well.

For the inversion stage of DNN-FWI, the synthetic data set was calculated using 50 discretized frequencies, distributed over the [3.0Hz, 45.0Hz ] interval, and sampled at a constant sampling step of 0.86 Hz. At this stage, the training of DNN-FWI employs a multi-scale inversion strategy from low frequency (3.0Hz) to high frequency (45.0Hz), with 30 iterations per frequency. To further improve the resolution of the inversion, a conventional FWI inversion was performed using frequencies 3.0, 11.57, 20.14, 28.71, 37.28, and 45.0Hz, with 15 iterations per frequency, as shown in figure 12 for the reconstructed results of the DNN-FWI method after 1500 iterations of the second stage, and, the reconstructed result after 90 iterations of the conventional FWI method as shown in fig. 13, for vertical profiles at different locations (e.g. 1.2km, 2.8km, 3.8km), then the vertical profile at 1.2km is shown in figure 14-a, FIG. 14-b shows a vertical section at 2.8km, FIG. 14-c shows a vertical section at 3.8km, curve 1 represents TM, i.e., a vertical section corresponding to a true model of the mineral model, curve 2 represents SM, i.e., a vertical section corresponding to a smoothed model of the mineral model, curve 3 represents DM, i.e., DNN — FWI, and curve 4 represents FM, i.e., stagei, inversion result. The above-mentioned fig. 12, fig. 13, fig. 14-a, fig. 14-b and fig. 14-c show that the DNN-FWI method provided by the present application can accurately reconstruct a high contrast model of a mineral model, and that DNN-FWI has a regularization property of preserving sharp structures.

In another possible embodiment, the conventional FWI method based on the L-BFGS algorithm is exemplified here in order to prove the advantage of the DNN-FWI method over the conventional FWI method. To preserve the boundary information and mitigate the effects of multiple local minima of the FWI problem, TV (Total Variation) regularization is applied to a conventional FWI method, where the objective function of TV regularization can be calculated according to the following formula:

f_reg(m)＝f(m)+λ||m||_TV (12)

where λ is the regularization parameter, i.e. for control data simulationSum term f (m) and TV canonical term | m | | ceiling_TVA regularization parameter of the trade-off between | | | | | non-woven phosphor_TVRepresenting the TV regularization term.

In practical applications, the TV regularization term can be calculated according to the following formula:

wherein | m | Y calculation_TVA term representing the regularization of the TV,

a function value representing the physical parameter m at the (i, j) th point.

Wherein it can be calculated according to the following formula

Wherein the content of the first and second substances,

a function value representing a physical parameter m at the (i, j) th point,

represents the value of the physical parameter m in the x direction at the (i, j) th point,

denotes the value of a physical parameter m at the (i, j) th point in the z direction, m_i+1,jDenotes the corresponding parameter value, m, at the (i +1, j) th point_i,j+1Represents the corresponding parameter value, m, at the (i, j +1) th point_i,jDenotes the corresponding parameter value at the (i, j) th point, x denotes the x direction, z denotes the z direction, which is perpendicular to the subsurface direction, i and j denote the abscissa and ordinate of the discrete point, Δ z denotes the depth direction discrete step size, and β is a positive real number.

It should be noted that the factor Δ x Δ z is omitted to the right in the discretization equation (14) for the TV regularization term described above, since it can be absorbed by the regularization parameter λ. The small positive real number beta is introduced to overcome the irreducibility of the TV regularization term at the origin, and in practical application, the positive real number beta can be set to 1.0e^-4At this time, the TV regularization term can be related to the parameter m as follows_i,jThe derivative of (a) is approximated as:

wherein the content of the first and second substances,

a function value representing a physical parameter m at the (i, j) th point,

a function value representing the physical parameter m at the (i-1, j) th point,

a function value representing the physical parameter m at the (i, j-1) th point,

represents the value of the physical parameter m in the z direction at the (i, j) th point,

represents the value of the physical parameter m in the z direction at the (i, j-1) th point,

denotes the value of the physical parameter m at the (i-1, j) th point in the x direction, x denotes the x direction, z denotes the z direction, which is perpendicular to the subsurface direction, i and j denote the abscissa and ordinate of the discrete pointThe coordinate, Δ z, represents a discrete step in the depth direction.

Optionally, there is another valid option for the TV regularization term, that is, the TV term is used as a constraint, and the detailed description of the embodiment of the present invention is omitted here.

In addition, PSNR (Peak Signal to Noise Ratio) index is used to quantify the inversion results of the different methods. It is defined as:

wherein m is^tRepresenting a real model, m^rRepresenting an inverse model, the MSE (m) can be calculated according to the following equation^t,m^r)：

Wherein m is^tRepresenting a real model, m^rRepresenting an inverse model, N representing the number of discrete points in the x-direction, M representing the number of discrete points in the depth z-direction,

represents the value of the physical parameter m in the real model at the (i, j) th point,

the values of the physical parameter m in the inverse model at the (i, j) th point are shown, and i and j represent the abscissa and ordinate of the discrete point. N and M can be set according to practical application conditions, and are not limited to be described in the embodiment of the invention.

And calculating MAX (m) according to the following formula^t)：

Wherein m is^tA real model is represented by a model of the object,

the values of the physical parameter m in the real model at the (i, j) th point are shown, and i and j represent the abscissa and ordinate of the discrete point.

In practice, the mineral model is reconstructed according to DNN-FWI and conventional FWI methods. Among them, the initial model, which is a smooth model shown in fig. 15, has a good smoothness, and when a real model is calculated using a Ricker wavelet, the seismic record is shown in fig. 16, and when the initial model is calculated using a Ricker wavelet, the seismic record is shown in fig. 17, wherein the principal frequency of Ricker is 25Hz, and z is 10 m. From fig. 16 and 17, it can be seen that without low frequency information or a suitable regularization strategy, it would be difficult to reconstruct a mineral model from this initial model.

For the DNN-FWI method, the maximum number of pre-trained iterations was also 30000, and both the DNN-FWI and conventional FWI methods used all of the configurations in the above mineral model experiments. As shown in fig. 18, is the reconstructed result after 1500 iterations of DNN-FWI, and, as shown in fig. 19, is the reconstructed result after 1500 iterations of conventional FWI. Wherein the regularization parameter is 2.0e^-3The calculated performance of each method is shown in FIG. 20, with a total calculated time of 48439s for DNN-FWI and 78124s for L-BFGS.

Further, as shown in fig. 21-a, the vertical section at x of 1.2km, fig. 21-b, and fig. 21-c, the vertical section at x of 2.8km, and fig. 21-c, the vertical section at x of 3.8km, where curve 1 represents the vertical section corresponding to the TM, curve 2 represents the vertical section corresponding to the SM, the smooth model, curve 3 represents the inversion result of the DM, DNN — FWI, and curve 4 represents the inversion result of the FM, stagei. It can be seen that the DNN-FWI process provided herein has superior performance to conventional FWI processes. Wherein, for the DNN-FWI method, the PSNR index is 31.20, and the PSNR index of the L-BFGS is 30.19, wherein the better result of the DNN-FWI is probably due to the regularization and feature extraction characteristics of the DNN convolutional layer; in addition, the DNN-FWI method shows good computational efficiency performance due to its implicit regularization, feature extraction and hierarchical structure of the algorithm, while the L-BFGS method requires more form time, resulting in more function values and gradient estimation due to the multiple line search steps required for some non-linear iterations. And 6408 matrix decompositions are performed on the L-BFGS, including the positive problem and the accompanying problem, and the DNN-FWI is performed only 3000 times, so that the DNN-FWI provided by the application has higher calculation efficiency compared with the traditional FWI.

In summary, the present application provides a full waveform inversion method that solves the FWI problem under the DNN framework, and thus DNN-FWI can be considered as a reparameterization method. Compared to the conventional FWI process, the DNN-FWI process has several advantages: (1) the model representation with the piecewise polynomial may be used as a priori information to guide the inversion method to select a solution in the desired function space, so DNN-FWI may be suitable for automated regularization methods that solve ill-defined problems; (2) DNN-FWI involves only the first derivative and therefore has good computational efficiency. Furthermore, DNN-FWI can be easily implemented in an efficient manner also under the framework of deep learning toolkits such as PyTorch or TensorFlow.

In addition, in order to evaluate the effectiveness of DNN-FWI, a deep learning model, namely a machine learning model, comprising 32 layers is established. The network of the machine learning model mainly comprises Conv and Deconv operations, numerical experiments based on Marmousi2 and a mineral model are carried out on the basis of the machine learning model to prove the feasibility and the efficiency of the machine learning model, and the numerical results show that the proposed DNN-FWI is effective and can invert complex underground structures with relatively high precision. Therefore, the DNN-FWI provided by the application solves the problems that the existing full-waveform inversion cannot solve the nonlinear inverse problem and a large amount of calculation exists, improves the calculation efficiency and achieves the beneficial effect of inverting the complex underground structure with high precision.

Example two:

corresponding to the full waveform inversion method shown in fig. 1, an embodiment of the present invention further provides a full waveform inversion apparatus, which is applied in a server. As shown in fig. 22, the apparatus includes an obtaining module 221, a parameterization module 222, an input-output module 223 and a reconstruction module 224, which are connected in sequence, wherein the functions of the modules are as follows:

an obtaining module 221, configured to obtain an initial physical parameter; wherein the initial physical parameter is a velocity parameter of the underground structure;

a parameterization module 222, configured to perform parameterization on the initial physical parameter to obtain a physical parameter;

an input/output module 223, configured to input the physical parameters into a machine learning model trained in advance, so that the machine learning model outputs a full waveform inversion result according to the physical parameters; the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relation;

and the reconstruction module 224 is used for reconstructing the velocity structure chart of the underground structure according to the full waveform inversion result.

According to the full waveform inversion device provided by the embodiment of the invention, initial physical parameters are firstly obtained; wherein the initial physical parameter is a velocity parameter of the underground structure; carrying out parameterization processing on the initial physical parameters to obtain physical parameters, and inputting the physical parameters into a machine learning model trained in advance so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; the full waveform inversion result and the weight coefficient of the machine learning model meet a preset relation; and reconstructing a velocity structure map of the subsurface structure from the full waveform inversion results. According to the method and the device, the initial physical parameters are parameterized through the weight coefficients of the machine learning model, so that the full waveform inversion problem is converted into the reconstruction problem of the network parameters of the machine learning model, the problems that the existing full waveform inversion cannot solve the nonlinear inverse problem and a large amount of calculation exists are solved, the calculation efficiency is improved, and the beneficial effect of inverting the complex underground structure with high precision is achieved.

In one possible implementation, the reconstructing module 224 is further configured to: forward modeling is carried out on the full waveform inversion result to obtain a full waveform inversion matrix; and reconstructing to obtain a speed structure chart of the underground structure based on the full waveform inversion matrix.

In another possible embodiment, the full waveform inversion result is a feature inversion result, and the reconstruction module 224 is further configured to: forward modeling is carried out on the feature inversion result to obtain a feature inversion matrix; and reconstructing to obtain a characteristic diagram of the underground structure based on the characteristic inversion.

In another possible embodiment, the parameterization module 222 is further configured to: carrying out parameterization processing on the initial physical parameters according to the following formula to obtain the physical parameters:

m(w)＝m＝G(w)

In another possible embodiment, the machine learning model is a model trained based on a deep neural network, and the apparatus further includes: acquiring an original training physical parameter set; wherein the original training set of physical parameters comprises velocity parameters of a plurality of subsurface media in the subsurface structure; and inputting the original training physical parameter set into a deep neural network for training to obtain a machine learning model.

In another possible embodiment, the inputting the original training physical parameter set to the deep neural network for training to obtain the machine learning model includes: inputting the original training physical parameter set into a deep neural network to obtain a full waveform inversion result of the original training physical parameter set; calculating a function value of an objective function of the deep neural network based on a full waveform inversion result of the original training physical parameter set; and adjusting the parameters of the deep neural network through the function value of the objective function to obtain a machine learning model.

The implementation principle and the technical effect of the full waveform inversion apparatus provided by the embodiment of the present invention are the same as those of the full waveform inversion method embodiment, and for brief description, reference may be made to the corresponding contents in the full waveform inversion method embodiment for some points not mentioned in the embodiment of the full waveform inversion apparatus.

An embodiment of the present invention further provides an electronic device, as shown in fig. 23, which is a schematic structural diagram of the electronic device, where the electronic device includes a processor 231 and a memory 232, the memory 232 stores machine executable instructions capable of being executed by the processor 231, and the processor 231 executes the machine executable instructions to implement the full waveform inversion method.

In the embodiment shown in fig. 23, the electronic device further comprises a bus 233 and a communication interface 234, wherein the processor 231, the communication interface 234 and the memory 232 are connected by the bus.

The Memory 232 may include a high-speed Random Access Memory (RAM) and may also include a non-volatile Memory (non-volatile Memory), such as at least one disk Memory. The communication connection between the network element of the system and at least one other network element is realized through at least one communication interface 234 (which may be wired or wireless), and the internet, a wide area network, a local network, a metropolitan area network, and the like can be used. The bus may be an ISA bus, PCI bus, EISA bus, or the like. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one double-headed arrow is shown in FIG. 23, but that does not indicate only one bus or one type of bus.

The processor 231 may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method may be performed by integrated logic circuits of hardware or instructions in the form of software in the processor 231. The Processor 231 may be a general-purpose Processor, and includes a Central Processing Unit (CPU), a Network Processor (NP), and the like; the device can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, a discrete Gate or transistor logic device, or a discrete hardware component. The various methods, steps and logic blocks disclosed in the embodiments of the present invention may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present invention may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software module may be located in ram, flash memory, rom, prom, or eprom, registers, etc. storage media as is well known in the art. The storage medium is located in a memory, and the processor 231 reads the information in the memory 232 and completes the steps of the full waveform inversion method of the foregoing embodiment in combination with the hardware thereof.

An embodiment of the present invention further provides a computer-readable storage medium, where a computer-executable program is stored on the computer-readable storage medium, and when the computer-executable program is called and executed by a processor, the computer-executable program causes the processor to implement the full waveform inversion method, and specific implementation may refer to the foregoing method embodiment, and is not described herein again.

The full waveform inversion method, the full waveform inversion apparatus, and the computer program product of the electronic system provided in the embodiments of the present invention include a computer-readable storage medium storing program codes, where instructions included in the program codes may be used to execute the full waveform inversion method described in the foregoing method embodiments, and specific implementations may be referred to as method embodiments and will not be described herein again.

The computer program product provided in the embodiment of the present invention includes a computer-readable storage medium storing a program code, where instructions included in the program code may be used to execute the method described in the foregoing method embodiment, and specific implementation may refer to the method embodiment, which is not described herein again.

It is clear to those skilled in the art that, for convenience and brevity of description, the specific working process of the apparatus described above may refer to the corresponding process in the foregoing method embodiment, and is not described herein again.

In addition, in the description of the embodiments of the present invention, unless otherwise explicitly specified or limited, the terms "mounted," "connected," and "connected" are to be construed broadly, e.g., as meaning either a fixed connection, a removable connection, or an integral connection; can be mechanically or electrically connected; they may be connected directly or indirectly through intervening media, or they may be interconnected between two elements. The specific meanings of the above terms in the present invention can be understood in specific cases to those skilled in the art.

The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a non-volatile computer-readable storage medium executable by a processor. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.

In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, and are only for convenience of description and simplicity of description, but do not indicate or imply that the device or element being referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus, should not be construed as limiting the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.

Finally, it should be noted that: the above-mentioned embodiments are only specific embodiments of the present invention, which are used for illustrating the technical solutions of the present invention and not for limiting the same, and the protection scope of the present invention is not limited thereto, although the present invention is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: any person skilled in the art can modify or easily conceive the technical solutions described in the foregoing embodiments or equivalent substitutes for some technical features within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the embodiments of the present invention, and they should be construed as being included therein. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A full waveform inversion method, applied to a server, the method comprising:

acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the subsurface structure;

inputting the physical parameters into a pre-trained machine learning model so that the machine learning model outputs a full waveform inversion result according to the physical parameters; wherein the full waveform inversion result and the weight coefficient of the machine learning model satisfy a preset relationship, that is, the full waveform inversion result is the weight coefficient related to the physical parameter;

reconstructing a speed structure chart of the underground structure according to the full waveform inversion result;

the step of reconstructing the velocity structure diagram of the underground structure according to the full waveform inversion result comprises the following steps:

carrying out forward modeling on the full waveform inversion result to obtain a full waveform inversion matrix; performing inverse calculation on the full waveform inversion result, performing forward modeling on each physical parameter obtained by the inverse calculation to obtain a plurality of inversion matrix values of the physical parameters, and obtaining a full waveform inversion matrix according to the plurality of inversion matrix values of the physical parameters;

and reconstructing to obtain a speed structure chart of the underground structure based on the full-waveform inversion matrix.

2. The full waveform inversion method of claim 1, wherein the full waveform inversion result is a characteristic inversion result, and the step of reconstructing the velocity map of the subsurface structure from the full waveform inversion result comprises:

carrying out forward modeling on the feature inversion result to obtain a feature inversion matrix;

3. The full waveform inversion method of claim 1, wherein the step of parameterizing the initial physical parameters to obtain physical parameters comprises:

carrying out parameterization processing on the initial physical parameters according to the following formula to obtain the physical parameters:

m(w)＝m＝G(w)

wherein m represents the initial physical parameter, g (w) represents a parameterized polynomial, w represents a weight coefficient, and m (w) represents the physical parameter.

4. The full waveform inversion method of claim 1, wherein the machine learning model is a model trained based on a deep neural network, the method further comprising:

and inputting the original training physical parameter set to the deep neural network for training to obtain the machine learning model.

5. The full waveform inversion method of claim 4, wherein the step of inputting the original set of trained physical parameters to the deep neural network for training to obtain the machine-learned model comprises:

inputting the original training physical parameter set into the deep neural network to obtain a full waveform inversion result of the original training physical parameter set;

and adjusting the parameters of the deep neural network through the function value of the objective function to obtain the machine learning model.

6. A full waveform inversion apparatus, applied to a server, the apparatus comprising:

the acquisition module is used for acquiring initial physical parameters; wherein the initial physical parameter is a velocity parameter of the subsurface structure;

the input and output module is used for inputting the physical parameters into a machine learning model trained in advance so that the machine learning model outputs a full-waveform inversion result according to the physical parameters; wherein the full waveform inversion result and the weight coefficient of the machine learning model satisfy a preset relationship, that is, the full waveform inversion result is the weight coefficient related to the physical parameter;

the reconstruction module is used for reconstructing a velocity structure chart of the underground structure according to the full waveform inversion result;

wherein the reconstruction module further comprises:

7. The full waveform inversion apparatus of claim 6, wherein the machine learning model is a model trained based on a deep neural network, the apparatus further comprising:

8. An electronic device comprising a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the full waveform inversion method of any one of claims 1 to 5.

9. A computer-readable storage medium having a computer-executable program stored thereon, which when invoked and executed by a processor, causes the processor to implement the full waveform inversion method of any one of claims 1 to 5.