Seismic data convolution cross-correlation intelligent FWI method

Abstract

The invention is suitable for the technical field of oil-gas exploration, and provides a seismic data convolution cross-correlation intelligent FWI method, which integrates a proposed seismic data convolution cross-correlation FWI target equation into a deep learning framework, and combines an automatic encoder technology to significantly reduce the nonlinearity of the traditional FWI, effectively avoid the cycle jump problem, and improve the accuracy of the FWI. Therefore, the speed model inversion precision of the FWI is improved. Finally, the FWI technology which can still carry out stable and accurate high-resolution speed modeling under the conditions that seismic source wavelets are unknown, the initial speed is unknown and low-frequency data are missing is realized, and the method has important significance on seismic exploration.

Description

Technical Field

[0001] The invention belongs to the technical field of oil and gas exploration, and in particular relates to an intelligent FWI method for convolution cross-correlation of seismic data. Background Art

[0002] Full waveform inversion (FWI) is an important high-resolution seismic exploration inversion technology, which is widely used in underground structure imaging and physical parameter estimation in the field of oil and gas exploration. The core idea of ​​FWI is to invert the physical parameters of the underground medium by minimizing the error between the observed seismic data and the simulated data. Its main process includes: first, use the initial model for forward simulation to obtain the predicted seismic wave field; then calculate the error between the predicted data and the actual observed data, and use the error to update the physical model of the underground medium; this process is repeated iteratively until the error converges to obtain the final high-resolution underground model.

[0003] Traditional FWI methods face a series of challenges in practical applications. First, FWI is highly dependent on accurate source wavelets, but in actual seismic exploration, source wavelets are often difficult to estimate accurately, and the amplitude of simulated seismic data is difficult to match with the observed data, which affects the accuracy and stability of the inversion. In addition, the objective function of FWI is highly nonlinear. When the travel time difference between the observed data and the predicted data exceeds half a cycle, a cycle jump problem will occur, resulting in the emergence of local minima, which affects the accuracy of the inversion. Although a good initial model and low-frequency data can mitigate the impact of the cycle jump problem, this prior information is difficult to obtain in practice, causing FWI to fall into a local solution in practical applications and it is difficult to obtain a global optimal solution. Summary of the invention

[0004] The purpose of the embodiment of the present invention is to provide an intelligent FWI method for convolution cross-correlation of seismic data, aiming to solve the problems raised in the above background technology.

[0005] The embodiment of the present invention is implemented as follows: a seismic data convolution cross-correlation intelligent FWI method, comprising the following steps: Step 1: Establish a deep learning framework to predict speed model; The intelligent FWI network structure based on the autoencoder is established, and its input is a random speed set V from the OpenFWI dataset; this speed set is input to the encoder, latent space and decoder; finally, the intelligent FWI framework outputs the predicted speed model ; Step 2: Define the loss function of the deep learning framework; Intelligent FWI framework outputs predicted velocity model back, Introduced into the wave equation solver to generate predicted earthquake data ; Then, using the observed data With forecast data Perform fitting to achieve network parameters Updates; Step 3: Conduct network training and output results; This model does not require the establishment of any training data set, and only relies on the physical constraints of the wave equation and data fitting. Its network parameter training method is completely unsupervised training. Specifically, the Adam optimizer commonly used in the field of deep learning can be used to complete network parameter optimization and finally obtain the inverted velocity model.

[0006] A further technical solution is characterized in that, in step 1, an autoencoder is used to represent the velocity model, and the mathematical expression is as follows: (1); in, is the function represented by the autoencoder, is the output velocity model, is the autoencoder input, is an updateable parameter in the network.

[0007] A further technical solution is that in step 2, the mathematical expression of the objective function of the autoencoder intelligent FWI network is as follows: (2); in, represents the convolution operation; represents the cross-correlation operation; The coordinates of the geophones that collect seismic data; is the receiver position of a single reference seismic trace; Represents simulated data With observational data Convolution of reference seismic traces; Represents the reference seismic trace for the simulated data With observational data The source wavelet information in the observed and simulated seismic data after convolution with the reference seismic trace is balanced, eliminating the influence of the uncertainty of the source wavelet of the simulated data on FWI. In order to eliminate the amplitude information of the seismic data, the convolved seismic data is normalized. and are the normalized convolution simulation data and convolution observation data, respectively. The operation of finding the modulus of the data. The cross-correlation calculation of the normalized convolution simulation data and the convolution observation data can measure the phase fit between the data, eliminating the dependence of traditional FWI on data amplitude fitting.

[0008] A further technical solution is that in step 3, the intelligent data convolution cross-correlation FWI objective function using the physical constraint deep learning framework is as follows: (3); This objective function adds an initial velocity model Prior information regularization constraints, For this purpose, the weight coefficient of the regularization term is used to optimize the network parameters of the autoencoder through the physical constraint deep learning framework, and finally the speed model is output.

[0009] An intelligent FWI method for seismic data convolution and cross-correlation provided by an embodiment of the present invention aims to eliminate the uncertainty and amplitude dependence of source wavelets in FWI by proposing a new seismic data convolution and cross-correlation target equation, thereby improving the stability and accuracy of the FWI inversion process. The method integrates the proposed seismic data convolution and cross-correlation FWI target equation into the deep learning framework, and combines it with the autoencoder technology to significantly reduce the nonlinearity of traditional FWI, effectively avoid the cycle jump problem, and thus improve the velocity model inversion accuracy of FWI. Finally, an FWI technology that can still perform stable and accurate high-resolution velocity modeling is realized when the source wavelet is unknown, the initial velocity is unknown, and the low-frequency data is missing. Its ability to overcome the source wavelet error and cycle jump problems provides a key technology for high-precision and high-resolution seismic exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is an intelligent FWI framework based on autoencoder; Figure 2 is the true Marmousi model and the initial model of FWI, where a is the true Marmousi model and b is the initial model of FWI; Figure 3 are the true and false source wavelets; Figure 4 for the spectrum of truth and falsehood; Figure 5 The shot gather records of seismic data collected with different unknown sources, where a is the source at x=2 km, b is the source at x=4 km, and c is the source at x=6 km; Figure 6 The traditional FWI inversion results using wrong source wavelet; Figure 7 The traditional FWI inversion results using real source wavelet; Figure 8 The present invention uses the convolution cross-correlation of the erroneous source wavelet to intelligentize the FWI inversion results. DETAILED DESCRIPTION

[0011] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0012] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0013] An embodiment of the present invention provides a seismic data convolution cross-correlation intelligent FWI method, comprising the following steps: Step 1: Establish a deep learning framework to predict speed model; The established FWI network structure based on autoencoder is as follows: Figure 1 As shown in Figure 1, its input is a random speed set V from the OpenFWI dataset; this speed set is input to the encoder, latent space, and decoder; finally, the intelligent FWI framework outputs the predicted speed model ; Step 2: Define the loss function of the deep learning framework; Intelligent FWI framework outputs predicted velocity model back, Introduced into the wave equation solver to generate predicted earthquake data ; Then, using the observed data With forecast data Perform fitting to achieve network parameters Updates; Step 3: Conduct network training and output results; The intelligent FWI technology developed by the present invention, which is independent of source wavelets and data amplitude, does not require the establishment of any training data set, but only relies on the physical constraints of wave equations and data fitting, and its network parameter training method is completely unsupervised training. The present invention uses the Adam optimizer commonly used in the field of deep learning to complete network parameter optimization and finally obtain the inverted velocity model.

[0014] As a preferred embodiment of the present invention, in step 1, an autoencoder is used to represent the velocity model, and the mathematical expression is as follows: (1); in, is the function represented by the autoencoder, is the output velocity model, is the autoencoder input, is an updateable parameter in the network.

[0015] As a preferred embodiment of the present invention, in step 2, the mathematical expression of the objective function based on the autoencoder intelligent FWI network is as follows: (2); in, represents the convolution operation; represents the cross-correlation operation; The coordinates of the geophones that collect seismic data; is the receiver position of a single reference seismic trace; Represents simulated data With observational data Convolution of reference seismic traces; Represents the reference seismic trace for the simulated data With observational data The source wavelet information in the observed and simulated seismic data after convolution with the reference seismic trace is balanced, eliminating the influence of the uncertainty of the source wavelet of the simulated data on FWI. In order to eliminate the amplitude information of the seismic data, the convolved seismic data is normalized. and are the normalized convolution simulation data and convolution observation data, respectively. The operation of finding the modulus of the data. The cross-correlation calculation of the normalized convolution simulation data and the convolution observation data can measure the phase fit between the data, eliminating the dependence of traditional FWI on data amplitude fitting.

[0016] As a preferred embodiment of the present invention, in step 3, the intelligent data convolution cross-correlation FWI objective function using the physical constraint deep learning framework is as follows: (3); This objective function adds an initial velocity model Prior information regularization constraints, For this purpose, the weight coefficient of the regularization term is used to optimize the network parameters of the autoencoder through the physical constraint deep learning framework, and finally the speed model is output.

[0017] As a preferred embodiment of the present invention, the method is applied to the Marmousi model, which contains 121×426 grid points and a spatial sampling interval of 20 m. Figure 2 a is the real Marmousi model, which contains complex fault structures. Figure 2 b is the initial model of FWI, which does not contain any detailed construction information.

[0018] In the test, the Ricker wavelet that does not contain frequencies below 3 Hz is used as the real source wavelet ( Figure 3The blue line in the middle generates 40 shot gatherings as observed seismic data, with 40 sources and 426 geophones evenly distributed on the surface. To simulate the actual data FWI application scenario, the source wavelet is unknown. This test uses Figure 3 The error wavelet (red line) in is used as the source wavelet of FWI. Figure 3 The real and erroneous source wavelets in the image can be clearly found to have differences in phase and waveform. By calculating their spectra, such as Figure 4 As shown (the blue line represents the real one, and the red line represents the wrong one), it can be found that their frequency components are also different. The real source wavelet lacks low-frequency data to simulate the real seismic data acquisition scenario, and the wrong source wavelet contains low-frequency data. Figure 5 The following table shows the shot gather records of seismic data collected at different earthquake sources in this test, where a is the earthquake source at x=2 km, b is the earthquake source at x=4 km, and c is the earthquake source at x=6 km.

[0019] This test used 2000 iterations for velocity inversion. Figure 6 The traditional FWI inversion results using the wrong source wavelet are shown. This result shows that due to the error in the source wavelet, there are obvious errors in the shallow layer of the inversion velocity model, and the deep structure is not effectively restored. Figure 7 The traditional FWI inversion results using the real source wavelet are shown. Although the velocity structure of the shallow layer is well restored, the deep structure has obvious inversion errors caused by the cycle jump problem. In contrast, the seismic data convolution cross-correlation FWI proposed in this invention can obtain high-precision inversion results using the wrong source wavelet, such as Figure 8 As shown, compared with the true velocity model ( Figure 2 a) Highly similar, overcoming the challenges of FWI caused by erroneous source wavelets and missing low-frequency data.

[0020] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent FWI method for convolution cross-correlation of seismic data, characterized in that: Step 1: Establish a deep learning framework to predict speed model; The intelligent FWI network structure based on the autoencoder is established, and its input is a random speed set V from the OpenFWI dataset; this speed set is input to the encoder, latent space and decoder; finally, the intelligent FWI framework outputs the predicted speed model ; Step 2: Define the loss function of the deep learning framework; Intelligent FWI framework outputs predicted velocity model back, Introduced into the wave equation solver to generate predicted earthquake data ; Then, using the observed data With forecast data Perform fitting to achieve network parameters Updates; Step 3: Conduct network training and output results; The Adam optimizer is used to optimize the network parameters and finally obtain the inverted velocity model.

2. The seismic data convolution cross-correlation intelligent FWI method according to claim 1, characterized in that: In step 1, an autoencoder is used to represent the velocity model, and the mathematical expression is as follows: （1）； in, is the function represented by the autoencoder, is the output velocity model, is the autoencoder input, is an updateable parameter in the network.

3. The seismic data convolution cross-correlation intelligent FWI method according to claim 2, characterized in that: In step 2, the mathematical expression of the objective function based on the autoencoder intelligent FWI network is as follows: （2）； in, represents the convolution operation; represents the cross-correlation operation; The coordinates of the geophones that collect seismic data; is the receiver position of a single reference seismic trace; Represents simulated data With observational data Convolution of reference seismic traces; Represents the reference seismic trace for the simulated data With observational data Convolution; In order to eliminate the amplitude information of seismic data, the convolved seismic data are normalized; In formula 2 and are the normalized convolution simulation data and convolution observation data, respectively. For the data modulus operation, the normalized convolution simulation data and the convolution observation data are cross-correlated.

4. The seismic data convolution cross-correlation intelligent FWI method according to claim 3, characterized in that: In step 3, the intelligent data convolution cross-correlation FWI objective function using the physical constraint deep learning framework is as follows: （3）； This objective function adds an initial velocity model Prior information regularization constraints, For this purpose, the weight coefficient of the regularization term is used to optimize the network parameters of the autoencoder through the physical constraint deep learning framework, and finally the speed model is output.

Images