A seismic waveform inversion method and system based on deep learning image registration

Abstract

The present application relates to the technical field of seismic data processing, in particular to a seismic waveform inversion method and system based on deep learning image registration, the method comprising: constructing a reflectivity model based on least squares migration of small offset observation data and a background velocity field, obtaining a simulated seismic record image according to Born forward modeling of the reflectivity and the background velocity field; obtaining a seismic record deformation field by deep learning image registration based on the simulated seismic record image and a seismic observation image; calculating a normalized deformation field, and judging whether the normalized deformation field meets a preset condition; if yes, the iteration process is terminated; otherwise, the next step is executed; updating the background velocity field by a conjugate gradient algorithm according to the seismic record deformation field and a companion source, and repeating the above steps. The present application utilizes the high efficiency and high precision of deep learning, and avoids the problem of local minimum value of cycle skipping easily occurring in waveform matching.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and specifically to a seismic waveform inversion method and system based on deep learning image registration. Background Art

[0002] Seismic exploration is a geophysical exploration method that uses artificially excited elastic waves to exploit differences in the elasticity of the underground medium, observe and analyze the propagation patterns of seismic waves generated by artificial earthquakes, and infer the properties and morphology of underground rock formations. Velocity modeling is a key step in seismic exploration data processing, and its accuracy directly affects the quality of seismic imaging. Full waveform inversion (FWI), as a nonlinear parameter inversion method, uses synthetic data from all seismic wave information (including travel time, amplitude, multiple waves, scattered waves, etc.) to fit observed data to estimate underground velocity density parameters. This method can obtain low, medium, and high full-wavenumber velocity information, and the inversion results can accurately characterize and describe complex underground structural morphology.

[0003] However, traditional full-waveform inversion methods use the least-squares norm (L2) to measure the difference between observed and calculated data. This measurement method can lead to cycle skipping when there is a large discrepancy between the simulated and observed data, especially when the simulated data shifts by more than half a cycle relative to the observed data. This requires a sufficiently good initial model for FWI. If the model can predict the data within half a cycle, FWI converges to the correct solution. Otherwise, FWI will be affected by cycle skipping and be led to a local minimum.

[0004]

[0005] Among them, t represents the time direction; h represents the spatial direction; (t, h) represents that the data is a two-dimensional matrix; d cal (t, h) represents simulated data, which is seismic data obtained by computer numerical simulation; d obs (t, h) represents the observed data, which are real seismic data collected in the field by geophysical prospecting vessels; Indicates data difference, d cal (t, h) and d obs The squared difference of (t, h).

[0006] To avoid the inversion falling into the local minimum, some scholars have proposed methods such as cross-correlation, dynamic time warping, and optimal transport. When calculating the one-dimensional travel time difference of the wave field, the cross-correlation function is often used to calculate the time shift as the optimization target, that is:

[0007]

[0008] Where Δτ represents the arrival time difference between the observed data and the simulated data as the waveform time of the image sequence; s represents the data index of an artificial source (shot point); r represents the data index of a seismometer; J t It represents the sum of the squares of the arrival time differences of the waveforms of all seismic traces in the work area.

[0009] From a higher level, the process of finding differences and optimizing seismic records can be solved by image registration. The registration problem is simply to align one image to another so that the aligned images are as similar as possible. That is, given a floating image I m and a fixed image I f , predict the deformation field Φ(t, h), I w represents the floating image after the floating image is restored by the deformation field Φ(t, h).

[0010] I w =I m *Φ(t,h) (3)

[0011] The deformation field Φ(t, h) represents the position movement of the sampling point in the t direction and the h direction.

[0012] Registration methods can be categorized as traditional methods and machine learning-based methods. Traditional image registration methods use local image similarity-based optimization methods or feature registration methods. The former uses continuous iteration to find the optimal solution for spatial transformation, and is more effective when dealing with image registration problems with small deformations. The latter mostly uses feature points for matching, which performs better when the image is severely deformed, but the computational cost is high. Traditional methods require repeated manual adjustment of parameters and are not suitable for automated large-scale data operations. Machine learning-based registration methods refer to registration methods that use neural network training. In this method, parameters are shared. First, a model is trained using a large amount of data, and then the trained model is used to register new images. Although the training process of this method is computationally expensive, the registration process is an order of magnitude faster than traditional methods.

[0013] Currently, full waveform inversion (FWI) technology faces several key technical issues in practice:

[0014] (1) When low-frequency data is missing and the L2 objective function is used to calculate the data residual, cycle jumps (waveform dislocation) are likely to occur. Velocity sections or diffraction points appear underground, and the reflected waves may have complex cross-overlapping, which further aggravates the cycle jump problem.

[0015] (2) The amplitude of seismic records is affected by velocity, Lame constant, and density coupling, and has a strong uncertainty. Currently, due to the limitations of limited observation aperture and acoustic wave equation simulation, only the phase accuracy of the wave field can be guaranteed. On this basis, it is difficult to obtain an accurate velocity structure based on the L2 objective function due to the influence of amplitude.

[0016] (3) Methods for calculating phase differences include cross-correlation (CC), dynamic time warping (DTW), and optimal transport (OT). If cross-correlation is used to calculate phase differences, it is necessary to divide the time window and pick the maximum time movement difference. If the analysis time window is too large, the resolution of the cross-correlation phase difference result will be very low. If the analysis time window is too small, the stability will be poor. The dynamic time warping (DTW) algorithm can only calculate one-dimensional differences in the time direction. When the shallow velocity changes laterally, the vertex of the diffraction wave hyperbola moves in time and space between the simulated data and the observed data. The one-dimensional method cannot describe this lateral movement. The optimal transport objective function has a high computational cost (OT). The calculation time of the transport scheme for two-dimensional images may exceed the forward shot gathering time. Summary of the Invention

[0017] The purpose of the present invention is to provide a seismic waveform inversion method and system based on deep learning image registration in order to solve at least one of the above technical problems.

[0018] The present invention provides a seismic waveform inversion method based on deep learning image registration, comprising the following steps:

[0019] Step S102, performing least square migration based on the small offset observation data and the background velocity field to construct a reflection coefficient model, and performing Born forward modeling based on the reflection coefficient and the background velocity field to obtain a simulated seismic recording image;

[0020] Step S104, based on the simulated earthquake recording image and the earthquake observation image, using deep learning image registration to obtain the earthquake recording deformation field;

[0021] Step S106, calculate the normalized deformation field using the following formula:

[0022] Φ_nor(t,h)=Φ(t,h) / wavelet_length(t,h) (4)

[0023] Where Φ_nor(t, h) represents the normalized deformation field; Φ(t, h) represents the deformation field of the seismic record; wavelet_length(t, h) represents the wavelength of the main frequency wave of the current data, t represents the time direction, and h represents the spatial direction;

[0024] Determine whether the normalized deformation field meets the preset conditions; if so, terminate the iterative process; otherwise, execute step S108;

[0025] Step S108, calculating the accompanying earthquake source according to the earthquake record deformation field, loading the accompanying earthquake source into the wave equation for wave field propagation, using cross-correlation to calculate the low wave number gradient component to update the background velocity field, and repeating steps S102 to S106.

[0026] Furthermore, in step S102, the least squares shift is implemented by using a shift deconvolution method.

[0027] Furthermore, in step S104, the simulated earthquake recording image is used as a floating image, and the earthquake observation image is used as a fixed image, and deep learning image registration is performed to obtain the earthquake recording deformation field.

[0028] Furthermore, step S104 includes the following steps:

[0029] Input the image to be registered, and output the deformation vector field after learning through a multi-scale parallel fully convolutional neural network; wherein the image to be registered includes the floating image and the fixed image;

[0030] The floating image is based on the deformation vector field and uses a spatial transformation network to generate a deformable image;

[0031] The difference between the fixed image and the deformed image is calculated according to the loss function, and the parameters of the earthquake record deformation field are updated, and multiple iterations are performed to obtain the final earthquake record deformation field.

[0032] Furthermore, a three-layer multi-scale parallel downsampling module is used to reduce the image size of the image to be registered and obtain a multi-scale downsampling feature map;

[0033] The multi-scale downsampled feature map is sequentially passed through a pyramid dilated convolution module and an adaptive channel attention module for feature extraction and feature processing;

[0034] The feature map processed by the adaptive channel attention module is input into the direction perception layer and the upsampling layer to obtain the deformation vector field of the input image pair.

[0035] Furthermore, the loss function adopts the following formula:

[0036] L(I f , I m ,Φ)=L(I f , I m ,Φ)+λL smoth (I f ,I m ,Φ) (5)

[0037] Among them, I f Indicates a fixed image; I mrepresents the floating image; Φ represents the earthquake record deformation field Φ(t, h); L(I f , I m , Φ) represents I m The grayscale similarity of the transformed image; λ represents the spatial smoothness coefficient; L smoth (I f ,I m ,Φ) represents the smoothing process of image grayscale similarity.

[0038] Furthermore, in step S106:

[0039] The preset condition is: the normalized deformation field Φ_nor(t,h)<1%.

[0040] Furthermore, step S108 includes the following steps:

[0041] The accompanying earthquake source is calculated according to the following formula:

[0042]

[0043] Where t represents the time direction; h represents the spatial direction; Δτ represents the arrival time difference between the observed data and the simulated data as the waveform time of the image sequence; Δs represents the arrival time difference between the observed data and the simulated data as the waveform space of the image sequence; represents the first-order time derivative of the observed data; represents the second-order time derivative of the observed data; d cal Represents simulation data; W is the proportional coefficient, which is used to balance the proportional relationship between time and space; represents the time-delayed accompanying source; It indicates that the accompanying earthquake source in the earthquake record has a lateral movement of the diffraction wave apex due to the lateral change of velocity;

[0044] The companion source is loaded into the wave equation for wave field propagation, and the low wave number gradient component is calculated by cross-correlation to update the background velocity field, specifically through the following formula:

[0045] The gradient is:

[0046]

[0047] Where, f(t, h) represents the source wave field; g t (t+Δτ, h) represents the accompanying source wave field; represents the gradient, represents the cross-correlation between the two at each time in the underground; J represents the target function, and m represents the particle velocity.

[0048] The present invention also provides a seismic waveform inversion system based on deep learning image registration, comprising:

[0049] A simulated seismic recording image acquisition module is used to perform least squares migration based on small offset observation data and background velocity field to construct a reflection coefficient model, and to perform Born forward modeling based on the reflection coefficient and the background velocity field to obtain a simulated seismic recording image;

[0050] An earthquake record deformation field acquisition module is used to acquire the earthquake record deformation field by using deep learning image registration based on the simulated earthquake record image and the earthquake observation image;

[0051] The seismic waveform inversion module is used to calculate the normalized deformation field using the following formula,

[0052] Φ_nor(t,h)=Φ(t,h) / wavelet_length(t,h) (4)

[0053] Where Φ_nor(t, h) represents the normalized deformation field; Φ(t, h) represents the deformation field of the seismic record; wavelet_length(t, h) represents the wavelength of the main frequency wave of the current data, t represents the time direction, and h represents the spatial direction;

[0054] The seismic waveform inversion module is further used to determine whether the normalized deformation field meets a preset condition; if so, the iterative process is terminated; otherwise, the background velocity field is updated through the velocity update module;

[0055] The velocity update module is used to calculate the accompanying earthquake source according to the earthquake record deformation field, load the accompanying earthquake source into the wave equation for wave field propagation, and use cross-correlation to calculate the low wave number gradient component to update the background velocity field.

[0056] The present invention also provides an electronic device comprising: a processor, a memory, and a computer program stored in the memory and runnable on the processor, wherein when the computer program is executed by the processor, any seismic waveform inversion method based on deep learning image registration is implemented.

[0057] The beneficial effects of the present invention are:

[0058] (1) For the local extrema problem caused by frequency jumps in classical full waveform inversion, the present invention performs deep learning global matching to improve the dependence on the initial velocity model and adopts a deformation field to avoid the uncertainty problem caused by matching amplitude.

[0059] (2) For the calculation of deformation vector fields, the present invention adopts a multi-scale parallel downsampling model and an adaptive channel attention module, which is not only suitable for the calculation of large deformation fields but also has high registration accuracy for small deformations. Moreover, after training, the present invention is significantly more efficient than traditional image registration methods.

[0060] (3) The deformation vector field of the present invention is a two-dimensional matrix. Compared with the one-dimensional time delay obtained by the current cross-correlation, dynamic time warping, transport and other methods, it is more reasonable and easier to obtain a more accurate velocity model.

[0061] (4) The present invention uses Born forward modeling to obtain reflected wave data for matching, which is not only suitable for transmitted waves but also for reflected waves, avoiding the problem of insufficient speed update depth of classical full waveform inversion.

[0062] This method leverages the high efficiency and precision of deep learning to avoid the cycle skipping problem that can easily occur with waveform matching, leading to local minima. It can also construct a more stable velocity model than conventional methods, significantly improving the quality of seismic migration imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The process of the seismic waveform inversion method based on deep learning image registration in one embodiment of the present invention is as follows Figure 1 ;

[0064] Figure 2 The process of the seismic waveform inversion method based on deep learning image registration in one embodiment of the present invention is as follows Figure 2 ;

[0065] Figure 3 This is a comparison chart of non-rigid image registration shot gather tests according to an embodiment of the present invention;

[0066] Figure 4 This is a network structure diagram of a multi-scale downsampling module according to an embodiment of the present invention;

[0067] Figure 5 A schematic diagram of a multi-scale parallel fully convolutional neural network according to an embodiment of the present invention;

[0068] Figure 6 An initial model observation data shot gather according to one embodiment of the present invention;

[0069] Figure 7 A de-migration simulated shot gather according to an embodiment of the present invention;

[0070] Figure 8 The initial model data shot gather deformation field of one embodiment of the present invention;

[0071] Figure 9 An initial model data shot gather associated with a seismic source in one embodiment of the present invention;

[0072] Figure 10 is an initial model velocity update gradient in one embodiment of the present invention;

[0073] Figure 11 : Figure 11 (a) is a true velocity profile of a two-dimensional depression model; Figure 11 (b) is the initial velocity model; Figure 11 (c) is the velocity model after model inversion update;

[0074] Figure 12 : Figure 12 (a) is the migration result of the true velocity model; Figure 12 (b) is the migration result of the initial velocity model; Figure 12 (c) is the migration result of the inverted velocity model;

[0075] Figure 13 This is a structural diagram of a seismic waveform inversion system based on deep learning image registration according to an embodiment of the present invention. DETAILED DESCRIPTION

[0076] The present invention will now be discussed with reference to exemplary embodiments. It should be understood that the embodiments discussed are only intended to enable those skilled in the art to better understand and implement the present invention, rather than to imply any limitation on the scope of the present invention.

[0077] As used herein, the term "including" and variations thereof are to be interpreted as open-ended terms meaning "including, but not limited to." The term "based on" is to be interpreted as "based, at least in part, on." The terms "one embodiment" and "an embodiment" are to be interpreted as "at least one embodiment."

[0078] Example 1

[0079] Figure 1 The process of the seismic waveform inversion method based on deep learning image registration in one embodiment of the present invention is as follows Figure 1 .like Figure 1 As shown, according to one embodiment of the present invention, a seismic waveform inversion method based on deep learning image registration includes the following steps:

[0080] Step S102, performing least squares migration based on the small offset observation data and the background velocity field to construct a reflection coefficient model, and performing Born forward modeling based on the reflection coefficient and the background velocity field to obtain a simulated seismic recording image;

[0081] Step S104, based on the simulated earthquake recording image and the earthquake observation image, deep learning image registration is used to obtain the earthquake recording deformation field;

[0082] Step S106, calculate the normalized deformation field using the following formula:

[0083] Φ_nor(t,h)=Φ(t,h) / wavelet_length(t,h) (4)

[0084] Where Φ_nor(t, h) represents the normalized deformation field; Φ(t, h) represents the deformation field of the seismic record; wavelet_length(t, h) represents the wavelength of the main frequency wave of the current data, t represents the time direction, and h represents the spatial direction;

[0085] Determine whether the normalized deformation field meets the preset conditions; if so, the iterative process terminates; otherwise, execute step S108;

[0086] Step S108, calculate the accompanying earthquake source according to the earthquake record deformation field, load the accompanying earthquake source into the wave equation for wave field propagation, use cross-correlation to calculate the low wave number gradient component to update the background velocity field, and repeat steps S102 to S106.

[0087] In this embodiment, the seismic waveform inversion method is combined with deep learning image registration technology and traditional seismic data processing technology.

[0088] First, using small-offset observation data and background velocity fields, least-squares migration and Born forward modeling are used to simulate seismic images, providing basic data for subsequent image registration. Preferably, in this embodiment, least-squares migration is implemented using a migration deconvolution approach. This migration deconvolution approach constructs a frequency-domain wavefield propagation operator within the least-squares migration imaging method and, based on deconvolution imaging conditions, constructs mutually conjugate forward and migration operators, resulting in more accurate imaging results. This approach can effectively eliminate crosstalk noise in multiple wave imaging results, improving the resolution and signal-to-noise ratio of the imaging results.

[0089] Next, through deep learning image registration technology, this method can accurately obtain the earthquake deformation field, which is a key step in achieving seismic waveform inversion. Preferably, in this embodiment, the simulated earthquake recording image is used as a floating image, and the observation image (seismic observation image) acquired by seismic acquisition is used as a fixed image, and deep learning image registration is performed to obtain the earthquake deformation field.

[0090] After obtaining the deformation field of the seismic record, a normalized deformation field is further obtained through normalization processing, and the preset conditions are used to determine whether the iterative process is terminated, which reflects the flexibility and adaptability of the algorithm. If the preset conditions are not met, the iteration continues, and the wave field propagation is carried out by calculating the accompanying earthquake source and loading it into the wave equation, and then the low-wavenumber gradient component is calculated by cross-correlation to update the background velocity field, thereby achieving gradual optimization of the waveform inversion. Preferably, in this embodiment, the preset conditions are: the normalized deformation field Φ_nor(t, h) <1%; when the observed data and the simulated data are consistent, the normalized deformation field is 0, and when the observed data and the simulated data are significantly different, the normalized deformation field exceeds 1. Here, setting Φ_nor(t, h) < 1% means that the iteration is terminated when the deformation range is less than 1%. Otherwise, go to step S108 for velocity update.

[0091] The entire process fully utilizes the advantages of deep learning in image registration and combines it with expertise in seismic data processing to form an efficient and accurate seismic waveform inversion method.

[0092] The present invention significantly improves the accuracy and efficiency of seismic data processing: through deep learning image registration technology, the deformation field of seismic records can be accurately obtained, providing a more accurate basis for waveform inversion; by iteratively updating the background velocity field, the waveform inversion results are continuously optimized, so that the final seismic image obtained is closer to the actual situation; the present invention has good flexibility and adaptability, and can adjust parameters according to specific data characteristics and needs to meet the seismic data processing needs in different scenarios.

[0093] According to one embodiment of the present invention, step S104 includes the following steps:

[0094] Step S1042: input the image to be registered, and output a deformation vector field after learning through a multi-scale parallel fully convolutional neural network; wherein the image to be registered includes a floating image and a fixed image;

[0095] The images to be registered include floating images (simulated seismic recordings) and fixed images (seismic observations). A fully convolutional neural network (FCN) is a deep learning model for image-to-image translation, suitable for tasks such as image segmentation and image registration. Its multi-scale parallel architecture processes images in parallel at different scales, helping to capture deformation information at different scales. The deformation vector field output by the network after learning describes the displacement vector of each pixel from the floating image to the fixed image, i.e., the deformation information between the images.

[0096] Step S1044: The floating image is transformed into a deformed image using a spatial transformation network based on the deformation vector field.

[0097] The Spatial Transformer Network is a learnable module that spatially transforms an image based on a deformation vector field as input. Using the deformation vector field to transform the floating image yields a deformed image that is morphologically closer to the fixed image.

[0098] Step S1046 , calculating the difference between the fixed image and the deformed image according to the loss function and updating the parameters of the earthquake record deformation field, performing multiple iterations to obtain the final earthquake record deformation field.

[0099] Through optimization methods such as backpropagation and gradient descent, the parameters of the multi-scale parallel fully convolutional neural network are updated according to the gradient of the loss function, making the deformation vector field output by the network more accurate. Through iterative optimization, the difference between the fixed image and the deformed image is gradually reduced, and finally an accurate earthquake record deformation field is obtained. Preferably, in this embodiment, the loss function adopts the following formula:

[0100] L(I f , I m ,Φ)=L(I f ,I m ,Φ)+λL smoth (I f , I m , Φ) (5)

[0101] Among them, I f Indicates a fixed image; I m represents the floating image; Φ represents the earthquake record deformation field Φ(t, h); L(I f , I m , Φ) represents I m The grayscale similarity of the transformed image; λ represents the spatial smoothness coefficient; L smoth (I f , I m , Φ) represents the smoothing of image grayscale similarity.

[0102] The present invention achieves precise registration of simulated seismic recording images and seismic acquisition observation images by introducing a multi-scale parallel fully convolutional neural network and a spatial transformation network. It can capture deformation information at different scales, gradually reduce the differences between images through iterative optimization, and ultimately obtain an accurate seismic recording deformation field, providing strong technical support for seismic waveform inversion and data analysis, and improving the accuracy and efficiency of seismic data analysis.

[0103] According to one embodiment of the present invention, in step S1042:

[0104] A three-layer multi-scale parallel downsampling module is used to reduce the image size and obtain a multi-scale downsampling feature map for the registered image;

[0105] The multi-scale downsampled feature maps are sequentially passed through the pyramid void convolution module and the adaptive channel attention module for feature extraction and feature processing;

[0106] The feature map processed by the adaptive channel attention module is input into the direction perception layer and the upsampling layer to obtain the deformation vector field of the input image pair.

[0107] In this implementation, based on the structure of a fully convolutional neural network (FCN), a three-layer multi-scale parallel downsampling module is used to reduce the image size and obtain a multi-scale downsampling feature map. Multi-scale feature extraction is performed through a pyramid dilated convolution module, and then three layers of adaptive channel attention are used for feature processing. Finally, a dense vector deformation field of the input image pair is obtained through a direction perception layer and an upsampling layer.

[0108] Multi-scale parallel downsampling model: To address the need for extracting larger deformation features in the seismic image registration task, a pyramid pooling module (PPM) is used to form a multi-scale parallel downsampling module to extract multi-scale features. This module first performs average pooling on the input image with steps of 1, 2, and 4, respectively, to obtain feature maps of three different scales. The module then uses a batch normalization layer and a ReLu activation function to simultaneously perform bilinear interpolation on the three feature maps and the input image to obtain feature maps with a resolution of 1 / 2 of the original image at different scales. Finally, the three feature maps of different scales and the original feature map are feature-concatenated to output a cascaded feature map with 4xc channels. This technology primarily employs a multi-scale strategy, enabling the sampling of large-scale and small-scale deformation features.

[0109] Pyramid Convolution Module: After multi-scale parallel downsampling, ordinary convolution cannot simultaneously process feature maps containing information from many different scales. Furthermore, downsampling uses pooling layers to concentrate features, which can easily cause information loss and affect registration. Therefore, a pyramid dilated convolution module is used for feature extraction to address this issue. It requires setting a dilation factor to control the dilation holes in the convolution kernel. When the dilation factor is 1, dilated convolution operates the same as ordinary convolution. When the dilation factor is greater than 1, each operator in the convolution kernel skips a corresponding number of points, forming a larger and sparser convolution kernel. This technology uses a multi-layer convolution structure combined with a multi-scale strategy to capture image deformation characteristics at different scales.

[0110] Adaptive Attention Module: The attention mechanism guides the model to prioritize feature maps with high information content, improving model performance while avoiding increasing network depth. Because the model extracts features through a multi-scale parallel downsampling module and a pyramid-shaped atrous convolutional layer, the number of feature maps expressing large deformations (such as strong reflection axes) is significantly higher than that of small features (such as high-frequency seismic axes). To improve registration accuracy by expressing features with small deformations, an adaptive channel attention module is used. This module applies adaptive channel attention to the feature images after three downsamplings, appropriately increasing the weight of small deformations and improving registration accuracy.

[0111] According to one embodiment of the present invention, step S108 includes the following steps:

[0112] The accompanying earthquake source is calculated according to the following formula:

[0113]

[0114] Where t represents the time direction; h represents the spatial direction; Δτ represents the arrival time difference between the observed data and the simulated data as the waveform time of the image sequence; Δs represents the arrival time difference between the observed data and the simulated data as the waveform space of the image sequence; represents the first-order time derivative of the observed data; represents the second-order time derivative of the observed data; d cal Represents simulation data; W is the proportional coefficient, which is used to balance the proportional relationship between time and space;

[0115] represents the time-delayed accompanying source; It indicates that the accompanying earthquake source in the earthquake record has a lateral movement of the diffraction wave apex due to the lateral change of velocity;

[0116] The accompanying earthquake source is loaded into the wave equation for wave field propagation, and the low wave number gradient component is calculated by cross-correlation to update the background velocity field. Specifically, the following formula is used:

[0117]

[0118] Where, f(t, h) represents the source wave field; g t (t+Δτ, h) represents the accompanying source wave field; represents the gradient, represents the cross-correlation between the two at each time in the underground; J represents the target function, and m represents the particle velocity.

[0119] The present invention utilizes a deep learning network to learn the deformation vector field between seismic images, and realizes high-precision inversion of underground velocity structure through accompanying earthquake source calculation and wave equation wave field propagation. It can significantly improve the accuracy and efficiency of seismic waveform inversion, and provide strong technical support for seismic exploration.

[0120] Example 2

[0121] During seismic full waveform inversion, the difference between observed and simulated records needs to be minimized, and the key is to solve the geometric transformation from simulated data to observed data. Figure 2 The process of the seismic waveform inversion method based on deep learning image registration in one embodiment of the present invention is as follows Figure 2 ; Figure 3 This is a comparison chart of a non-rigid image registration shot gather test in one embodiment of the present invention. The left side of the picture shows the original seismic shot gather, the middle of the picture shows the distorted shot gather, and the right side of the picture shows the distorted restored shot gather. Figure 4 This is a network structure diagram of a multi-scale downsampling module according to an embodiment of the present invention; Figure 5 Schematic diagram of a multi-scale parallel fully convolutional neural network according to an embodiment of the present invention. Figure 2-5 As shown, according to one embodiment of the present invention, the present invention provides a seismic waveform inversion method based on deep learning image registration, in which the observed seismic record is set as the reference image, and the simulated seismic record is set as the floating image, wherein the objective function is defined as the matching residual of the floating and reference images, and a deep learning fully convolutional neural network is used to calculate the geometric transformation field for full waveform inversion.

[0122] The program includes four key steps:

[0123] Step 1: Use the simulated data to calculate the velocity update, perform least squares migration using the small-offset observation data and the background velocity field to establish a reflection coefficient model, and perform Born forward modeling using the reflection coefficient and the background velocity field to obtain a simulated seismic record image.

[0124] Step 2: Registration, using the simulated earthquake record image as a floating image and the observation image collected by earthquake as a fixed image for deep learning image registration to obtain the earthquake record deformation field Φ(t, h), which is the key point of the present invention.

[0125] Step 3: Calculate the normalized deformation field Φ_nor(t, h) = Φ(t, h) / wavelet_length(t, h), where wavelet_length(t, h) represents the wavelength of the dominant frequency wave in the current data. When the observed and simulated data are consistent, the normalized deformation field is 0. When the observed and simulated data differ significantly, the normalized deformation field exceeds 1. Here, Φ_nor(t, h) is set to < 1%, indicating that the iteration terminates if the deformation range is less than 1%. Otherwise, proceed to Step 4 to update the velocity.

[0126] Step 4: Velocity update: calculate the conjugate source of the deformation field of the seismic record, load the accompanying source into the wave equation for wave field propagation, use cross-correlation to calculate the low-wavenumber gradient component to update the background velocity field, and separate and calculate the high-frequency seismic image to update the high-frequency reflection coefficient.

[0127] According to one embodiment of the present invention, step 1 includes the following steps:

[0128] Step 1.1: Use least squares migration to improve the consistency of amplitude and waveform between the observed seismic image and the simulated seismic image; the least squares migration is implemented using a migration deconvolution method.

[0129] Step 1.2: Use the Born forward modeling technique to input the reflection coefficient image and background velocity model to simulate the seismic record image.

[0130] According to one embodiment of the present invention, step 2 includes the following steps:

[0131] Step 2.1: First, input the images to be registered (floating image and fixed image). After learning through a multi-scale parallel fully convolutional neural network, a deformation vector field is output. This vector field represents the displacement of corresponding positions in the floating image and the fixed image.

[0132] In step 2.1, based on the structure of the Fully Convolutional Network (FCN), a three-layer multi-scale parallel downsampling module is used to reduce the image size and obtain a multi-scale downsampled feature map. Multi-scale feature extraction is performed through the pyramid void convolution module, and then three layers of adaptive channel attention are used for feature processing. Finally, a dense vector deformation field of the input image pair is obtained through the direction perception layer and the upsampling layer. Specifically,

[0133] Multi-scale parallel downsampling model: To address the need for extracting larger deformation features in the seismic image registration task, a pyramid pooling module (PPM) is used to form a multi-scale parallel downsampling module to extract multi-scale features. This module first performs average pooling on the input image with steps of 1, 2, and 4, respectively, to obtain feature maps of three different scales. The module then uses a batch normalization layer and a ReLu activation function to simultaneously perform bilinear interpolation on the three feature maps and the input image to obtain feature maps with a resolution of 1 / 2 of the original image at different scales. Finally, the three feature maps of different scales and the original feature map are feature-concatenated to output a cascaded feature map with 4xc channels. This technology primarily employs a multi-scale strategy, enabling the sampling of large-scale and small-scale deformation features.

[0134] Pyramid convolution module: After multi-scale parallel downsampling, ordinary convolution is difficult to simultaneously process feature maps containing information of many different scales, and the downsampling uses the pooling layer to concentrate the features, which easily causes information loss and affects the registration effect. Therefore, the pyramid void convolution module is used for feature extraction to solve the above problems. It requires setting a dilation factor to control the holes in the convolution kernel. When the dilation factor is 1, the void convolution is the same as the ordinary convolution operation; when the dilation factor is greater than 1, each operator in the convolution kernel will skip the corresponding number of points, forming a larger and sparse convolution kernel. This technology uses a multi-layer convolution structure in conjunction with a multi-scale strategy to capture image deformation features of different scales.

[0135] Adaptive Attention Module: The attention mechanism guides the model to prioritize feature maps with high information content, improving model performance while avoiding increasing network depth. Because the model extracts features through a multi-scale parallel downsampling module and a pyramid-shaped atrous convolutional layer, the number of feature maps expressing large deformations (such as strong reflection axes) is significantly higher than that of small features (such as high-frequency seismic axes). To improve registration accuracy by expressing features with small deformations, an adaptive channel attention module is used. This module applies adaptive channel attention to the feature images after three downsamplings, appropriately increasing the weight of small deformations and improving registration accuracy.

[0136] Step 2.2: The floating image is then transformed into a new deformed image using the spatial transformation network based on the deformation vector field.

[0137] Step 2.3: Finally, the difference between the fixed image and the deformed image is calculated based on the loss function (including similarity measurement and smoothness constraint) and the model parameters are updated. After multiple iterations, the final deformable model is obtained.

[0138] The main goal of image registration is to obtain the deformation field Φ(t, h) and find the floating image I m and fixed image I f , the pixel-by-pixel correspondence between them. The present invention designs the grayscale value of the image to maximize an objective function about the degree of matching between images, L(I f , I m , Φ) represents I m The grayscale similarity of the transformed image, L smoth (I f , I m , Φ) represents the smoothing of image grayscale similarity, and λ is a coefficient representing the degree of spatial smoothness.

[0139] L(I f , I m ,Φ)=L(I f , I m ,Φ)+λL smoth (If , I m , Φ)

[0140] According to one embodiment of the present invention, step 4 includes: calculating the adjoint source according to the deformation field Φ(t, h) obtained in step 2; specifically,

[0141]

[0142] Where t represents the time direction; h represents the spatial direction; Δτ represents the arrival time difference between the observed data and the simulated data as the waveform time of the image sequence; Δs represents the arrival time difference between the observed data and the simulated data as the waveform space of the image sequence; represents the first-order time derivative of the observed data; represents the second-order time derivative of the observed data; d cal Represents simulation data; W is the proportional coefficient, which is used to balance the proportional relationship between time and space;

[0143] After derivation, the earthquake record with the source can be obtained, where the first term in the above formula is represents the time-delayed companion source, the second term It indicates the accompanying source of the lateral movement of the diffraction wave apex due to the lateral change of velocity in seismic records;

[0144] The gradient is:

[0145]

[0146] Where, f(t, h) represents the source wave field; g t (t+Δτ, h) represents the companion source wave field; Represents the gradient, and represents the cross-correlation of the two at each time in the underground.

[0147] After determining the objective function in step 4, the nonlinear optimization problem is linearized using the Taylor expansion, and the result is obtained step by step through iteration. This step uses the conjugate gradient method to update the velocity. The conjugate gradient direction is constructed by combining the opposite direction of the gradient and the direction of the previous update. The update step size is obtained by parabolic interpolation.

[0148] The beneficial effects of the present invention are:

[0149] (1) For the local extrema problem caused by frequency jumps in classical full waveform inversion, the present invention performs deep learning global matching to improve the dependence on the initial velocity model and adopts a deformation field to avoid the uncertainty problem caused by matching amplitude.

[0150] (2) For the calculation of deformation vector fields, the present invention adopts a multi-scale parallel downsampling model and an adaptive channel attention module, which is not only suitable for the calculation of large deformation fields but also has high registration accuracy for small deformations. Moreover, after training, the efficiency of this method is significantly improved compared with traditional image registration methods.

[0151] (3) The deformation vector field of the present invention is a two-dimensional matrix. Compared with the one-dimensional time delay obtained by the current cross-correlation, dynamic time warping, transport and other methods, it is more reasonable and easier to obtain a more accurate velocity model.

[0152] (4) The reflected wave data is obtained by Born forward modeling for matching. The present invention is not only suitable for transmitted waves but also for reflected waves, thus avoiding the problem of insufficient speed update depth of the classic full waveform inversion.

[0153] Taking all factors into consideration, the present invention utilizes the high efficiency and high precision of deep learning to avoid the frequency skipping problem that is prone to occur in waveform matching and falls into the local minimum; the present invention can construct a more stable velocity model than conventional methods, thereby significantly improving the quality of seismic migration imaging.

[0154] Example 3

[0155] like Figure 6-Figure 12 As shown, this embodiment is based on the actual application of embodiment 2.

[0156] Figure 11 (a) is a true velocity profile of a 2D depression model with a length of 10,000 m and a depth of 8,000 m. A 2D seismic observation system is arranged with a shot spacing of 200 m and an offset length of 10,000 m. Figure 9 The initial model data shot gather associated with the earthquake source in one embodiment of the present invention is: Figure 9 right Figure 11 (a) The model simulates the acoustic wave equation of a point source. The source function is a Ricker wavelet with a main frequency of 8 Hz. The observation data are 40 shot points evenly distributed on the surface, with a shot point spacing of 250 m. Each shot point contains 801 geophones spaced 12.5 m apart, and the earthquake record formed by surface observations.

[0157] The test adopted Figure 2 The technical flow chart shown below is used to perform velocity inversion experiments on the data. Figure 11 (b) is the initial velocity model, which is the initial velocity model used for testing. The fully convolutional neural network structure used in this experiment is shown in the figure. Figure 5 shown.

[0158] Figure 6 The initial model observation data shot gather of one embodiment of the present invention is an original shot gather from which the transmitted wave has been removed. The curvature of the reflection event is complex and there are a large number of intersections, resulting in high complexity.

[0159] Figure 7 The de-migration simulated shot gather of one embodiment of the present invention is obtained by performing least squares migration in step S202 and forward modeling using the initial velocity Born, and has a shape similar to the original shot gather.

[0160] Figure 8 The initial model data shot gather deformation field of one embodiment of the present invention is calculated using a fully convolutional neural network. It describes the transformation relationship between the de-migration simulation data and the original shot gather fixed data. As can be seen from the figure, the deformation field changes very little at near-offset distances, while the deformation field values ​​at bilateral far-offset distances are larger.

[0161] Figure 9 The deformation field is used to calculate the accompanying source. In this experiment, the accompanying source will be propagated in the reverse direction of the detector wave field and the velocity gradient field will be calculated.

[0162] Figure 10 is an initial model velocity update gradient in one embodiment of the present invention; Figure 11 (c) is the velocity model after model inversion update. After completing 5 velocity iterations using this method, the reflection wave waveform inversion results are as follows: Figure 11 As shown in (c), it can be seen that this method can better restore the low-frequency structural information of the velocity model. In this experiment, the three velocity models were migrated and imaged. Figure 12 (a) is the migration result of the true velocity model, Figure 12 (b) is the migration result of the initial velocity model, which has poor imaging; Figure 12 (c) is the migration result of the inverted velocity model. It can be seen that the migration result of the inverted velocity of this experiment has been significantly improved compared with the initial model imaging 11(b), and most structures have been well imaged.

[0163] Example 4

[0164] Figure 13 This is a structural diagram of a seismic waveform inversion system based on deep learning image registration according to an embodiment of the present invention. Figure 13 As shown, a seismic waveform inversion system based on deep learning image registration includes:

[0165] A simulated seismic recording image acquisition module 10 is used to perform least squares migration based on small offset observation data and background velocity field to construct a reflection coefficient model, and to perform Born forward modeling based on the reflection coefficient and background velocity field to obtain a simulated seismic recording image;

[0166] The earthquake record deformation field acquisition module 20 is used to acquire the earthquake record deformation field based on the simulated earthquake record image and the earthquake observation image by using deep learning image registration;

[0167] The seismic waveform inversion module 30 is used to calculate the normalized deformation field using the following formula:

[0168] Φ_nor(t,h)=Φ(t,h) / wavelet_length(t,h)

[0169] Where Φ_nor(t, h) represents the normalized deformation field; Φ(t, h) represents the deformation field of the seismic record; wavelet_length(t, h) represents the wavelength of the main frequency wave of the current data, t represents the time direction, and h represents the spatial direction;

[0170] The seismic waveform inversion module 30 is also used to determine whether the normalized deformation field meets the preset conditions; if so, the iterative process is terminated; otherwise, the background velocity field is updated through the velocity update module;

[0171] The velocity update module 40 is used to calculate the accompanying earthquake source according to the earthquake record deformation field, load the accompanying earthquake source into the wave equation for wave field propagation, and use cross-correlation to calculate the low wave number gradient component to update the background velocity field.

[0172] An electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, any seismic waveform inversion method based on deep learning image registration is implemented.

[0173] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems and devices can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0174] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention herein is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in this application.

[0175] It should be understood that the size of the serial numbers of each step in the content of the invention and the embodiments of the present invention does not absolutely mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

Claims

1. A seismic waveform inversion method based on deep learning image registration, characterized in that: The following steps are involved: Step S102, performing least square migration based on the small offset observation data and the background velocity field to construct a reflection coefficient model, and performing Born forward modeling based on the reflection coefficient and the background velocity field to obtain a simulated seismic recording image; Step S104, based on the simulated earthquake recording image and the earthquake observation image, using deep learning image registration to obtain the earthquake recording deformation field; Step S106, calculate the normalized deformation field using the following formula: ， in, represents the normalized deformation field; represents the earthquake record deformation field; Indicates the wavelength of the current data's main frequency wave, t indicates the time direction, and h indicates the space direction; Determine whether the normalized deformation field meets the preset conditions; if so, terminate the iterative process; otherwise, execute step S108; Step S108, calculating the accompanying earthquake source according to the earthquake record deformation field, loading the accompanying earthquake source into the wave equation for wave field propagation, using cross-correlation to calculate the low wave number gradient component to update the background velocity field, and repeating steps S102 to S106.

2. The seismic waveform inversion method based on deep learning image registration according to claim 1, characterized in that: In step S102, the least squares shift is implemented by using a shift deconvolution method.

3. The seismic waveform inversion method based on deep learning image registration according to claim 1, characterized in that: In step S104, the simulated earthquake recording image is used as a floating image, and the earthquake observation image is used as a fixed image, and deep learning image registration is performed to obtain the earthquake recording deformation field.

4. The seismic waveform inversion method based on deep learning image registration according to claim 3, characterized in that: Step S104 includes the following steps: Input the image to be registered, and output the deformation vector field after learning through a multi-scale parallel fully convolutional neural network; wherein the image to be registered includes the floating image and the fixed image; The floating image is based on the deformation vector field and uses a spatial transformation network to generate a deformable image; The difference between the fixed image and the deformed image is calculated according to the loss function, and the parameters of the earthquake record deformation field are updated, and multiple iterations are performed to obtain the final earthquake record deformation field.

5. The seismic waveform inversion method based on deep learning image registration according to claim 4, characterized in that: Using a three-layer multi-scale parallel downsampling module to reduce the image size and obtain a multi-scale downsampling feature map; The multi-scale downsampled feature map is sequentially passed through a pyramid dilated convolution module and an adaptive channel attention module for feature extraction and feature processing; The feature map processed by the adaptive channel attention module is input into the direction perception layer and the upsampling layer to obtain the deformation vector field of the input image pair.

6. The seismic waveform inversion method based on deep learning image registration according to claim 4, characterized in that: The loss function adopts the following formula: ， in, Represents a fixed image; Represents a floating image; Represents the deformation field of earthquake records ; express Grayscale similarity of the transformed images; Indicates the spatial smoothness coefficient; Indicates smoothing of image grayscale similarity.

7. The seismic waveform inversion method based on deep learning image registration according to claim 1, characterized in that: In step S106: The preset condition is: the normalized deformation field .

8. A seismic waveform inversion system based on deep learning image registration, characterized in that: include: A simulated seismic recording image acquisition module is used to perform least squares migration based on small offset observation data and background velocity field to construct a reflection coefficient model, and to perform Born forward modeling based on the reflection coefficient and the background velocity field to obtain a simulated seismic recording image; An earthquake record deformation field acquisition module is used to acquire the earthquake record deformation field by using deep learning image registration based on the simulated earthquake record image and the earthquake observation image; The seismic waveform inversion module is used to calculate the normalized deformation field using the following formula, ， in, represents the normalized deformation field; represents the earthquake record deformation field; Indicates the wavelength of the current data's main frequency wave, t indicates the time direction, and h indicates the space direction; The seismic waveform inversion module is further used to determine whether the normalized deformation field meets a preset condition; if so, the iterative process is terminated; otherwise, the background velocity field is updated through the velocity update module; The velocity update module is used to calculate the accompanying earthquake source according to the earthquake record deformation field, load the accompanying earthquake source into the wave equation for wave field propagation, and use cross-correlation to calculate the low wave number gradient component to update the background velocity field.

9. An electronic device, characterized in that: include: A processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the seismic waveform inversion method based on deep learning image registration as described in any one of claims 1 to 7.

Images