A multi-scale global information fusion neural network seismic data low-frequency reconstruction method

Abstract

The present application relates to a kind of multi-scale global information fusion neural network seismic data low frequency reconstruction method, solve the accurate low frequency reconstruction of any complex actual seismic record, improve the convergence of full waveform inversion problem.This method incorporates the multi-scale coding-decoding structure neural network of grid local sharing module and non-local attention module, while effectively extracting the local features of input data, fully consider the global information correlation of long-wavelength time series signal such as low frequency signal, and highlight the importance of global information for low frequency reconstruction, and combined with the simulation seismic data training set obtained on the public standard stratum velocity model, automatically learn the nonlinear mapping relationship between the missing low frequency seismic data and its corresponding low frequency data, realize the intelligent extrapolation of seismic data frequency band from high frequency to low frequency, accurately obtain the low frequency data carrying the real long-wavelength information of underground medium, effectively full waveform inversion velocity modeling accuracy.

Description

Technical Field

[0001] This invention belongs to the field of exploration geophysics and artificial intelligence technology, specifically relating to a low-frequency reconstruction method for exploration seismic data using a multi-scale global information fusion neural network. Background Technology

[0002] Oil and gas resources are important non-renewable resources, often referred to as the "lifeblood" of modern industry.

[0003] Seismic exploration is the most important means of oil and gas resource exploration and development. In recent years, due to factors such as the gradual depletion of easily detectable resources, the shift in exploration targets from shallow to medium-deep layers, unconventional oil and gas exploration, and complex exploration environments, actual seismic exploration is facing many problems, including high data acquisition costs, poor seismic data quality, and difficulties in seismic imaging and interpretation. Currently, exploring new low-cost, high-resolution, and reliable seismic acquisition technologies has become a research hotspot in recent years. At the same time, how to overcome the technical bottlenecks of exploration principles based on fixed physical models and assumptions, explore intelligent seismic data processing methods, and improve the signal-to-noise ratio and resolution of seismic data has also become one of the important technical issues of concern to the geophysicists both domestically and internationally in recent years.

[0004] The core of seismic exploration technology research is to improve the accuracy and identification capability of detecting interfaces between subsurface media and target geological bodies. Seismic exploration technology mainly uses artificially generated seismic waves and receives seismic signals through detectors deployed on the surface. It is an important method for delineating geological body interfaces and identifying oil and gas reservoirs. This method utilizes artificially generated elastic waves propagating in subsurface rock masses. When there is a difference in wave impedance between two strata, scattered waves are generated and transmitted back to the surface, where they are received by detectors. By analyzing the kinematic (travel time) and dynamic (amplitude, phase) characteristics of the seismic signals received by the detectors, the physical properties and structural information of the subsurface media are obtained, thereby achieving an accurate understanding of the geological conditions at different locations and depths underground.

[0005] In the seismic data processing workflow for exploration, velocity modeling of the subsurface medium is a crucial step, directly determining the accuracy of subsurface structural imaging and the interpretation of oil and gas reservoir locations. To date, methods for velocity modeling using exploration seismic data can be categorized into three types: velocity analysis, tomography, and full waveform inversion. Velocity analysis and tomography utilize only kinematic information from seismic data while neglecting dynamic information, resulting in low velocity modeling accuracy. They can only roughly construct the background velocity field, failing to capture detailed structures and subsurface interfaces. Full waveform inversion, on the other hand, utilizes all kinematic and dynamic information from seismic data and is a vital method for obtaining high-precision velocity distributions in the subsurface medium.

[0006] In the 1980s, Lailly (1983) and Tarantola (1984) first proposed the full waveform inversion method. This method utilizes all the kinematic and dynamic information of the seismic wavefield to construct a least-squares objective function. Within a local optimization framework, it continuously matches simulated and observed data to ultimately find a convergent distribution of physical parameters. They also proposed using the adjoint state method to calculate the gradient, avoiding the enormous computational burden of directly calculating the Fréchet derivative. Full waveform inversion is a local optimization problem based on the first-order Born weak scattering approximation, making it highly dependent on the accuracy of the initial model and the low-frequency information in the data. However, providing a good initial model for seismic data in complex geological structures or containing strong low-frequency noise interference is very difficult. Furthermore, because it is difficult to generate low-frequency signals during source excitation and detectors struggle to respond to low-frequency components in seismic signals, low-frequency signals are rarely present or have very weak energy in actual seismic data. The significant difference between the initial model and the actual distribution of physical property parameters, as well as the periodic jump problem caused by the lack of low-frequency information, are the most critical factors affecting the convergence of the full waveform inversion.

[0007] To address this issue, the concept of multi-scale inversion is employed, decomposing seismic data into different scales through various transformations for scale-by-scale inversion. This allows for high-precision modeling of subsurface physical parameters even when the initial model is inaccurate. Claerbout and Jannanne et al. pointed out that the background field information of the subsurface medium is related to the low-frequency components in the field seismic data, while the information of the perturbation field (wave impedance) corresponds to the high-frequency components of the seismic data. Bunks proposed using a low-pass filter to decompose seismic data into different frequency scales. Employing an inversion strategy from low-frequency to high-frequency bands in the time domain can improve the convergence of the full waveform inversion objective function, reduce the occurrence of local extrema, and alleviate the period jump problem. Sirgue and Pratt conducted in-depth research on the multi-scale characteristics of frequency domain full waveform inversion and provided specific frequency selection strategies. Reconstructing low-frequency signals reflecting the large-scale structure of the subsurface medium from observational data that does not contain low frequencies to provide a better initial model for full waveform inversion is an important research direction in full waveform inversion research. Shin and Cha transformed seismic data into the Laplace domain and obtained data containing rich low-frequency information through exponential decay. Some researchers proposed that the envelope of seismic data can reflect the large-scale structural information of the subsurface medium. Seismic data with the envelope obtained through Hilbert transform contains rich low-frequency information. Chi et al. proposed an objective function based on the L2 norm of the envelope residual and derived its gradient using the adjoint state method. In numerical tests, they successfully inverted the large-scale structure of the trial model and compared it with objective functions such as envelope square and envelope logarithm. Wu et al. proposed a modulation-convolution model to invert the large-scale structure of the subsurface medium from the envelope signal of seismic data, giving physical meaning to envelope inversion. Li et al. developed beat theory, using multiplication of single-frequency signals to generate low-frequency signals, and extracted the low-frequency signals through nonlinear filtering to achieve accurate inversion of the large-scale structure of the subsurface medium. Zhang et al., based on the convolution model of seismic data, proposed using a sparse blind deconvolution algorithm combined with artificially designed low-frequency wavelets to construct observational data containing low-frequency information. Liu et al. proposed that signal intensity can decompose the original data into high-frequency and low-frequency bands, constructed an objective function based on the signal intensity residual, and successfully inverted the large-scale structure of the subsurface medium. Phase information in seismic data exhibits a better linear correspondence with subsurface velocity models. Sun and Schuster proposed a time-domain phase inversion objective function, demonstrating that phase inversion methods can achieve good inversion results even with poor initial velocity models. Choi and Alkhalifah introduced the concept of a phase derivative objective function in the frequency domain, obtaining phase information by calculating the derivative of the angular frequency of seismic data. Alkhalifah and Choi analyzed the strong nonlinearity of the objective function in detail, pointing out that unwrapping the phase can reduce the nonlinearity of the full-waveform inversion objective function and decrease the dependence of full-waveform inversion on the velocity model. The period jump problem can also be intuitively explained and solved by improving waveform matching. Warner and Guasch constructed an L2 norm objective function using the residuals of simulated data after Wiener filtering and original observation data, aiming to continuously update the Wiener filter operator so that accurate modeling of subsurface parameters can be achieved when the filter operator is gradually updated to a Delta function; Zhu and Fomel proposed an objective function based on the phase difference of an adaptive matched filter; Wang et al. proposed a dynamic time warping method from the field of signal recognition to successfully alleviate the periodic jump problem; Hu Yong et al. improved the waveform matching degree by reducing the phase difference between simulated and observed data through time-frequency domain phase correction, thus guiding the initial model to update in the correct direction.

[0008] Besides the two methods mentioned above for solving the periodic jump problem, there are many other approaches. In practical exploration, for situations where the offset is not large enough, researchers have proposed a full-waveform reflection inversion method. Using global optimization algorithms or statistical methods for full-waveform inversion can prevent the objective function from converging to local minima. By utilizing optimal transport theory, the waveform matching problem is treated as an optimal strategy problem for transporting objects of different masses to different locations, which can effectively address the high accuracy requirement of the initial model in full-waveform inversion. In addition, the very popular deep learning technology in recent years has also been widely applied to full-waveform inversion for initial model building.

[0009] However, many methods for reducing the nonlinearity of full-waveform inversion and improving convergence accuracy at the algorithmic level, such as objective function construction and wavefield separation, have limited applicability and require high accuracy from actual exploration seismic data. Furthermore, traditional low-frequency reconstruction methods yield low-frequency information that differs from the missing true low-frequency components in the actual data, potentially leading to false anomalies in the inversion results. Therefore, a crucial technological approach is urgently needed to directly recover accurate low-frequency signals carrying true subsurface long-wavelength information. This approach combines big data-driven artificial intelligence neural network methods to find the nonlinear mapping between missing low-frequency observational seismic data and true low-frequency data from a complete training set. This would enable accurate low-frequency reconstruction of any complex actual seismic record and improve the convergence of full-waveform inversion. Summary of the Invention

[0010] The purpose of this invention is to provide a low-frequency reconstruction method for exploration seismic data using a multi-scale global information fusion neural network, in order to solve the problems of accurate low-frequency reconstruction of any complex actual seismic record and improve the convergence of full waveform inversion.

[0011] This invention proposes a neural network with a multi-scale encoding and decoding structure that integrates a grid-based local sharing module and a non-local attention module. While effectively extracting local features of the input data, it fully considers the global information correlation of long-wavelength time-series signals such as low-frequency signals, and emphasizes the importance of global information for low-frequency reconstruction. Combined with a training set of simulated seismic data obtained from a publicly available standard formation velocity model, it automatically learns the nonlinear mapping relationship between missing low-frequency seismic data and its corresponding low-frequency data, realizing intelligent extrapolation of seismic data frequency bands from high frequency to low frequency. This accurately obtains low-frequency data carrying real long-wavelength information of the subsurface medium, effectively improving the accuracy of full-waveform inversion velocity modeling.

[0012] The objective of this invention is achieved through the following technical solution:

[0013] A low-frequency reconstruction method for seismic data using a multi-scale global information fusion neural network includes the following steps:

[0014] a. Design the input and output data formats for neural network training; the input data format for the neural network is a 1-dimensional time-series signal, and the output data format is a 1-dimensional time-series signal.

[0015] b. Design a grid-based local sharing module;

[0016] c. Design a non-local attention module;

[0017] d. Design multi-scale network structures;

[0018] e. Training dataset construction;

[0019] f. Test set construction: Using the standard Marmousi model and setting the same observation system, wavelet and forward modeling algorithm parameters as in step e, perform forward modeling simulation to obtain the test dataset;

[0020] g. Employ a loss function based on a multi-scale global information fusion neural network using mean square error;

[0021] h. Initialize network parameters, determine the learning rate, learning rate decay rate, number of input samples per cycle, and number of learning iterations;

[0022] i. Input the training set data into the multi-scale global information fusion neural network to train the network parameters, adjust the hyperparameters and data volume, and perform multiple training sessions. Compare and analyze the accuracy of the low-frequency reconstruction results with the output results and labels using single-channel waveform, single-channel amplitude spectrum, single-channel phase spectrum, and single-shot FK spectrum. Take the network model parameters with the smallest loss function and the highest low-frequency data reconstruction accuracy as the final training result.

[0023] J. Input the test set into the trained multi-scale global information fusion neural network model to perform low-frequency data reconstruction test, and perform speed inversion of the output data using the Marmousi model, comparing the accuracy of the inversion with the real model.

[0024] k. Set the network layer parameters output by the original scale nonlocal attention module in the trained multi-scale global information fusion neural network model to be adjustable, and freeze the network layer parameters before this module. Use a small amount of actual exploration seismic data containing effective low-frequency components for transfer learning training, and then apply it to a large amount of actual seismic data without low-frequency components.

[0025] Further, step b specifically includes the following steps:

[0026] b1. Design a grid of a certain length, and distribute the feature map of the input module in 3 grids without overlap. That is, the data in the grid is 1 / 3 of the length of the input data, and the data in each grid is continuous. The input data is folded into 3 channels on the time axis through the grid splitting operation.

[0027] b2. Perform convolutional feature extraction along the channel direction to enhance the relevance of global features;

[0028] b3. Concatenate the feature maps of the 3-channel data to restore them to the same dimensions as the input data of this module.

[0029] Further, step c specifically includes the following steps:

[0030] c1. Perform three linear mappings on the feature map input to this module to compress the number of channels and obtain three feature maps: Q, K, and V.

[0031] c2. Multiply the transpose of Q by K and output the autocorrelation coefficient matrix through the Softmax function. Then multiply the autocorrelation coefficient matrix by V to obtain the result after global attention enhancement.

[0032] c3. By adjusting the number of channels through linear mapping and adding it to the original input data, the residual connection effect is achieved, and the final module output data is obtained.

[0033] Further, step d specifically includes the following steps:

[0034] d1. The original input signal is downsampled by 1 / 2 four times to obtain four low-scale input data. The original scale signal and the four low-scale signals are input into the network respectively and processed through three consecutive 1D dilated convolutional layers for local feature extraction. The lower the scale of the input data, the larger the receptive field of the extracted features after convolution.

[0035] d2. Multi-scale feature fusion is performed by progressively passing the local features extracted at each scale down from high scale to low scale.

[0036] After fusing the feature maps input from the previous scale, the d3 and 4 low-scale layers incorporate global relevance information through a non-local attention module. The resulting feature maps are then concatenated with the previous scale and followed by a non-local attention module to highlight global feature information.

[0037] d4. Data fusion is performed at the original scale layer, and then the number of data channels is adjusted through multiple convolutional layers and a linear transformation layer to obtain the final output data.

[0038] Further, step e specifically includes the following steps:

[0039] e1. The training dataset is generated using the classic Marmousi formation velocity model in the field of exploration seismicity.

[0040] e2. Using a window approximately 1 / 9 the size of the standard Marmousi model, extract 17 formation velocity models, including 9 simple structural sub-models and 8 complex structural sub-models. Interpolate these 17 velocity models to the same size as the standard Marmousi model and add a water layer above each model.

[0041] e3. Perform forward modeling based on the finite difference method of the acoustic wave equation using both complete frequency band and missing low-frequency Ricker wavelets to obtain training datasets. The seismic data obtained by forward modeling using missing low-frequency Ricker wavelets is used as input data. The seismic data obtained by forward modeling using complete frequency band Ricker wavelets is low-pass filtered to obtain labels.

[0042] e4. Collect a small amount of actual exploration seismic data with effective low-frequency components and a large amount of missing low-frequency components.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] 1. Compared with the classic Unet neural network in the field of deep learning, this invention focuses more on global features for the low-frequency components of time-series seismic signal reconstruction. It designs a multi-scale encoding and decoding structure to realize multi-scale feature extraction and feature fusion, and proposes to combine multi-layer one-dimensional dilated convolution to improve the receptive field of feature extraction.

[0045] 2. A grid-based local sharing module and a non-local attention module were proposed and designed to further enhance the effective extraction of contextual information of time-series signals, enhance the global correlation of extracted feature maps, and at the same time eliminate the interference of useless redundant information on low-frequency reconstruction information.

[0046] 3. A transfer learning method using a small amount of actual exploration seismic data containing effective low-frequency components is adopted. After the neural network has learned the underlying nonlinear mapping relationship from high-frequency data to low-frequency data using simulated data, the network's learning ability is further refined to the features of actual exploration data. This enables accurate low-frequency reconstruction of complex actual exploration seismic data, making the data features learned by the network closer to the actual data and improving the generalization of network applications. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 Mesh-based local shared module structure diagram;

[0049] Figure 2 Non-local attention module structure diagram;

[0050] Figure 3 Multi-scale global information fusion neural network structure diagram;

[0051] Figure 4a Selection of a simple velocity model for training set generation;

[0052] Figure 4b Selection of complex velocity models for training set generation;

[0053] Figure 4c The time-domain waveform of the Ricker wavelet used to generate input data for the training set;

[0054] Figure 4d Used to generate the Rack wavelet spectrum of the training set input data;

[0055] Figure 4e The time-domain waveform of the Rack wavelet used to generate training set labels;

[0056] Figure 4f The Lake wavelet spectrum used to generate training set labels;

[0057] Figure 5a Comparison of single-channel prediction results on the test set with the time-domain waveforms of the labels;

[0058] Figure 5b Comparison of single-channel prediction results on the test set with the tag amplitude spectrum;

[0059] Figure 5c Comparison of single-channel prediction results on the test set with the label phase spectrum;

[0060] Figure 5d Comparison of single-shot prediction results on the test set with those in the label time domain;

[0061] Figure 5e Comparison of single-shot input data, prediction results, and labeled FK spectra in the test set;

[0062] Figure 6a Marmousi model inversion results for low-frequency missing data;

[0063] Figure 6b The results of the Marmousi model inversion were obtained using the predicted low-frequency data. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0065] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0066] The multi-scale global information fusion neural network seismic data low-frequency reconstruction method of this invention is implemented on the Anaconda platform of Python language compiler. The program is written in the PyTorch machine learning library framework. According to the corresponding computational requirements, a CUDA parallel computing architecture is built, and the graphics card used is RTX 4090.

[0067] The core of this method is to propose a low-frequency reconstruction method for exploration seismic data based on a multi-scale global information fusion neural network. During actual seismic exploration data acquisition, the high frequency of the source-excited signal lacks sufficient low-frequency components, the detector's response to low-frequency signals is weak, and the prevalent low-frequency noise contaminates the weak low-frequency information. This results in insufficient low-frequency information in the acquired field seismic data, which is crucial for the correct convergence of full-waveform inversion velocity modeling. Therefore, this invention constructs an artificial neural network structure suitable for multi-scale information fusion of frequency band extrapolation of exploration seismic data. It designs a grid-based local sharing module and a non-global attention module that can effectively extract global correlation features of the data. It fully explores the nonlinear mapping relationship between high-frequency and low-frequency data in the full-band data, improves the attention to effective feature information while discarding redundant noise information. It uses the publicly available standard model of exploration geophysics, the Marmousi formation velocity model, to construct training and test sets. It adopts a transfer learning mode of supervised pre-training plus fine-tuning of parameters with actual seismic data. The network parameters are trained multiple times and the effect is verified using the test set. The training with the highest prediction accuracy of actual data is used as the final model parameters.

[0068] The present invention provides a low-frequency reconstruction method for seismic data using a multi-scale global information fusion neural network, comprising the following steps:

[0069] a. Design the input and output data formats for neural network training. Input data d in and output data d outThe dimension can be represented as [C,L,W], where C represents the number of channels, L represents the time length of a 1D time series signal, and W represents the number of 1D time series signals in a sample. Taking into account both the computational efficiency and accuracy of low-frequency reconstruction of seismic data, the data form of the input neural network used in this invention is a single-channel 1D time series signal, and the output data form is a single-channel 1D time series signal. That is, the dimension of the input and output data can be represented as [B,1,L,1].

[0070] b. Design a grid-based local sharing module. Based on the dimension [C, L, 1] of the input feature map M, design a grid of length L / 3. Distribute the input data non-overlappingly across the three grids, meaning the data within each grid is L / 3 of the input data length, and the data within each grid is continuous. The input data is folded into 3 channels along the time axis through grid splitting, at which point the data dimension becomes [3C, L / 3, 1]. Use C 3×1 convolutional kernels in the channel direction to perform convolutional feature extraction, enhancing the relevance of global features. Finally, stitch the feature maps of the 3 channels together to restore the same dimension [C, L, 1] as the input data of this module, as shown below. Figure 1 As shown;

[0071] c. Design a non-local attention module. The input feature map M of dimension [C,L,1] is processed by C / 2 1×1 convolutional kernels using 3 linear mappings to compress the number of channels, resulting in three feature maps Q, K, and V of dimension [C / 2,L,1].

[0072]

[0073] Among them, W Q W K and W V Let x represent the coefficients of the linear transformation corresponding to a single 1×1 convolution kernel. i and x j Let Q, K, and V represent global information and the currently focused location information, respectively. Flattening Q, K, and V (merging the last two dimensions of the feature map) results in Q, K, and V data dimensions of [C / 2, L×1]. Channel rearrangement of K and V results in K and V data dimensions of [L×1, C / 2]. Multiplying Q and K yields a feature map F with dimensions [L×1, L×1]. Applying the Softmax function to F results in a feature map FL×1, L×1 with dimensions [L×1, L×1]. V The meaning of this feature map is the same as that of the autocorrelation coefficient matrix; multiplying Fv and V yields a feature map A with dimensions [L×1, C / 2]. The calculation process of A can be expressed as follows:

[0074] Where T represents transpose; A is transformed into [C / 2, L×1] by channel re-transformation, and the second dimension is expanded into two dimensions [C / 2, L, 1]. These are then processed by C 1×1 convolutional kernels to transform the dimensions back to [C, L, 1]. Finally, the input feature map X is added to A to obtain the output feature map of the module, as shown below. Figure 2 As shown.

[0075] d. Design a multi-scale network structure. The original input signal d with dimension [C,L,1]... in To avoid information loss during downsampling, this invention designs a channel-extended downsampling method that separates odd and even sampling points, specifically tailored to the characteristics of time-series 1D seismic signals. Specifically, during downsampling, odd and even sampling points are extracted separately and then concatenated along the channel dimension. Therefore, the four low-scale input data obtained after downsampling have dimensions of [2C,L / 2,1], [4C,L / 4,1], [8C,L / 8,1], and [16C,L / 16,1], respectively. and The original scale signal and four low-scale signals are input into the network and processed through three consecutive 5×1 dilated convolutional layers for local feature extraction. The dilation rates of the three layers are 1, 2, and 3, respectively, with the lower-scale input data having a larger receptive field after convolution. Then, the local features extracted from each scale are progressively passed down from high to low scales for multi-scale feature fusion. After fusing the feature maps from the previous scale, the four low-scale layers undergo a non-local attention module to incorporate global correlation information. The resulting feature maps are then channel-fused with the previous scale, followed by another non-local attention module to highlight global feature information. Finally, the data is fused at the original scale layer, and the number of data channels is adjusted through three 3×1 convolutional kernel layers and one 1×1 convolutional kernel linear transformation layer to obtain the final output data d. out ,like Figure 3 As shown.

[0076] e. Training Dataset Construction. To make the training dataset closer to the characteristics of real exploration seismic data, thereby enhancing the generalization of the neural network training results, this invention uses the classic standard Marmousi formation velocity model in the field of exploration seismicity to generate the training dataset, with a grid size of 128×384; using a window with a grid size of 30×100, nine simple models are uniformly and non-overlappingly extracted from the standard Marmousi model ( Figure 4a Eight typical complex models containing special geological bodies such as faults, low-velocity zones, anticlines, and strong scattering bodies were artificially selected. Figure 4bThe extracted 17 model dimensions were linearly interpolated to the same grid size as the standard Marmousi model, and a water layer of 5 grid thickness was added above each model. The observation system was set up with each grid point on the water surface acting as a geophone, resulting in 384 seismic records. Each model had 96 sources evenly distributed on the water surface, resulting in 96 single-shot records. The forward modeling of the simulated seismic records was based on the two-dimensional acoustic medium velocity-displacement wave equation.

[0077]

[0078] Where v represents the velocity of sound, u represents the wave field value, t represents the wave field propagation time, K represents the bulk modulus, ρ represents the medium density, and x and z represent the coordinates in the horizontal and depth directions, respectively; the forward modeling algorithm uses a finite difference method with 10th-order accuracy in staggered grid space and 4th-order accuracy in time, combined with PML boundary conditions of 30 layers; the grid spacing is 20m, the sampling interval is 0.002s, and the total sampling time is 4.8s; the wavelet used to generate the input data is the Ricker wavelet with a 7Hz main frequency and a 5Hz truncation frequency high-pass filter, and its spectrum, as shown in the figure. Figure 4c , Figure 4d As shown, the generated tag uses a 7Hz main frequency full-band Ricker wavelet and its spectrum, as follows. Figure 4e , Figure 4f As shown, the forward-modeled seismic records are low-pass filtered with a 5Hz cutoff frequency; a small amount of actual exploration seismic data with effective low-frequency components and a large amount of missing low-frequency components are collected to provide data for transfer learning.

[0079] f. Test set construction. Using the standard Marmousi model and setting the same observation system, wavelet, and forward modeling algorithm parameters as in step e, perform forward modeling simulations to obtain the test dataset.

[0080] g. Employ a loss function based on the multi-scale global information fusion neural network using mean square error. This loss function J MSE It can be represented as;

[0081]

[0082] Where N represents the total number of input data; y represents the label value; y' represents the input data; k represents the sequence number of the input data and the corresponding label, and its value range is all integers in the interval [1, N];

[0083] h. Initialize network parameters with a learning rate of 0.001 and a learning rate decay rate of 0.01. Set the number of training samples to 32 for each input and perform 1000 training iterations. Use the Adam optimizer to iteratively update the network parameters using gradient descent.

[0084] i. Input the training set data into the multi-scale global information fusion neural network to train the network parameters, adjust the hyperparameters and data volume, and perform multiple training sessions. Compare and analyze the accuracy of the low-frequency reconstruction results with the output results and labels using single-channel waveform, single-channel amplitude spectrum, single-channel phase spectrum, and single-shot FK spectrum. Take the network model parameters with the smallest loss function and the highest low-frequency data reconstruction accuracy as the final training result.

[0085] j. Input the test set into the trained multi-scale global information fusion neural network model for low-frequency data reconstruction testing, and perform speed inversion using the standard Marmousi model on the output data. Compare the difference between the real model and the inversion results by calculating the error coefficients to evaluate the inversion accuracy.

[0086]

[0087] in, V represents the error coefficient. inv and v t Let i and j represent the initial model and the real model, respectively. Let i and j represent the coordinates of the grid points in the model in the x and z directions, respectively. Let X and Z represent the total number of grid points in the x and z directions of the model, respectively.

[0088] k. Set the network layer parameters output by the original scale nonlocal attention module in the trained multi-scale global information fusion neural network model to be adjustable, and freeze the network layer parameters before this module. Use a small amount of actual exploration seismic data containing effective low-frequency components for transfer learning training, and then apply it to a large amount of actual seismic data without low-frequency components.

[0089] Example 1

[0090] This embodiment provides an application of a low-frequency reconstruction method for exploration seismic data lacking effective low-frequency components, based on a multi-scale global information fusion neural network, for low-frequency compensation. The details are as follows:

[0091] Table 1 Test Set Data Parameters

[0092] velocity model Number of cannons Number of tracks Sampling time Sampling interval Input data frequency band Tag band Standard Marmousi 96 384 4.8s 0.002s 5~100Hz 0~5Hz

[0093] The test set data comes from forward modeling of the standard Marmousi model, consisting of 96 single-shot records, 384 tracks per shot, a total sampling time of 4.8 s, a sampling interval of 0.002 s, and an input data bandwidth of 5–100 Hz. The input data is fed into a trained multi-scale global information fusion neural network to obtain low-frequency output results. Figure 5a , 5bThe output data shown in Figure 5c highly overlaps with the tag's waveform in the time domain, amplitude spectrum in the 0-5Hz low-frequency band, and phase spectrum in the 0-5Hz low-frequency band. The predicted value has a very small error with the tag, and the prediction accuracy is high. It can accurately recover the corresponding low-frequency signal using high-frequency band data that is missing low frequencies. Figure 5d and Figure 5e The single-shot prediction results shown are highly similar to those of the single-shot label in the two-dimensional data and FK spectrum in the time domain, and have good lateral continuity. This proves that the output obtained by inputting a one-dimensional time-series signal can maintain good lateral continuity, so that the effective low-frequency information can be accurately recovered. Figure 6a The results of the standard Marmousi model inversion using missing low-frequency data show obvious false anomalies in the shallow layer. This is because the lack of low-frequency components in the data leads to periodic jumps, resulting in incorrect gradient updates. Figure 6b As shown, the inversion results based on the standard Marmousi model for predicting low frequencies are accurate and basically consistent with the real model, with no false anomalies. The calculated error coefficient p between the inversion results and the real model is extremely small at 0.52%, proving that the low-frequency information predicted using the method proposed in this invention is true and accurate.

[0094] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A low-frequency reconstruction method for seismic data using a multi-scale global information fusion neural network, characterized in that, Includes the following steps: a. Design the input and output data formats for neural network training; the input data format for the neural network is a 1-dimensional time-series signal, and the output data format is a 1-dimensional time-series signal. b. Design a grid-based local sharing module; c. Design a non-local attention module; d. Design multi-scale network structures; e. Training dataset construction; f. Test set construction: Using the standard Marmousi model and setting the same observation system, wavelet and forward modeling algorithm parameters as in step e, perform forward modeling simulation to obtain the test dataset; g. Employ a loss function based on a multi-scale global information fusion neural network using mean square error; h. Initialize network parameters, determine the learning rate, learning rate decay rate, number of input samples per cycle, and number of learning iterations; i. Input the training set data into the multi-scale global information fusion neural network to train the network parameters, adjust the hyperparameters and data volume, and perform multiple training sessions. Compare and analyze the accuracy of the low-frequency reconstruction results with the output results and labels using single-channel waveform, single-channel amplitude spectrum, single-channel phase spectrum, and single-shot FK spectrum. Take the network model parameters with the smallest loss function and the highest low-frequency data reconstruction accuracy as the final training result. J. Input the test set into the trained multi-scale global information fusion neural network model to perform low-frequency data reconstruction test, and perform speed inversion of the output data using the Marmousi model, comparing the accuracy of the inversion with the real model. k. Set the network layer parameters output by the original scale nonlocal attention module in the trained multi-scale global information fusion neural network model to be adjustable, and freeze the network layer parameters before this module. Use a small amount of actual exploration seismic data containing effective low-frequency components for transfer learning training, and then apply it to a large amount of actual seismic data without low-frequency components.

2. The low-frequency reconstruction method for seismic data using a multi-scale global information fusion neural network according to claim 1, characterized in that, Step b specifically includes the following steps: b1. Design a grid of a certain length, and distribute the feature map of the input module in 3 grids without overlap. That is, the data in the grid is 1 / 3 of the length of the input data, and the data in each grid is continuous. The input data is folded into 3 channels on the time axis through the grid splitting operation. b2. Perform convolutional feature extraction along the channel direction to enhance the relevance of global features; b3. Concatenate the feature maps of the 3-channel data to restore them to the same dimensions as the input data of this module.

3. The low-frequency reconstruction method for seismic data using a multi-scale global information fusion neural network according to claim 1, characterized in that, Step c specifically includes the following steps: c1. Perform three linear mappings on the feature map input to this module to compress the number of channels and obtain three feature maps: Q, K, and V. c2. Multiply the transpose of Q by K and output the autocorrelation coefficient matrix through the Softmax function. Then multiply the autocorrelation coefficient matrix by V to obtain the result after global attention enhancement. c3. By adjusting the number of channels through linear mapping and adding it to the original input data, the residual connection effect is achieved, and the final module output data is obtained.

4. The low-frequency reconstruction method for multi-scale global information fusion neural network seismic data according to claim 1, characterized in that... Step d specifically includes the following steps: d1. The original input signal is downsampled by 1 / 2 four times to obtain four low-scale input data. The original scale signal and the four low-scale signals are input into the network respectively and processed through three consecutive 1D dilated convolutional layers for local feature extraction. The lower the scale of the input data, the larger the receptive field of the extracted features after convolution. d2. Multi-scale feature fusion is performed by progressively passing the local features extracted at each scale down from high scale to low scale. After fusing the feature maps input from the previous scale, the d3 and 4 low-scale layers incorporate global relevance information through a non-local attention module. The resulting feature maps are then concatenated with the previous scale and followed by a non-local attention module to highlight global feature information. d4. Data fusion is performed at the original scale layer, and then the number of data channels is adjusted through multiple convolutional layers and a linear transformation layer to obtain the final output data.

5. The low-frequency reconstruction method for seismic data using a multi-scale global information fusion neural network according to claim 1, characterized in that, Step e specifically includes the following steps: e1. The training dataset is generated using the classic Marmousi formation velocity model in the field of exploration seismicity. e2. Using a window approximately 1 / 9 the size of the standard Marmousi model, extract 17 formation velocity models, including 9 simple structural sub-models and 8 complex structural sub-models. Interpolate these 17 velocity models to the same size as the standard Marmousi model and add a water layer above each model. e3. Perform forward modeling based on the finite difference method of the acoustic wave equation using both complete frequency band and missing low-frequency Ricker wavelets to obtain training datasets. The seismic data obtained by forward modeling using missing low-frequency Ricker wavelets is used as input data. The seismic data obtained by forward modeling using complete frequency band Ricker wavelets is low-pass filtered to obtain labels. e4. Collect a small amount of actual exploration seismic data with effective low-frequency components and a large amount of missing low-frequency components.

Images