Self-supervised deep learning seismic data low frequency reconstruction method and system

Abstract

The application discloses a self-supervised deep learning seismic data low-frequency reconstruction method and system, extracts multi-source information of different frequency bands from observed seismic data, and constructs a network training data set after preprocessing; a double-channel U-Net network is built based on a Pytorch platform, a self-supervised learning task is designed to train the network; finally, the seismic data with missing low-frequency data is resampled and energy balanced, and then input into the trained network, and then the low-frequency information of the seismic data is reconstructed. The application can reconstruct the missing low-frequency seismic data information from the medium-high frequency seismic data. Since the low-frequency data is less sensitive to the cycle skipping problem than the high-frequency data, the proposed method can overcome the cycle skipping problem and provide a good data basis for traditional full waveform inversion, thereby further improving the inversion accuracy of the final velocity model.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration technology, specifically relating to a self-supervised deep learning method and system for low-frequency reconstruction of seismic data. Background Technology

[0002] Oil and natural gas are among the most important fossil fuels, closely related not only to national political stability and economic development but also to people's daily lives. In today's rapidly developing world, oil and natural gas resources remain of significant strategic importance, making the exploration of effective exploration and efficient development methods and advanced technologies crucial. Seismic exploration is the most important method in oil and gas exploration, characterizing the structural features of subsurface geological bodies by analyzing the propagation patterns of excited seismic waves in the subsurface medium. Specifically, seismic waves propagate through different strata, and the reflected or refracted waves received at the surface carry subsurface information. By processing and analyzing the changes in the propagation time, amplitude, phase, and frequency of seismic waves, subsurface morphology and structures can be determined. Analyzing the detected geological information to obtain the physical properties of the subsurface medium and subsequently assessing the oil and gas content of reservoirs is the core objective of seismic exploration for oil and gas. In seismic exploration, seismic inversion, especially full-waveform inversion, can establish lithological and physical property models of the subsurface medium, thereby guiding oil and gas content prediction, and has become a focus of scholarly attention.

[0003] In seismic inversion, it is often impossible to directly obtain parameters such as P-wave velocity, S-wave velocity, density, and quality factor of the subsurface medium from seismic data through closed-form solutions. A common approach is to estimate these parameters by setting an objective function and using iterative optimization. Full-waveform inversion, established within this framework, theoretically yields a velocity model with half the resolution of the minimum wavelength, making it the most accurate seismic inversion method currently available, as it fully utilizes all the information carried by seismic data. However, full-waveform inversion using seismic data faces the local extremum problem caused by cycle-skipping, preventing the inversion from converging to the global extremum. Research indicates that providing an accurate initial model or seismic data with sufficient low-frequency information can effectively overcome the cycle-skipping problem. During the inversion process, low-frequency information can be used first for inversion, and the result can be used as the initial model for full-waveform inversion of mid-to-high-frequency data, thus overcoming the cycle-skipping problem. Therefore, sufficient low-frequency seismic data is crucial for full-waveform inversion.

[0004] In practice, due to unavoidable factors such as exploration equipment and noise, actual seismic data often lacks sufficient low-frequency information (below 5Hz). Accurately reconstructing the low-frequency information of seismic data is a crucial step in full waveform inversion. To obtain seismic data with sufficient low-frequency information and overcome the cycle skipping problem, some works have been proposed. For example, the possibility of obtaining low-frequency data from high-frequency data through computation has been explored. Deep learning algorithms, such as Generative Adversarial Networks (GANs) and U-Net networks, have been used for bandwidth extension, and the predicted data is used to supplement the full waveform inversion data. However, all of these methods directly extrapolate low-frequency information from mid-to-high-frequency information, failing to fully utilize the slowly changing signal information reflected in the envelope, and there is still room for improvement in extrapolation accuracy. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a self-supervised deep learning method and system for low-frequency reconstruction of seismic data. This method extracts multi-source information (taking envelope information as an example) from seismic data of different frequency bands based on observed seismic data, constructs a self-supervised learning task to avoid dependence on target frequency band label data, and uses a dual-channel U-Net network to reconstruct missing low-frequency information from effective mid-to-high frequency band information using a self-supervised deep learning method. This helps to recover a more accurate long-wavelength velocity model, which serves as the initial velocity model for traditional full waveform inversion (FWI) to solve the technical problem of cycle-skipping.

[0006] The present invention adopts the following technical solution:

[0007] A self-supervised deep learning method for low-frequency reconstruction of seismic data is proposed. This method acquires seismic data of different frequency bands from observed seismic data, extracts multi-source information from the seismic data, and obtains training and test set data after time downsampling and energy equalization. A dual-channel U-Net network is constructed, and a low-frequency data reconstruction task on the training set data is used as a self-supervised learning task to train the constructed dual-channel U-Net network. The test set data containing multi-source information is then fed into the trained dual-channel U-Net network to reconstruct the missing low-frequency seismic data.

[0008] Specifically, after time downsampling and energy equalization, the training set data and test set data are as follows:

[0009] Design zero-phase digital filters for the frequency bands corresponding to Band1, Band2, and Band3 involved in the training and test data.

[0010] Based on the designed zero-phase digital filters corresponding to the Band1, Band2, and Band3 frequency bands, the seismic data u of the Band1, Band2, and Band3 frequency bands are obtained.Band1 (t), u Band2 (t), u Band3 (t); applying an envelope operator to the seismic data u Band(n) (t) to obtain corresponding envelope information G uBand(n) (t), n=1, 2, 3, and constructing a data set for a deep learning network to reconstruct low-frequency seismic data;

[0011] performing time-domain downsampling and energy equalization processing on the seismic multi-source information of three frequency bands to obtain U Band1 (t), U Band2 (t), U Band3 (t), G UBand1 (t), G UBand3 (t), and constructing training set data and test set data for two-channel U-Net network training.

[0012] Further, the frequency band ranges of the Band1, Band2 and Band3 frequency band data are:

[0013]

[0014] wherein, Band1 is the input frequency band of the network training set and the verification set, and Band2 is the target frequency band of the network training set and the verification set; Band3 is the input frequency band of the test set.

[0015] Further, the envelope operator is defined as:

[0016]

[0017] wherein, u(t) is a one-dimensional time signal, G(u(t)) is the envelope operator of the signal u(t), and H(·) is a Hilbert operator.

[0018] Further, the performing of time-domain downsampling and energy equalization processing on the seismic multi-source information of three frequency bands is specifically as follows:

[0019] the input frequency band of the training set is [f1,f2], the target frequency band is [nf1,f2], 0<n<1, and the input frequency band of the test set data is [nf1,nf2];

[0020] performing m-fold downsampling in the time domain on the multi-source information, m>1, and in the time direction, the original N t sampling points become points, so as to obtain shot data of size; removing five traces of data on the left and right of the shot point and performing energy equalization on each trace of data in proportion to the distance from the shot point, so as to obtain training set data of size;

[0021] The multi-source information of seismic data in Band 3 is reprocessed with m×n times time downsampling and energy equalization,

[0022] For the seismic data with the input frequency band [nf1,nf2], 0<n<1 in the test set, m×n times time downsampling is performed to keep the number of waveform samples per sampling point consistent with that of the training set data, and the obtained dimension is data per shot; after zero-padding the data of each shot to the dimension of , five traces of data on the left and right of each shot point are removed, that is, five traces of data on the left and right of each shot point are assigned a value of 0, and then energy equalization is performed on each trace of data in proportion to the distance from the shot point to obtain test set data.

[0023] Further, the data after energy equalization is:

[0024] d b,i =d i ·0.01·P s,j -P r,i

[0025] Wherein, d i is the seismic data corresponding to each shot, P s,j is the grid position corresponding to the j-th shot point, and P r,i is the grid position corresponding to the i-th trace data.

[0026] Specifically, the dual-channel U-Net network comprises 15 convolution layers, 3 maximum pooling layers, 3 transposed convolution layers and 3 concatenation layers, the activation function adopts ReLU, seismic data is input as one channel of the dual-channel U-Net network, after passing through 2 convolution layers, it is concatenated with the output of the U-Net network for seismic data, and then the final output is obtained after passing through 3 convolution layers.

[0027] Specifically, during the training process of the constructed dual-channel U-Net network, one channel input is mid-high frequency band seismic data U Band1 (t), and the other channel input is the envelope of mid-high frequency band seismic data the network output is low frequency band seismic data U Band2 (t).

[0028] Further, the objective function for dual-channel U-Net network training is:

[0029]

[0030] Wherein, N is the number of samples in the network training set, is the input effective frequency band data of the training set, is the input effective frequency band envelope data of the training set, J represents the target frequency band seismic data in the training set, and J is the objective function for training the dual-channel U-Net network.

[0031] The Adam optimizer is used to iteratively update the network parameters θ with a step size η = 0.001.

[0032]

[0033] Where, θ i This represents the parameters of the network in the i-th iteration.

[0034] Secondly, embodiments of the present invention provide a self-supervised deep learning seismic data low-frequency reconstruction system, comprising:

[0035] The data module acquires seismic data of different frequency bands from observed seismic data, extracts multi-source information from the seismic data, and obtains training set data and test set data after time downsampling and energy equalization;

[0036] The network module constructs a dual-channel U-Net network and uses a low-frequency data reconstruction task on the training set as a self-supervised learning task to train the constructed dual-channel U-Net network.

[0037] The reconstruction module feeds test set data containing multi-source information into a trained dual-channel U-Net network to reconstruct the missing low-frequency seismic data.

[0038] Compared with the prior art, the present invention has at least the following beneficial effects:

[0039] A self-supervised deep learning method for low-frequency reconstruction of seismic data is proposed. This method avoids reliance on labeled seismic data in the target frequency band, improving its feasibility. By learning the mapping relationship between mid-to-high frequency and low-frequency seismic data through auxiliary tasks, and addressing waveform mismatch issues during the prediction and generalization stages, high-precision low-frequency seismic data can be reconstructed. Furthermore, multi-source information is incorporated into the low-frequency reconstruction process, reflecting more layers of information about the seismic signal and serving as a supplement to the seismic data. For example, the envelope information of a signal has a smaller frequency band than the original signal. Based on the self-supervised deep learning algorithm, by incorporating envelope information, the method fully utilizes the slowly changing information of the signal reflected in the envelope. The dual-channel U-Net network fully combines the seismic data and its envelope information, enabling more accurate reconstruction of seismic data lacking low-frequency information. This improves the accuracy of the reconstructed low-frequency seismic data and effectively overcomes the cycle skipping problem.

[0040] Furthermore, three frequency bands of data required for network training are obtained through different zero-phase digital filters. The envelope operator is applied to the three frequency band seismic data to obtain the corresponding envelope information. The multi-source information is downsampled and energy equalized to provide the required multi-source information for the self-supervised deep learning seismic data low-frequency reconstruction method and provide data support for network training.

[0041] Furthermore, before training the network, to ensure the simplicity of the deep learning network structure, time-domain downsampling was performed on the seismic multi-source information in three frequency bands, while satisfying the sampling theorem. Since the energy of near-offset channels is stronger than that of far-offset channels, energy equalization processing of the seismic multi-source information in the three frequency bands can prevent overfitting at near-offset and underfitting at far-offset during network training. Time-domain downsampling and energy equalization processing of the multi-source information help the network learn the mapping relationship from high-frequency to low-frequency information, thereby extrapolating to obtain high-precision low-frequency data, which helps overcome the cycle skipping problem in full waveform inversion.

[0042] Furthermore, a dual-channel U-Net network is used for self-supervised learning, mapping high-frequency multi-source information to target low-frequency seismic data. The dual-channel input enables the U-Net network to perform low-frequency extrapolation using multi-source information, improving the accuracy of reconstructed low-frequency seismic data. In the dual-channel U-Net network structure, the output data size is consistent with the input data size.

[0043] Furthermore, to evaluate the performance of the network model parameters, an objective function is constructed. During network training, aiming to minimize the objective function, the Adam optimizer is used with a constant learning rate to train the network to learn the mapping relationship from input to output from the dataset, iteratively updating the network weight parameters to obtain the optimal parameters for extrapolating low-frequency seismic data using the deep learning network. The Adam optimization algorithm is particularly effective for deep learning tasks.

[0044] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0045] In summary, this invention employs a self-supervised deep learning method for low-frequency reconstruction of seismic data, mitigating the cycle skipping problem. It establishes a more accurate long-wavelength velocity model as the initial velocity model for traditional full-waveform inversion methods, further improving the final inversion results. Multi-source information is integrated into the low-frequency reconstruction process, fully utilizing multi-layered information to improve the accuracy of extrapolating low-frequency seismic data. Since low-frequency data is often missing in actual datasets, the self-supervised deep learning method avoids reliance on low-frequency data labels. An auxiliary task is used to learn the mapping relationship between high-frequency multi-source information and low-frequency seismic data, thereby reconstructing low-frequency seismic data using a dual-channel U-Net network. In the low-frequency reconstruction process, the full utilization of multi-source information and the self-supervised deep learning method results in better improvement of low-frequency data reconstruction accuracy, overcoming the cycle skipping problem in full-waveform inversion, and obtaining a high-precision velocity model.

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the process of the present invention;

[0048] Figure 2 This is a schematic diagram of the dual-channel U-Net convolutional neural network structure;

[0049] Figure 3 A schematic diagram of the velocity model for testing a self-supervised deep learning seismic data low-frequency reconstruction method based on multi-source information;

[0050] Figure 4 This is a schematic diagram comparing the 3-5Hz low-frequency earthquake data predicted by the network. (a) is the actual 3-5Hz data, (b) is the network-predicted 3-5Hz data, (c) is the network prediction error, and (d) is the prediction result of the 42nd single channel. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0053] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0054] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0055] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0056] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0057] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0058] This invention provides a self-supervised deep learning method for low-frequency reconstruction of seismic data. It involves filtering observed seismic data to obtain seismic data in different frequency bands, extracting multi-source information from these data, and constructing a network training dataset after preprocessing. A self-supervised learning task is then constructed based on the multi-source information of the seismic data, using a dual-channel U-Net neural network to learn the mapping relationship from effective frequency band seismic data to low-frequency seismic data. The multi-source information of the seismic data lacking low-frequency data is resampled in time and energy equalized before being input into the trained dual-channel U-Net neural network to reconstruct the low-frequency seismic data.

[0059] Please see Figure 1 This invention discloses a self-supervised deep learning method for low-frequency reconstruction of seismic data, comprising the following steps:

[0060] S1. Design zero-phase digital filters for the frequency bands corresponding to Band1, Band2, and Band3 involved in the training and test data.

[0061] The frequency band ranges for Band1, Band2, and Band3 are:

[0062]

[0063] Wherein, Band1 is the input frequency band for the network training and validation sets, Band2 is the target frequency band for the network training and validation sets, and Band3 is the input frequency band for the test set.

[0064] S2. Obtain seismic data u for three frequency bands based on the three-band zero-phase digital filters designed in step S1. Band1 (t), u Band2 (t), u Band3 (t); Apply the envelope operator to the seismic data u Band1 (t), u Band3 (t), obtain its corresponding envelope information.

[0065] The envelope operator is defined as follows:

[0066]

[0067] Where u(t) is a one-dimensional time signal, G(u(t)) is the envelope operator of the signal u(t), and H(·) is the Hilbert operator.

[0068] Envelope operator applied to seismic data u Band(n) After obtaining the corresponding envelope information for (t), n = 1, 2, 3, the envelope information is obtained. Therefore, a deep learning network is constructed to reconstruct a dataset of low-frequency earthquake data.

[0069] S3. Time-domain downsampling and energy equalization processing are performed on the multi-source seismic information of the three frequency bands to obtain U. Band1 (t), U Band2 (t), U Band3 (t), Then, a training set and a test set are constructed for training the dual-channel U-Net network;

[0070] Multi-source information is downsampled by a factor of m (m>1) over time, data from the left and right sides near the shot point are removed, and energy equalization is applied to each data point proportional to its distance from the shot point. Band(n) After preprocessing (t), n=1,2,3, the seismic data U is obtained. Band(n) (t), n=1,2,3.

[0071] Before training the dual-channel U-Net network, N s Cannon N r ×N t The data of varying sizes undergoes a data preprocessing process, including two operations: time-wise downsampling and energy equalization.

[0072] First, the data for each shot is downsampled by a factor of m in the time direction, changing the time direction from the original N. t Sampling points become Point, get Size data per shot.

[0073] Subsequently, to avoid overfitting at near offset and underfitting at far offset, the five data points near the shot point were removed. That is, the five data points near each shot point were assigned a value of 0, and then energy equalization was performed on each data point in proportion to its distance from the shot point.

[0074] For N s Forward modeling data for each shot, with N corresponding to each shot's data. r road Dimensional data is denoted as N d =N r The grid position corresponding to the shot point of each shot is denoted as . The grid position corresponding to each detector is denoted as . For the j-th shot point and the i-th data line, the envelope energy equalization ratio is set to balance_ratio = 0.01|P s,j -P r,i | Let the balanced data be _____ Then we have:

[0075] d b,i =d i ·0.01·|P s,j -P r,i | (3)

[0076] wherein j=1,2,...,N s , i=1,2,...,N r .

[0077] Finally, after data preprocessing, sized training set is obtained to train the dual-channel U-Net network.

[0078] To solve the problem of waveform mismatch at a single sampling point, a test set is constructed after re-performing m×n-fold time downsampling and energy equalization preprocessing on the multi-source information of seismic data in the Band3 frequency band. The overall frequency band of the test set is reduced by n times compared with that of the training set, and the number of waveforms at a single sampling point also becomes n times that of the training set data. The multi-source information of seismic data in Band3 frequency band is subjected to m×n-fold downsampling to keep the number of waveforms at a single sampling point consistent with that of the training set, and the seismic data after resampling and energy equalization preprocessing is used as the test set.

[0079] Preprocessing is performed on the seismic data u Band3 (t) to solve the problem of waveform mismatch at a single sampling point in the prediction and generalization stages.

[0080] The input frequency band of the training set is [f1,f2], the target frequency band is [nf1,f2], 0<n<1, and the input frequency band of the test set data is [nf1,nf2]. The seismic data with input frequency band [nf1,nf2], 0<n<1 of the test set is subjected to m×n-fold time downsampling to keep the number of waveforms at a single sampling point consistent with that of the training set data, and the obtained dimension is data per shot. To facilitate network training, zero-padding is performed on each shot data to dimensions. Subsequently, five traces of data on the left and right near the shot point are removed, that is, the five traces of data on the left and right of each shot point are assigned a value of 0, and then energy equalization is performed on each trace of data in proportion to the distance from the shot point.

[0081] S4. A dual-channel U-Net convolutional neural network is built on the Pytorch platform, and the network structures corresponding to the two different channels are different;

[0082] See Figure 2 , in the dual-channel U-Net network structure, seismic data is input as one of the channels. After passing through the U-Net network, the entire U-Net network includes 15 convolutional layers, 3 max-pooling layers, 3 transposed convolutional layers and 3 concatenation layers, and the activation function adopts ReLU.

[0083] The envelope information of seismic data is input as one channel, after passing through 2 convolutional layers, it is concatenated with the output of the U-Net network of seismic data, and then the final output is obtained after passing through 3 convolutional layers.

[0084] S5, constructing a low-frequency data reconstruction task on the training set as a self-supervised learning task, and training the dual-channel U-Net convolutional network obtained in step S4 by using the training set data to learn the mapping relationship from multi-source information of effective frequency-band seismic data to low-frequency-band seismic data;

[0085] In the network training process, the input of one channel is medium and high frequency band seismic data U Band1 (t), and the input of the other channel is the envelope of medium and high frequency band seismic data The network output is low frequency band seismic data U Band2 (t).

[0086] The constructed self-supervised learning task is a low-frequency data reconstruction task on the training set, which enables the dual-channel U-Net network to mine the correlation of the data itself, and solves the problem of waveform non-correspondence at a single sampling point in the prediction and generalization stages.

[0087] The training set is used to train the dual-channel U-Net convolutional network under self-supervised learning, and the envelope information is integrated to learn the mapping relationship from effective frequency band seismic data to low-frequency seismic data. For example, taking the input frequency band Band1∈[f1,f2] of the training set and the target frequency band Band2∈[nf1,f2], 0<n<1 as an example, the seismic data envelope information of the frequency band Band1∈[f1,f2] is obtained through an envelope operator. One channel of the dual-channel U-Net convolutional neural network inputs Band1 frequency band seismic data, the other channel inputs Band1 frequency band seismic data envelope information, and the target data is the seismic data of frequency band Band2.

[0088] During network training, shot gather data is used for network training. For size training set, 80% of the shot data is used as the network training set, and the remaining shot gather data is used as the network validation set; U Band1 (t) is the input seismic data of the training set, is the input envelope data of the training set, Net(·) is the network mapping operator, is the low-frequency data obtained by network mapping, then:

[0089]

[0090] where θ is the network parameter.

[0091] Setting the objective function minimized by the network as the mean square loss function, we have:

[0092]

[0093] where N is the number of samples in the network training set, The effective frequency band data is the input to the training set. The effective frequency band envelope data is the input to the training set. The training set contains seismic data in the target frequency band.

[0094] The network uses the Adam optimizer, iteratively updating the network parameters θ with a step size η = 0.001.

[0095]

[0096] Where, θ i This represents the parameters of the network in the i-th iteration.

[0097] S6. The seismic data in band Band 3 and its envelope information are resampled in the time domain and energy equalized before being input into the trained dual-channel U-Net network to reconstruct the missing low-frequency data U. Band4 (t).

[0098] In another embodiment of the present invention, a self-supervised deep learning seismic data low-frequency reconstruction system is provided. This system can be used to implement the above-mentioned self-supervised deep learning seismic data low-frequency reconstruction method. Specifically, the self-supervised deep learning seismic data low-frequency reconstruction system includes a data module, a network module, and a reconstruction module.

[0099] The data module obtains earthquake data of different frequency bands from observed earthquake data, extracts multi-source information from the earthquake data, and obtains training set data and test set data after time downsampling and energy equalization.

[0100] The network module constructs a dual-channel U-Net network and uses a low-frequency data reconstruction task on the training set as a self-supervised learning task to train the constructed dual-channel U-Net network.

[0101] The reconstruction module feeds test set data containing multi-source information into a trained dual-channel U-Net network to reconstruct the missing low-frequency seismic data.

[0102] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a self-supervised deep learning seismic data low-frequency reconstruction method, including:

[0103] Seismic data of different frequency bands are obtained from observed seismic data, and multi-source information of the seismic data is extracted. After time downsampling and energy equalization, training set data and test set data are obtained. A dual-channel U-Net network is built, and a low-frequency data reconstruction task on the training set data is constructed as a self-supervised learning task to train the built dual-channel U-Net network. The test set data containing multi-source information is fed into the trained dual-channel U-Net network to reconstruct the missing low-frequency seismic data.

[0104] In another embodiment of the present invention, a storage medium is also provided, specifically a computer-readable storage medium (memory). This computer-readable storage medium is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0105] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the self-supervised deep learning seismic data low-frequency reconstruction method in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by the processor in the following steps:

[0106] Seismic data of different frequency bands are obtained from observed seismic data, and multi-source information of the seismic data is extracted. After time downsampling and energy equalization, training set data and test set data are obtained. A dual-channel U-Net network is built, and a low-frequency data reconstruction task on the training set data is constructed as a self-supervised learning task to train the built dual-channel U-Net network. The test set data containing multi-source information is fed into the trained dual-channel U-Net network to reconstruct the missing low-frequency seismic data.

[0107] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0108] Model Examples

[0109] The specific implementation process of this invention was applied to a portion of the Marmousi model. The model has a horizontal dimension of 5090m and a depth dimension of 1330m, and the velocity model is as follows: Figure 3 As shown.

[0110] In this model, the input frequency band of the training set is Band1 ∈ [10Hz, 25Hz], the target frequency band of the training set is Band2 ∈ [5Hz, 25Hz], the input frequency band of the test set is Band3 ∈ [5Hz, 12.5Hz], and the target frequency band of the test set is Band4 ∈ [2.5Hz, 12.5Hz]. The seismic data and their envelope information of the input frequency band Band3 ∈ [5Hz, 12.5Hz] of the test set are preprocessed and then input into the trained dual-channel U-Net network.

[0111] After restoring the original energy values ​​of the 2.5–12.5 Hz frequency band data output by the network, filtering is performed to obtain low-frequency data in the 3–5 Hz frequency band. The reconstructed 3–5 Hz frequency band seismic data is then compared with the actual 3–5 Hz frequency band data. Figure 4 As shown.

[0112] Figure 4 middle, Figure 4 (a) is the actual 3-5Hz seismic data of the 510th shot; Figure 4 (b) is the 3-5Hz frequency band seismic data reconstructed from the network; Figure 4 (c) represents the corresponding prediction error; Figure 4 (d) is a one-dimensional waveform comparison diagram of the 42nd path of the 510th gun. Figure 4 (d) It can be seen that the low-frequency earthquake data predicted by the network is very close to the real data in terms of amplitude and phase. The maximum error between the two is close to 0, indicating that based on multi-source information (envelope information), through a self-supervised learning task, the network can learn the mapping relationship between mid- and high-frequency earthquake data and low-frequency earthquake data. It has generalization on the test set and the reconstructed low-frequency earthquake data has high accuracy, providing a data foundation for subsequent inversion.

[0113] In summary, this invention presents a self-supervised deep learning method and system for low-frequency reconstruction of seismic data. For observed seismic data, it extracts multi-source information from seismic data in different frequency bands. Utilizing a dual-channel U-Net network, and based on a self-supervised deep learning method, it learns the mapping relationship from effective mid-to-high frequency data to low-frequency seismic data through a low-frequency reconstruction auxiliary task on the training set. This method can reconstruct missing low-frequency seismic data information from mid-to-high frequency seismic data. Since low-frequency data is less sensitive to cycle-skipping problems compared to high-frequency data, the proposed method can overcome this issue, providing a good data foundation for traditional full waveform inversion (FWI), thereby further improving the accuracy of the final velocity model inversion.

[0114] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0115] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0116] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0117] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0118] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0120] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0121] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A self-supervised deep learning method for low-frequency reconstruction of seismic data, characterized in that, Seismic data at different frequency bands were obtained from observed seismic data. Multi-source information of the seismic data was extracted, and after time downsampling and energy equalization, training set data and test set data were obtained, specifically as follows: The design of training and test set data involves , , Zero-phase digital filter for the corresponding frequency band; According to the design , , The corresponding frequency band zero-phase digital filter is obtained , , seismic data in frequency bands , , ; Applying the envelope operator to seismic data Then the corresponding envelope information is obtained. , A dataset for reconstructing low-frequency earthquake data using a deep learning network was constructed. Time-domain downsampling and energy equalization processing were performed on multi-source seismic information from three frequency bands to obtain... , , , , To construct the training and test sets for training a dual-channel U-Net network, , , The frequency band range of the frequency band data is: in, For the input frequency bands of the network training and validation sets, The target frequency bands for the network training and validation sets; The input frequency band for the test set; The specific steps for time-domain downsampling and energy equalization processing of multi-source seismic information in three frequency bands are as follows: The training set input frequency band is The target frequency band is The input frequency band for the test set data is ; Multi-source information is downsampled by a factor of m over time, where m > 1, and the time direction changes from the original... Sampling points become Point, get The data for each shot is calculated; the five data points to the left and right of the firing point are removed, and energy is balanced for each data point proportional to its distance from the firing point. Training set data of a certain size For the number of cannons, This represents the number of seismic channels per shot / the number of geophones. For frequency band Seismic data from multiple sources was re-timed. Double downsampling and energy equalization processing Input frequency band for test set data earthquake data Downsampling at multiple times the time, maintaining the same waveform count as a single sampling point in the training set data, yields a dimension of... Data for each shot; pad each shot data with 0 to... After dimensioning, the five channels to the left and right of each shot point are removed, i.e., each shot point's five channels are assigned a value of 0. Then, energy equalization is performed on each channel proportional to its distance from the shot point to obtain the test set data. A dual-channel U-Net network is constructed, consisting of 15 convolutional layers, 3 max-pooling layers, 3 transposed convolutional layers, and 3 concatenation layers. The activation function used is ReLU. Seismic data is used as one channel input to the dual-channel U-Net network. After passing through two convolutional layers, it is concatenated with the U-Net network output of the seismic data, and then passed through three more convolutional layers to obtain the final output. A low-frequency data reconstruction task on the training set is constructed as a self-supervised learning task to train the constructed dual-channel U-Net network. During training, one channel input is mid-to-high frequency seismic data. The other channel input is a mid-to-high frequency band seismic data envelopment. The network output is low-frequency seismic data. The test set data containing multi-source information is fed into the trained dual-channel U-Net network to reconstruct the missing low-frequency seismic data.

2. The self-supervised deep learning seismic data low-frequency reconstruction method according to claim 1, characterized in that, The envelope operator is defined as follows: in, It is a one-dimensional time signal. For signal The envelope operator, For Hilbert operators.

3. The self-supervised deep learning seismic data low-frequency reconstruction method according to claim 1, characterized in that, Data after energy balance for: in, For each shot, the corresponding seismic data. For the first j The grid positions corresponding to each shot point For the first i The grid position corresponding to the data.

4. The self-supervised deep learning method for low-frequency reconstruction of seismic data according to claim 1, characterized in that, The objective function for training the dual-channel U-Net network is: in, N The number of samples in the network training set. The effective frequency band data is the input to the training set. The effective frequency band envelope data is the input to the training set. For the target frequency band seismic data in the training set, The objective function for training the dual-channel U-Net network; Using the Adam optimizer, with step size Iteratively update network parameters : in, Indicates the first i The parameters of the next iteration of the network.

5. A self-supervised deep learning seismic data low-frequency reconstruction system, characterized in that, include: The data module acquires seismic data at different frequency bands from observed seismic data, extracts multi-source information from the seismic data, and obtains training set data and test set data after time downsampling and energy equalization. Specifically: The design of training and test set data involves , , Zero-phase digital filter for the corresponding frequency band; According to the design , , The corresponding frequency band zero-phase digital filter is obtained , , seismic data in frequency bands , , ; Applying the envelope operator to seismic data Then the corresponding envelope information is obtained. , A dataset for reconstructing low-frequency earthquake data using a deep learning network was constructed. Time-domain downsampling and energy equalization processing were performed on multi-source seismic information from three frequency bands to obtain... , , , , To construct the training and test sets for training a dual-channel U-Net network, , , The frequency band range of the frequency band data is: in, For the input frequency bands of the network training and validation sets, The target frequency bands for the network training and validation sets; The input frequency band for the test set; The specific steps for time-domain downsampling and energy equalization processing of multi-source seismic information in three frequency bands are as follows: The training set input frequency band is The target frequency band is The input frequency band for the test set data is ; Multi-source information is downsampled by a factor of m over time, where m > 1, and the time direction changes from the original... Sampling points become Point, get The data for each shot is calculated; the five data points to the left and right of the firing point are removed, and energy is balanced for each data point proportional to its distance from the firing point. Training set data of a certain size; For frequency band Seismic data from multiple sources was re-timed. Double downsampling and energy equalization processing Input frequency band for test set data earthquake data Downsampling at multiple times the time, maintaining the same waveform count as a single sampling point in the training set data, yields a dimension of... Data for each shot; pad each shot data with 0 to... After dimensioning, remove the five data points to the left and right of each shot point, that is, assign 0 values ​​to the five data points to the left and right of each shot point, and then perform energy balancing on each data point in proportion to its distance from the shot point to obtain the test set data. The network module constructs a dual-channel U-Net network. A low-frequency data reconstruction task on the training set is used as a self-supervised learning task. The constructed dual-channel U-Net network is trained on this network, which includes 15 convolutional layers, 3 max-pooling layers, 3 transposed convolutional layers, and 3 concatenation layers. The activation function used is ReLU. Seismic data is used as one channel input to the dual-channel U-Net network. After passing through two convolutional layers, it is concatenated with the U-Net network output of the seismic data, and then passed through three more convolutional layers to obtain the final output. During training, one channel input consists of mid-to-high frequency seismic data. The other channel input is a mid-to-high frequency band seismic data envelopment. The network output is low-frequency seismic data. ; The reconstruction module feeds test set data containing multi-source information into a trained dual-channel U-Net network to reconstruct the missing low-frequency seismic data.

Images