Seismic data velocity filtering processing method based on deep learning

Abstract

The invention discloses a seismic data velocity filtering processing method based on deep learning, and relates to the technical field of seismic data processing, and the method comprises the steps: S1, constructing a preliminarily trained wave field separation model; s2, making a standard up-going wave and a down-going wave; optimizing the obtained uplink waves and downlink waves to obtain a manufactured data set; s3, segmenting the manufactured data set to obtain final data blocks, and optimizing the preliminarily trained wave field separation model based on the final data blocks to obtain a trained wave field separation model; s4, filtering the obtained seismic data through the wave field separation model to obtain a corresponding filtering result, the method is applied to wave field separation and denoising processing of the seismic data, coherent noise is effectively suppressed through the speed difference between the target wave field and the non-target wave field, and therefore the seismic imaging quality and exploration precision are improved.

Description

Technical Field

[0001] This application relates to the field of seismic data processing technology, and in particular to a seismic data velocity filtering processing method based on deep learning. Background Technology

[0002] In seismic data processing, velocity filtering, as an important wavefield separation and denoising technique, is widely used in noise suppression, multiple attenuation, surface wave and linear noise removal, and other stages. Traditional velocity filtering methods mainly include transform domain-based filtering methods and time domain-based filtering methods.

[0003] Common transform domain velocity filtering algorithms mainly include frequency-wavenumber domain (fk) filtering and Radon transform filtering. fk filtering is prone to artifacts and spatial aliasing during processing, especially when seismic data amplitude differences are large, and may also cause waveform distortion. While Radon transform filtering can separate the velocity domain, it struggles to accurately preserve amplitude information. To overcome these shortcomings, researchers have proposed improved methods such as weighted least squares Radon transform, parabolic Radon transform based on singular value decomposition constraints, frequency-varying Radon transform, and short-time Fourier transform filtering. Although these methods improve filtering performance to some extent, they often require the introduction of inversion or additional constraints, significantly increasing computational complexity. Radon transform-based methods, in particular, are inherently inefficient, further limiting their application in large-scale data processing.

[0004] Time-domain velocity filtering methods mainly include median filtering and singular value filtering. These methods rely on picking up seismic phase axes to separate the target wavefield from the interfering wavefield. Their advantages include avoiding spatial aliasing and the ability to flexibly extract the wavefield with manual intervention; however, their disadvantages include low automation and a tendency to over-smooth the effective signal, reducing processing accuracy.

[0005] In recent years, migration-based filtering methods have emerged, which can achieve similar effects to velocity filtering in specific applications, such as suppressing linear noise or achieving VSP wavefield separation. However, due to the high hardware resource requirements of wave equation calculations, these methods have low computational efficiency and are difficult to adapt to the practical processing needs of large-scale seismic data. Furthermore, the applicability of these methods is relatively limited; for example, inverse migration-based algorithms can only obtain the wavefield after image focusing and cannot flexibly separate wavefield components of different velocity components in the original data, thus limiting their widespread application.

[0006] In summary, existing velocity filtering methods still suffer from problems such as artifacts, waveform distortion, inaccurate amplitude preservation, low computational efficiency, and limited applicability in seismic data processing. With the development of deep learning technology, researching an efficient, universal seismic data velocity filtering algorithm that can preserve signal characteristics has become an urgent need to improve seismic imaging quality and exploration accuracy. Summary of the Invention

[0007] The purpose of this application is to provide a deep learning-based velocity filtering method for seismic data, overcoming the problems of severe artifacts, waveform distortion, inaccurate amplitude energy, low computational efficiency, and limited applicability in existing seismic data velocity filtering methods. This method utilizes a deep learning network to intelligently separate the wavefields of different velocity components in seismic records, thereby effectively suppressing coherent noise, preserving the waveform and amplitude characteristics of the effective signal, and significantly reducing computational complexity while improving processing accuracy, thus meeting the high-efficiency processing requirements of large-scale seismic data.

[0008] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a seismic data velocity filtering processing method based on deep learning, the method comprising: S1. An encoder-decoder network based on the U-Net framework is constructed, and a residual module and a squeezing-excitation module are introduced to build a preliminary trained wavefield separation model. S2. Based on the pre-trained wavefield separation model, the convolution method is used to create standard up-going and down-going waves; and the obtained up-going and down-going waves are optimized based on an iterative optimization strategy to obtain the completed dataset. S3. Divide the completed dataset into final data blocks, and optimize the initially trained wavefield separation model based on the final data blocks to obtain the trained wavefield separation model. S4. The acquired seismic data is filtered using a wavefield separation model to obtain the corresponding filtering results.

[0009] This method is applicable to typical problems such as linear noise suppression, VSP uplink / downlink separation, and VSP P-wave / S-wave separation. To address the limitation of speed selection under low sampling rate conditions, this invention further introduces data resampling and time interpolation strategies to ensure the accuracy and applicability of the filtering process in different scenarios.

[0010] Furthermore, the wavefield separation model initially trained in S1 specifically includes: Network encoder, bottleneck section, decoder, and output unit; The network encoder consists of four downsampling stages, each consisting of two consecutive two-dimensional convolutional layers with ReLU activation function. It is then connected to a squeeze-excitation module for channel feature recalibration and a max pooling layer to reduce spatial resolution. The decoder consists of four upsampling stages. Each stage first improves the spatial resolution through transposed convolution, and then performs skip connections with the corresponding encoder features to obtain the concatenated features. The bottleneck part consists of a residual module composed of two convolutional layers. The residual module optimizes and reconstructs the concatenated features to obtain the decoded features. The output unit, containing two-dimensional convolutional layers, is used to map the decoded features to the target output.

[0011] Furthermore, the production of the standard uplink and downlink waves specifically includes: One-dimensional reflection coefficient sequences of corresponding types are randomly generated based on historical data; Furthermore, by increasing the amplitude value of the shallow reflection coefficient based on the one-dimensional reflection coefficient sequence, a preliminary enhanced one-dimensional reflection coefficient sequence is obtained; Based on the initially enhanced one-dimensional reflection coefficient sequence, multiple reflection points are randomly selected, and the energy of the selected multiple reflection points is randomly increased to obtain the final one-dimensional time series; The final one-dimensional time series is extended into a two-dimensional reflection series with an upward trend or a two-dimensional reflection series with a downward trend. The obtained two-dimensional reflection sequences with upward or downward trends are then augmented with data to obtain a two-dimensional reflection coefficient matrix and wavelets. The standard upwave and downwave are obtained by convolving the two-dimensional reflection coefficient matrix and the wavelet.

[0012] Furthermore, the process of optimizing the obtained uplink and downlink waves based on an iterative optimization strategy to obtain the completed dataset specifically includes: The uplink and downlink separation model is trained based on historical data; The obtained dataset is tested using a model that separates uplink and downlink waves to generate detection results. The completed dataset is iterated based on the detection results to obtain clean uplink and downlink waves for label creation; The separated pure upwave and downwave are added together, and the combined data is used as the input features of the network. Upwave labels and downwave labels are generated based on the pure upwave. The dataset is created by generating uplink labels and downlink labels based on clean uplink waves.

[0013] Furthermore, the step of dividing the completed dataset into final data blocks specifically includes: The optimized final dataset is generated through optimization strategies; The optimized final dataset is then divided into preliminary data blocks. Each initial data block is then normalized to generate data blocks with consistent feature scales. The final data block is generated based on data blocks with consistent feature scales.

[0014] Furthermore, the step of generating the optimized final dataset through the optimization strategy specifically includes: Based on the idea in unsupervised denoising tasks, we perform unsupervised learning on the labels of uplink or downlink waves and set the labels to be the same as the input data. In the process of fitting uplink or downlink data to the network, the final dataset is obtained by fitting the data in order of priority based on the data type.

[0015] Furthermore, the step of filtering the acquired seismic data using a wavefield separation model to obtain the corresponding filtering results specifically includes: By specifying the time-distance curve or linear time-distance curve for correction, the wave field within the target apparent velocity range is mapped to a field consisting only of up-going or down-going waves through time difference correction; then, a multiple up-going and down-going wave separation strategy is adopted to perform precise filtering of the data within a specific velocity range; finally, the original domain space is restored through inverse transformation, thereby achieving the separation and extraction of the target velocity components.

[0016] The seismic data velocity filtering method based on deep learning provided in this application has the following technical advantages compared with the prior art: The wave field separation accuracy is significantly improved, effectively suppressing multiple types of noise and accurately separating the target wave field from the interfering wave field: relying on the U-Net network that integrates the residual module and the squeeze-excitement (SE) module, the feature representation and reconstruction capabilities are enhanced, which can accurately extract target wave fields such as uplink and downlink waves, while efficiently suppressing coherent noise.

[0017] Abandoning the drawbacks of traditional transform domain methods that rely on inversion or constraints, this method achieves filtering by intelligently learning wave field velocity differences, effectively solving problems such as artifacts, spatial aliasing, waveform distortion, and inaccurate amplitude preservation that are easily generated by traditional methods.

[0018] Optimize computational efficiency, adapt to large-scale data processing, reduce computational complexity, and transform various velocity filtering problems into uplink and downlink wave separation tasks through a unified processing approach. This avoids the high computational cost of traditional Radon transform-type methods and the stringent hardware resource requirements of offset-based filtering methods.

[0019] By eliminating the reliance on phase axis picking in time-domain methods, and automatically identifying wavefield characteristics through deep learning models, manual operations are reduced, which not only reduces human error but also improves the automation level of the processing flow. This is especially suitable for seismic data processing in complex tectonic regions.

[0020] Covering a wide range of seismic processing needs: It can be flexibly applied to typical scenarios such as linear noise suppression, VSP uplink and downlink wave separation, VSP P-wave and S-wave separation, DAS common-mode noise suppression, and CDP gather optimization. For example, in CDP gather optimization, compared with weighted least squares parabolic Radon transform (WLSRT), it can avoid the defect of abnormal energy enhancement in the boundary region, and the noise suppression effect is better; in common-mode noise suppression, compared with traditional stacking methods, it can completely remove the noise residue caused by shallow strong energy interference.

[0021] By accurately filtering and preserving key features such as waveform and amplitude of effective signals, more reliable basic data can be provided for seismic imaging. This helps to more clearly depict underground structural morphology, reduce exploration errors, and provide more accurate technical support for oil and gas resource exploration and geological disaster prediction. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the original DAS-VSP data provided in one embodiment of this application; Figure 2 This is a comparative schematic diagram of a DAS-VSP data wavefield separation method provided in an embodiment of this application; wherein, Figure 2 Images (a)–(e) show the results of median filtering with a small window, including... Figure 2 (a) shows the downlink P-wave filtering result. Figure 2 (b) shows the downlink S-wave filtering result. Figure 2 (c) shows the uplink P-wave filtering result. Figure 2 (d) shows the uplink S-wave filtering result and Figure 2 (e) represents the filtering result of the residual wave field; Figure 2 The values ​​in (f)–(j) represent the results of the large-window median filtering, including... Figure 2 (f) shows the downlink P-wave filtering result. Figure 2 (g) represents the downlink S-wave filtering result. Figure 2 The middle (h) represents the uplink P-wave filtering result. Figure 2(i) Uplink S-wave filtering results Figure 2 In the middle (j), the result of the residual wave field filtering is shown. Figure 2 In the middle, (k)–(o) represent the filtering results of the deep learning method described in this invention, including... Figure 2 (k) represents the downlink P-wave filtering result. Figure 2 The middle part shows the filtering results of the downlink S-wave (l). Figure 2 The middle (m) represents the uplink P-wave filtering result. Figure 2 The middle (n) represents the result of the upward S-wave filtering. Figure 2 In the middle (o), the result of the residual wave field filtering is shown. Figure 3 This is a schematic diagram of the arc-shaped noise in the imaging results; Figure 4 This is a comparison chart of the arc-shaped noise suppression results provided in the embodiments of this application, wherein, Figure 4 In the middle (a), the effective signal obtained by narrow filter window fk filtering is shown. Figure 4 In the middle (d), the noise removed by the narrow filter window fk is represented; Figure 4 (b) shows the effective signal obtained by wide-window filtering (fk). Figure 4 (e) represents the noise removed by the wide filtering window fk; Figure 4 (c) represents the effective signal obtained by the proposed deep learning method. Figure 4 In the middle (f), the noise removed by the proposed deep learning method is represented; Figure 5 The common-mode noise suppression results provided in the embodiments of this application, wherein, Figure 5 (a) is a schematic diagram of the original data; Figure 5 (b) shows the common-mode noise suppression result of the superposition method; Figure 5 (c) shows the common-mode noise suppression results of the proposed method; Figure 5 (d) shows the common-mode noise suppression result achieved by the superposition method; Figure 5 (e) shows the common-mode noise suppression result achieved by the proposed method; Figure 6 This is a schematic diagram of CDP gather optimization, where, Figure 6 (a) represents the original data; Figure 6 (b) shows the results of the WLSRT method; Figure 6 (c) shows the result of the proposed method for optimizing the trace set; Figure 6 (d) shows the optimized noise gather result suppressed by the WLSRT method; Figure 6 In Figure (e), the noise gather optimization result suppressed by the proposed method is shown.

[0024] Figure 7 A schematic diagram of a UNet-like network for wavefield separation; Figure 8A schematic diagram of the iterative strategy technical roadmap: Figure 8 (a) is a flowchart for determining whether the network meets the accuracy requirement. Figure 8 (b) is the flowchart of the iteration strategy; Figure 9 A diagram illustrating the label's transformation process: Figure 9 In the middle (a), the original wave field is shown. Figure 9 (b) shows the first prediction result of the initial model; Figure 9 (c) shows the label results after the iterative strategy; Figure 9 (d) represents the result obtained by further employing an optimization strategy based on the iteration; Figure 9 In the middle (e), the difference between the original wavefield and the first prediction result of the initial model is shown. Figure 9 In the middle (f), the difference between the initial model's first prediction and the label result after the iterative strategy is shown. Figure 9 In the middle (g), the difference between the label result after the iterative strategy and the result obtained by further using the optimization strategy based on the iteration is shown. Figure 10 Flowchart of AI speed filtering operation: Figure 10 (a) is a simplified diagram. Figure 10 (b) is a schematic diagram of earthquake data; Figure 11 This is a flowchart of a seismic data velocity filtering method based on deep learning. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] like Figure 11 This embodiment illustrates a seismic data velocity filtering method based on deep learning, the method comprising: S1. An encoder-decoder network based on the U-Net framework is constructed, and a residual module and a squeezing-excitation module are introduced to build a preliminary trained wavefield separation model. S2. Based on the pre-trained wavefield separation model, the convolution method is used to create standard up-going and down-going waves; and the obtained up-going and down-going waves are optimized based on an iterative optimization strategy to obtain the completed dataset. S3. Divide the completed dataset into final data blocks, and optimize the initially trained wavefield separation model based on the final data blocks to obtain the trained wavefield separation model. S4. The acquired seismic data is filtered using a wavefield separation model to obtain the corresponding filtering results.

[0028] Optionally, the wavefield separation model initially trained in S1 specifically includes: Network encoder, bottleneck section, decoder, and output unit; The network encoder consists of four downsampling stages, each consisting of two consecutive two-dimensional convolutional layers with ReLU activation function. It is then connected to a squeeze-excitation module for channel feature recalibration and a max pooling layer to reduce spatial resolution. The decoder consists of four upsampling stages. Each stage first improves the spatial resolution through transposed convolution, and then performs skip connections with the corresponding encoder features to obtain the concatenated features. The bottleneck part consists of a residual module composed of two convolutional layers. The residual module optimizes and reconstructs the concatenated features to obtain the decoded features. The output unit, containing two-dimensional convolutional layers, is used to map the decoded features to the target output.

[0029] Specifically, to achieve high-precision wavefield separation, this embodiment constructs an encoder-decoder network based on the U-Net framework and introduces a residual module and a squeeze-excitation (SE) module to improve feature representation capabilities and reconstruction accuracy (e.g., Figure 7 (As shown). This network takes shot gather data as input and outputs the separated upgoing wave field. In the absence of other interference, the downgoing wave can be indirectly obtained simply by obtaining the accurate upgoing wave.

[0030] The network encoder consists of four downsampling stages, each containing two consecutive 2D convolutional layers (kernel size 3) using the ReLU activation function. These are followed by an SE module for channel feature recalibration and a max-pooling layer (stride 2) to reduce spatial resolution. The number of feature channels doubles after each downsampling. The bottleneck is the residual module, which contains two convolutional layers (kernel size 3). The decoder has a symmetrical structure and contains four upsampling stages. Each stage first improves spatial resolution through transposed convolution, then performs skip connections with the corresponding encoder features, and finally optimizes the reconstruction using the residual module. Finally, the 2D convolutional layers map the decoded features to the target output.

[0031] Optionally, the production of standard uplink and downlink waves specifically includes: One-dimensional reflection coefficient sequences of corresponding types are randomly generated based on historical data; Furthermore, by increasing the amplitude value of the shallow reflection coefficient based on the one-dimensional reflection coefficient sequence, a preliminary enhanced one-dimensional reflection coefficient sequence is obtained; Based on the initially enhanced one-dimensional reflection coefficient sequence, multiple reflection points are randomly selected, and the energy of the selected multiple reflection points is randomly increased to obtain the final one-dimensional time series; The final one-dimensional time series is extended into a two-dimensional reflection series with an upward trend or a two-dimensional reflection series with a downward trend. The obtained two-dimensional reflection sequences with upward or downward trends are then augmented with data to obtain a two-dimensional reflection coefficient matrix and wavelets. The standard upwave and downwave are obtained by convolving the two-dimensional reflection coefficient matrix and the wavelet.

[0032] Specifically, in order to improve the quality of training labels, this embodiment proposes two dataset construction strategies, both based on the pre-trained wavefield separation model.

[0033] (1) Pre-trained model and convolution label creation To avoid using traditional methods for creating datasets, we start directly from theoretical data and employ the convolution method to generate a large number of standard uplink and downlink waves. The approach to creating convolution labels is as follows: ① First, a one-dimensional reflection coefficient sequence with an amplitude range of 0-50 and a length of 6000 is randomly generated, with a density of 15-70 reflection points generated every 1000 time sampling points.

[0034] ② Increase the amplitude value of the shallow reflection coefficient to simulate stronger shallow energy. In particular, for downlink direct waves, we increased their energy by 50-100 times.

[0035] ③ Considering the presence of certain strong energy axes in the data, we randomly selected 10 reflection points and randomly increased their energy by 5-40 times.

[0036] ④ Extend the one-dimensional time series into a two-dimensional reflection sequence with an upward or downward trend. Two cases are provided: one using a linear transformation, randomly changing the slope of the waveform; the other providing a quadratic upward or downward trend, randomly changing the curvature and slope, as shown in the formula: ; In the formula, x represents the spatial coordinates, i represents the coordinates of the one-dimensional sequence sample points, and t represents the time coordinates. px The slope is linear. pxx Let be the curvature of the parabola.

[0037] ⑤ To account for the simultaneous existence of upward or downward waves with different slopes, we have also added some data, which includes samples of two sets of upward or downward waves with different slopes.

[0038] ⑥ Convolve the two-dimensional reflection coefficient matrix and the wavelet, with the wavelet's dominant frequency range being 10-50Hz, and set multiple wavelets.

[0039] The above process creates a preliminary dataset, which is then used to train an uplink / downlink separation model. A large number of models are then collected, and the dataset is gradually expanded based on the iterative and optimization strategies described below.

[0040] Optionally, the step of optimizing the obtained uplink and downlink waves based on an iterative optimization strategy to obtain the completed dataset specifically includes: The uplink and downlink separation model is trained based on historical data; The obtained dataset is tested using a model that separates uplink and downlink waves to generate detection results. The completed dataset is iterated based on the detection results to obtain clean uplink and downlink waves for label creation; The separated pure upwave and downwave are added together, and the combined data is used as the input features of the network. Upwave labels and downwave labels are generated based on the pure upwave. The dataset is created by generating uplink labels and downlink labels based on clean uplink waves.

[0041] Specifically, firstly, it is necessary to evaluate whether the separation performance of the pre-trained model on new data meets the requirements. Taking the traveling wave as an example, if... Figure 8 As shown in Figure a, if there is no obvious downflow wave residue in the separated upflow wave and no obvious upflow wave residue in the differential profile, the separation result is considered acceptable. If these two conditions are not met, iterative or optimization strategies are needed to further improve the separation accuracy.

[0042] Figure 8Figure b illustrates the specific process of iterative learning. First, the two models trained in the first training iteration are used to obtain initially separated uplink and downlink waves. Taking the uplink wave as an example, if downlink waves still remain in the separated uplink waves, the separation results need to be input into the network again for further separation to extract as pure as possible uplink waves. Since the extracted pure uplink and downlink waves are only used for labeling, the amplitude loss that may occur during repeated separation has little impact on the final training effect. Next, the separated pure uplink and downlink waves are added together, and the combined data is used as the input features of the network, with the pure uplink wave as the label. The initial model of the network is a pre-trained model, which can also be re-initialized. By applying the re-trained model to the original data, higher-precision wavefield separation can be achieved.

[0043] Optionally, the step of dividing the completed dataset into final data blocks specifically includes: The optimized final dataset is generated through optimization strategies; The optimized final dataset is then divided into preliminary data blocks. Each initial data block is then normalized to generate data blocks with consistent feature scales. The final data block is generated based on data blocks with consistent feature scales.

[0044] Specifically, in addition to the iterative strategy, an optimization strategy was introduced to further improve label quality. This strategy draws on ideas from unsupervised denoising tasks, utilizing part of the network's learning ability and learning preferences to achieve the goal.

[0045] Specifically, we perform unsupervised learning directly on the labels of either the rising or falling waves, setting the labels to be the same as the input data. Taking the rising wave as an example, during the process of fitting the rising wave data, the network will prioritize learning its primary feature—the rising wave—before fitting secondary features (such as the falling wave and noise). Therefore, by interrupting training at appropriate training points, we can obtain a model that can further refine the rising wave. We call this process "optimization". Figure 9 The graph illustrates the changes in downlink wave labels during the optimization process. As can be seen from the figure, after initial prediction, most uplink waves were removed, but some strong axes still retained noticeable residues. Through iterative strategies, these strong axis residues were significantly reduced, improving the quality of the downlink wave labels, although a small amount of uplink wave residue remained. After further optimization, the residual uplink waves were further eliminated, resulting in a relatively pure downlink wave label.

[0046] The completed dataset was divided into data blocks (slices) of size 128×128, and each slice was normalized separately to ensure that the feature scale of each slice was consistent during training.

[0047] The model training employs the square root L1 (sqrt L1) loss function, combined with the Adam optimizer for parameter optimization. During training, the batch size is set to 64, and the initial learning rate is 0.0001, to balance training stability and convergence speed.

[0048] During the dataset creation phase, we combined the initial model, optimization strategies, and iterative strategies to continuously generate high-quality labels, and fine-tuned the model through repeated training. Through this series of processes, the final separation model can achieve high-precision uplink and downlink wave separation, effectively preserving the integrity of the target wave signal while suppressing interference and noise, thus improving the model's generalization ability and stability in complex seismic data scenarios.

[0049] Optionally, generating the optimized final dataset through the optimization strategy specifically includes: Drawing inspiration from unsupervised denoising tasks, this approach leverages some of the network's learning capabilities and preferences to achieve the desired outcome.

[0050] Based on the idea in unsupervised denoising tasks, we perform unsupervised learning on the labels of uplink or downlink waves and set the labels to be the same as the input data. In the process of fitting uplink or downlink data to the network, the final dataset is obtained by fitting the data in order of priority based on the data type.

[0051] Unsupervised learning is performed on the labels of the up-wave or down-wave, and the labels are set to be the same as the input data.

[0052] Taking the traveling wave as an example, in the process of fitting the traveling wave data, the network will first learn its main feature - the traveling wave, and then fit the secondary features (such as the traveling wave and noise).

[0053] Therefore, an upward wave model can be obtained through unsupervised training.

[0054] Optionally, the step of filtering the acquired seismic data using a wavefield separation model to obtain the corresponding filtering result specifically includes: By specifying the time-distance curve or linear time-distance curve for correction, the wave field within the target apparent velocity range is mapped to a field consisting only of up-going or down-going waves through time difference correction; then, a multiple up-going and down-going wave separation strategy is adopted to perform precise filtering of the data within a specific velocity range; finally, the original domain space is restored through inverse transformation, thereby achieving the separation and extraction of the target velocity components.

[0055] Specifically, a unified solution is proposed: transforming all velocity difference problems into the task of separating uplink and downlink waves. This strategy allows us to focus on building a high-precision uplink / downlink separation model, thereby significantly improving its applicability and robustness in different application scenarios.

[0056] Figure 10 It demonstrates the process of converting two wave fields with different velocities into upward and downward waves.

[0057] according to Figure 10 The velocity difference between the red and blue waves is used to determine a boundary time-distance curve (e.g., ...). Figure 10 (As shown by the black dashed line in the middle), this velocity lies between the two wave velocities.

[0058] By correcting the time difference, the boundary line is adjusted to a horizontal position, thereby converting the two waves into an upward wave and a downward wave, respectively.

[0059] Subsequently, the trained uplink and downlink separation model is used to separate the corrected data, and finally the separation result is reverse-corrected to restore the original state of the signal.

[0060] This method not only simplifies the processing flow of complex velocity filtering problems, but also provides an efficient and unified solution for wavefield separation in multiple scenarios.

[0061] The red and blue solid lines represent wave components with two different velocities, and the black dashed lines represent the boundary time-distance curves used for linear transformation. Lossless data transformation can be achieved through time difference correction and reverse correction, but this process requires the movement of each sample point to be an integer multiple of the sampling rate, thus limiting the ability to finely divide velocity differences. To obtain more diverse boundary slopes, we introduce a data resampling strategy. Specifically, in the time direction, interpolation techniques are used to reduce the wave slope; in the spatial direction, an equal-interval decimation method is used to increase the wave slope. After data resampling, the formula for calculating the slope of the boundary line is: ; In the formula, B represents the boundary slope on the equally spaced grid, I is the number of interpolations in the time direction (1 indicates no interpolation), X is the number of equally spaced samples in the spatial direction (1 indicates no spatial resampling), and k is the boundary slope after resampling.

[0062] The purpose of this invention is to provide a seismic data velocity filtering algorithm based on deep learning to solve the problems of artifacts, waveform distortion, inaccurate amplitude preservation, low computational efficiency and insufficient applicability of existing velocity filtering methods, and to achieve efficient separation and identification of wavefields with different velocity components in seismic data.

[0063] To achieve the above objectives, the technical solution proposed by this invention mainly includes: This research focuses on constructing a large-scale training dataset and developing a deep learning model. Using the UNet network as the basic framework, the network structure is gradually optimized through systematic experiments and testing. Classic modules such as residual blocks, cascaded modules, and attention mechanisms are introduced, combined with methods like convolutional kernel adjustment, skip connection optimization, and network pruning to refine and lightweight the network design. During training, appropriate hyperparameter settings and efficient loss functions are employed. Simultaneously, a high-quality upwave and downwave separation dataset is constructed as the standard training data for the network, ensuring the model's effectiveness and stability in velocity filtering tasks.

[0064] This invention proposes an arbitrary visual velocity filtering method, developed and optimized for multi-scenario velocity filtering. First, by correcting with a specified time-distance curve or a linear time-distance curve, the target visual velocity range is mapped to the uplink / downlink separation domain. Then, multiple uplink / downlink separation strategies are employed to precisely filter the data within a specific velocity range. Finally, an inverse transformation is used to restore the original domain, achieving the separation and extraction of the target velocity components. In specific application scenarios, this method is suitable for typical problems such as linear noise suppression, VSP uplink / downlink separation, and VSP P-wave / S-wave separation.

[0065] To address the limitation of speed selection under low sampling rate conditions, this invention further introduces data resampling and time interpolation strategies to ensure the accuracy and applicability of the filtering process in different scenarios.

[0066] 1. DAS-VSP wavefield separation: Currently, mainstream DAS-VSP data is still single-component data. For this type of single-component case, conventional wavefield separation techniques relying on multiple components are not directly applicable. In practical processing, the difference in slope between the P-wave and S-wave can be used for differentiation. In near-polarized data, the slopes of the P-wave and S-wave are relatively fixed, while in far-polarized data, their slopes are significantly different from those in near-polarized data. In this embodiment, the arrival times of the waves are first manually calibrated before separation. The original data is as follows: Figure 1 As shown, we picked up the time-distance (TD) curves of the four types of waves and extracted the four types of waves in the order of DOWN-P, DOWN-S, UP-P, and UP-S.

[0067] This process is similar to median filtering, so we also used median filtering to extract the four waves sequentially. Both methods used the same TD curve. The separation results are as follows: Figure 2 As shown. In median filtering, the spatial window size is an important parameter. We tested two window values, 30 and 60. When the spatial window is small, the wavefield cannot be effectively separated, as shown... Figure 2As shown in Figure a, strong transverse waves are clearly residual within the longitudinal waves. When the spatial window is large, the separation of various wave types is relatively good, but this can lead to overly smooth data, and many effective signals still remain in the residual waves, such as... Figure 2 As shown in Figure j, the deep learning separation method proposed in this invention can achieve relatively clean separation of various waves. In the residual wave field, only a small amount of horizontal common-mode noise and a small amount of signals with large slopes are present.

[0068] 2. Imaging arc noise suppression: During the seismic data imaging process, due to complex structures, uneven coverage, or limitations of the imaging algorithm itself, large-angle arc-shaped time noise is often generated in the imaging results. Figure 3 An example is given, with the noise highlighted in red. If stratigraphic data is available, median filtering can be used, employing deep learning to extract the main signals along the stratigraphic trend. If stratigraphic data is unavailable, deep learning methods can be directly used to suppress steep-dipping noise. In this example, the seismic horizon is nearly horizontal, and the structural features are not obvious; therefore, steep-dipping noise was directly suppressed.

[0069] In contrast, the FK filtering method was used. We selected two sector-shaped filter windows: a smaller window to enhance noise suppression and a larger window to preserve as much of the effective signal as possible. The separation results are as follows: Figure 4 As shown in the figure, regardless of the filter window size, FK filtering causes significant loss of effective signal, as indicated by the red line. The red arrows also mark the arc-shaped noise that FK filtering failed to remove. In contrast, the deep learning filtering proposed in this invention preserves more of the effective signal and achieves better noise suppression.

[0070] 3. DAS common-mode noise suppression: DAS-VSP data often exhibits horizontal axis event-type noise, which is highly regular and shows consistent energy distribution across channels. Conventional methods often use superposition to suppress this noise, based on the principle of simulating the common mode distribution through statistical means. This method is generally simple and effective, but in complex field environments, interference or subsequent processing can disrupt the noise's regularity, often leading to inaccurate statistical results and failing to completely remove the noise. In such cases, velocity filtering can be used for suppression.

[0071] Figure 5 An example is shown. The superposition method can suppress most of the noise, such as... Figure 5As shown in b. However, due to the strong energy of shallow shear waves, the statistically derived noise characteristics are not accurate enough, and some common-mode noise is not completely suppressed, as indicated by the red line. Under the same data conditions, using the deep learning filtering method of this invention, common-mode noise is suppressed more thoroughly. Figure 5 Almost no obvious residues were observed in c.

[0072] 4. Common reflection point (CDP) gather optimization: The CDP (Continuous Data Point) gather after data offset typically contains a significant amount of noise, including linear interference and multiples. When the velocity model is accurate, the effective signal is generally flattened, while the noise exhibits characteristics significantly different from the effective signal. With sufficient coverage iterations, stacking can suppress this noise to some extent, but its presence still negatively impacts pre-stack inversion and AVO feature extraction, thus requiring removal. Since multiples in CDP gathers often exhibit parabolic patterns, parabolic Radon transform-like methods are commonly used in production. In this case, the Weighted Least Squares Parabolic Radon Transform (WLSRT) is selected as the comparison method. The denoising results are as follows: Figure 6 As shown. In the parabolic Radon transform method, the maximum time shift at the farthest offset is a key parameter. In this experiment, based on the data sampling and offset range, the maximum time shift was set to 71 milliseconds. The results show that, Figure 6 At the boundary location shown in b, the WLSRT method exhibits a significant energy enhancement phenomenon, a typical drawback of transform domain methods. However, the deep learning method of this invention demonstrates greater stability in the boundary region and superior noise suppression.

[0073] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0074] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0075] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0076] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (Read-Only Memory). Memory includes ROM, magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0077] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for seismic data velocity filtering based on deep learning, characterized in that, The method comprises: S1, an encoder-decoder network based on a U-Net framework, and a residual module and a squeeze-excitation module are introduced to construct a preliminary trained wave field separation model; S2, based on the preliminary trained wave field separation model, a convolution method is used to make standard upgoing waves and downgoing waves, and an iterative optimization strategy is used to optimize the obtained upgoing waves and downgoing waves to obtain a completed data set; S3, the completed data set is divided to obtain a final data block, and the preliminary trained wave field separation model is optimized based on the final data block to obtain a trained wave field separation model; S4, filtering processing is performed on the obtained seismic data through the wave field separation model to obtain a corresponding filtering result.

2. The seismic data velocity filtering processing method based on deep learning according to claim 1, characterized in that, The preliminary trained wave field separation model in S1 specifically comprises: a network encoder, a bottleneck part, a decoder and an outputter; the network encoder comprises four down-sampling stages, each stage comprises two continuous two-dimensional convolution layers, adopts a ReLU activation function, is then connected to the squeeze-excitation module for channel feature re-labeling, and reduces the spatial resolution through a maximum pooling layer; the decoder comprises four up-sampling stages, each stage first improves the spatial resolution through transposed convolution, then performs jump connection splicing with the corresponding encoder features to obtain spliced features; the bottleneck part comprises a residual module composed of two convolution layers, and the spliced features are optimized and reconstructed through the residual module to obtain decoded features; the outputter comprises a two-dimensional convolution layer, which is used to map the decoded features to a target output.

3. The seismic data velocity filtering processing method based on deep learning according to claim 1, characterized in that, The standard upgoing waves and downgoing waves are specifically made as follows: a corresponding type of one-dimensional reflection coefficient sequence is randomly generated based on historical data; a preliminary enhanced one-dimensional reflection coefficient sequence is obtained by increasing the amplitude value of the shallow reflection coefficient based on the one-dimensional reflection coefficient sequence; a final one-dimensional time sequence is obtained by randomly increasing the energy of the selected multiple reflection points based on the preliminary enhanced one-dimensional reflection coefficient sequence; the final one-dimensional time sequence is expanded into a two-dimensional reflection sequence with an upgoing trend or a two-dimensional reflection sequence with a downgoing trend; a two-dimensional reflection coefficient matrix and a wavelet are obtained by increasing data of the obtained two-dimensional reflection sequence with an upgoing trend or a two-dimensional reflection sequence with a downgoing trend; the two-dimensional reflection coefficient matrix and the wavelet are convolved to obtain standard upgoing waves and downgoing waves.

4. The seismic data velocity filtering processing method based on deep learning according to claim 1, characterized in that, The optimization of the obtained upgoing waves and downgoing waves based on the iterative optimization strategy to obtain the completed data set specifically comprises: training a model for separating upgoing waves and downgoing waves based on historical data; detecting the obtained completed data set through the model for separating upgoing waves and downgoing waves to generate a detection result; iterating the completed data set based on the detection result to obtain pure upgoing waves and downgoing waves for making labels; adding the separated pure upgoing waves and downgoing waves, and combining the added data as input features of the network, and generating upgoing wave labels and downgoing wave labels based on the pure upgoing waves; generating the completed data set based on the pure upgoing wave labels and downgoing wave labels.

5. The seismic data velocity filtering processing method based on deep learning according to claim 1, characterized in that, The division of the completed data set to obtain the final data block specifically comprises: The final data set is generated by optimizing the strategy; The final data set is divided into preliminary data blocks by optimization; Each preliminary data block is normalized to generate a data block with consistent feature scales; The final data block is generated based on all data blocks with consistent feature scales.

6. The seismic data velocity filtering processing method based on deep learning according to claim 1, characterized in that, The final data set is generated by optimizing the strategy, specifically including: Based on the idea in the unsupervised denoising task, the uplink or downlink label is learned unsupervisedly, and the label and the input data are set to be the same; In the process of fitting the uplink or downlink data, the final data set is obtained by fitting in turn based on the priority order of the data type.

7. The seismic data velocity filtering processing method based on deep learning according to claim 1, characterized in that, The wave field separation model is used to filter the obtained seismic data to obtain the corresponding filtering result, specifically including: Through the specified time-distance curve or linear time-distance curve correction, the wave field in the target apparent velocity range is mapped to only uplink or downlink field through the time difference correction; then a multiple uplink and downlink separation strategy is used to accurately filter the data in a specific velocity range; finally, through inverse transformation, it is restored to the original domain space to realize the separation and extraction of the target velocity component.

Images