Seismic data velocity modeling method and apparatus, and storage medium

Abstract

The present application relates to a seismic data velocity modeling method and apparatus, and a storage medium, and relates to the technical field of geophysical exploration. The method of the present application comprises: with regard to any seismic trace in a target work area, acquiring or constructing multi-modal data of the seismic trace, the multi-modal data comprising a velocity spectrum of the seismic trace and seismic trace data of the seismic trace extracted from a seismic trace set of the target work area, and the velocity spectrum of the seismic trace being calculated on the basis of the seismic trace set; and using the multi-modal data as an input of a pre-trained velocity prediction model, and using the velocity prediction model to generate a velocity curve of the seismic trace, such that velocity curves of all of seismic traces in the target work area constitute a velocity volume of the target work area. By means of a multi-modal data-driven machine learning method, the method of the present application fast and accurately predicts velocity curves.

Description

Seismic data velocity modeling methods, devices and storage media

[0001] Cross-references to related applications

[0002] This application claims the benefit of Chinese Patent Application No. 202411852941.8, filed on December 16, 2024, the contents of which are incorporated herein by reference. Technical Field

[0003] This application belongs to the field of geophysical exploration technology, specifically relating to a seismic data velocity modeling method, a seismic data velocity modeling device, a computer device, and a machine-readable storage medium. Background Technology

[0004] In the field of geophysical exploration, velocity modeling is a necessary and core step in seismic data processing. Time-domain velocity modeling is inseparable from the identification of velocity spectrum energy clusters in common middle point gathers (CMP), common reflection point gathers (CRP), and residual gathers. Ensuring the accuracy of the velocity field is crucial for subsequent processing. This process typically requires manual picking by experienced professionals, a labor-intensive, time-consuming, and repetitive task, and the results are heavily influenced by subjective factors. Considering the labor and time costs, manual velocity spectrum picking is usually low-density, then interpolated to obtain the velocity of each data point. Furthermore, manual picking is often affected by subjective factors, making it difficult to standardize picking criteria.

[0005] To improve the efficiency of velocity analysis, numerous automatic velocity analysis methods have been proposed. Currently, the main automatic velocity analysis methods are iterative optimization methods. These methods are inversion methods that use optimization algorithms and maximum similarity strategies to invert the optimal solution of the velocity curve in the velocity spectrum. These methods typically assume linear velocity, but noise and other interference signals highlight their nonlinearity. Furthermore, the iterative computation is large, and the picking efficiency still needs improvement, making these methods difficult to apply industrially. In recent years, with the significant improvement in computer performance, machine learning has also been applied to automatic velocity analysis, and unsupervised clustering methods have been widely adopted. Unsupervised clustering methods can search for seismic data features without labeling, grouping data with similar features to obtain an approximate distribution of the dataset. These methods have simple algorithms, are easy to implement, have low computational cost, high interpretability, and are applicable to seismic data with any characteristics, thus making them more widely applicable in industrial settings. However, unsupervised clustering algorithms often only consider the ability to track energy clusters in the velocity spectrum, without the constraint of prior geological background knowledge. This leads to the velocity analysis being affected by multiple interference signals and energy clusters, reducing the accuracy of the velocity analysis.

[0006] Therefore, there is an urgent need to improve the automatic speed analysis method. Summary of the Invention

[0007] The purpose of this application is to provide a seismic data velocity modeling method, a seismic data velocity modeling device, a computer device, and a machine-readable storage medium to overcome one or more defects in existing velocity analysis methods. The defects in existing velocity analysis methods include at least one of poor velocity picking accuracy and low velocity picking efficiency.

[0008] To achieve the above objectives, a first aspect of this application provides a seismic data velocity modeling method, comprising: acquiring or constructing multimodal data of any seismic trace in a target work area, the multimodal data including the velocity spectrum of the seismic trace and seismic trace data of the seismic trace extracted from a seismic trace set in the target work area, the velocity spectrum of the seismic trace being calculated based on the seismic trace set; using the multimodal data as input to a pre-trained velocity prediction model, and using the velocity prediction model to generate velocity curves of the seismic trace, so that the velocity curves of all seismic traces in the target work area constitute the velocity volume of the target work area; wherein, the velocity prediction model is obtained by training an initial machine learning model, the velocity prediction model is used to extract features from each modal data in the multimodal data, fuse all extracted features, and then predict velocity curves from the fused features, the features extracted from each modal data including energy cluster features extracted from the velocity spectrum and hyperbolic features extracted from the seismic trace data.

[0009] In a specific embodiment of this application, the velocity prediction model includes: a first feature extraction network for extracting energy cluster features from the velocity spectrum; a second feature extraction network for extracting hyperbolic features from the seismic trace data; a fusion network for fusing the energy cluster features with the hyperbolic features; and a sequence prediction network for predicting velocity sequences based on the feature maps obtained by the fusion network, thereby generating the velocity curve of the seismic trace.

[0010] In a specific embodiment of this application, training an initial machine learning model to obtain the velocity prediction model includes: using manually picked velocity curves as manual labels to train the initial machine learning model to obtain an initial velocity prediction model; using the current velocity prediction model to predict the velocity curves of each seismic trace in the historical seismic trace set; for any seismic trace in the historical seismic trace set, using the velocity curve generated by the current velocity prediction model as a priori constraint, segmenting the candidate region where the energy cluster is located from the velocity spectrum of the seismic trace, and tracking the position of the energy cluster with the maximum energy in the candidate region through an unsupervised clustering method to obtain the corrected velocity curve of the seismic trace; using the corrected velocity curves of each seismic trace and the manual labels to train the current velocity prediction model, and using the velocity prediction model when the predicted velocity curves meet the preset requirements as the final velocity prediction model.

[0011] In a specific embodiment of this application, the velocity curve generated by the current velocity prediction model is used as a priori constraint to segment the candidate region where the energy cluster is located from the velocity spectrum of the seismic trace. This includes: constructing the initial evolution curve of the Chan-vese segmentation model using the velocity curve generated by the current velocity prediction model; and segmenting the velocity spectrum of the seismic trace using the Chan-vese segmentation model to identify the candidate region where the energy cluster is located.

[0012] In a specific embodiment of this application, the construction of the initial evolution curve of the Chan-vese segmentation model using the velocity curve generated by the current velocity prediction model includes: shifting the velocity curve generated by the current velocity prediction model along the negative direction of the velocity coordinate axis by a first preset distance, and using the shifted velocity curve as the first initial evolution curve; shifting the velocity curve generated by the current velocity prediction model along the positive direction of the velocity coordinate axis by a second preset distance, and using the shifted velocity curve as the second initial evolution curve; the first initial evolution curve and the second initial evolution curve constitute the initial evolution curve of the Chan-vese segmentation model; wherein, the determination rule for the first preset distance and the second preset distance includes at least ensuring that the initial evolution curve includes all energy clusters.

[0013] In a specific embodiment of this application, the correction velocity curve of the seismic trace is obtained by tracking the position of the maximum energy of the energy cluster in the candidate region using an unsupervised clustering method. This includes: tracking the position of the maximum energy of the energy cluster in the candidate region using a mean-shift clustering method; interpolating the velocity curve of the seismic trace based on each energy point at the position of the maximum energy of the energy cluster; and using the velocity curve as the correction velocity curve.

[0014] In a specific embodiment of this application, before using the speed prediction model when the predicted speed curve meets the preset requirements as the final speed prediction model, the method further includes: adjusting the speed prediction model when the predicted speed curve meets the preset requirements using the artificial labels.

[0015] In a specific embodiment of this application, the method further includes: performing median filtering and two-dimensional Sobel filtering on the velocity spectrum in the multimodal data to trigger the execution of using the multimodal data as input to a pre-trained velocity prediction model, and using the velocity prediction model to generate the velocity curve of the seismic trace.

[0016] In a specific embodiment of this application, when training the current velocity prediction model using the corrected velocity curves and artificial labels of each seismic trace, the energy cluster features are superimposed with the historical seismic trace set for imaging, and seismic trace data of a specific seismic trace in the imaging profile obtained by superimposing the imaging are extracted. Then, the extracted seismic trace data of the specific seismic trace is fused with the feature map obtained by the fusion network and input into the sequence prediction network. The specific seismic trace is the seismic trace where the current seismic trace data input into the second feature extraction network during training is located.

[0017] A second aspect of this application provides a seismic data velocity modeling apparatus, comprising: a data acquisition module, configured to acquire or construct multimodal data of any seismic trace in a target work area, the multimodal data including the velocity spectrum of the seismic trace and seismic trace data extracted from a seismic trace set in the target work area, the velocity spectrum of the seismic trace being calculated based on the seismic trace set; and a velocity curve prediction module, configured to use the multimodal data as input to a pre-trained velocity prediction model, and generate velocity curves of the seismic trace using the velocity prediction model, so that the velocity curves of all seismic traces in the target work area constitute the velocity volume of the target work area; wherein the velocity prediction model is obtained by training an initial machine learning model, the velocity prediction model is used to extract features from each modal data in the multimodal data, fuse all extracted features, and then predict velocity curves from the fused features, the features extracted from each modal data including energy cluster features extracted from the velocity spectrum and hyperbolic features extracted from the seismic trace data.

[0018] A third aspect of this application provides a computer device comprising: a memory configured to store instructions; and a processor configured to retrieve the instructions from the memory and, when executing the instructions, to implement the seismic data velocity modeling method according to a first aspect of this application.

[0019] A fourth aspect of this application provides a machine-readable storage medium storing instructions that cause a machine to perform the seismic data velocity modeling method according to a first aspect of this application.

[0020] The above technical solution uses multimodal data as a driving machine learning method to predict velocity curves. The multimodal data includes seismic trace data extracted from the seismic trace set of the target work area and velocity spectrum calculated based on the seismic trace set. By identifying the velocity spectrum characteristics and seismic trace characteristics in the seismic data, the geological characteristics and energy distribution characteristics of the seismic data are comprehensively considered, thereby improving the accuracy of the predicted velocity curves and realizing efficient and accurate seismic data velocity modeling.

[0021] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings:

[0023] Figure 1 schematically illustrates a flowchart of a seismic data velocity modeling method according to an embodiment of this application;

[0024] Figure 2 schematically illustrates the network architecture of the speed prediction model used in a specific example;

[0025] Figure 3 schematically illustrates a flowchart of the speed prediction model construction process according to an embodiment of this application;

[0026] Figure 4 schematically shows the velocity spectrum of a CMP gather from a three-dimensional land data source in the Sichuan Basin.

[0027] Figure 5 schematically illustrates the initial prediction speed curves of a multimodal neural network trained with a small number of artificial labels;

[0028] Figure 6 schematically shows the initial evolution curve of the Chan-vese segmentation model constructed based on the velocity curve in Figure 5;

[0029] Figure 7 schematically illustrates the results of iterative evolution of the initial evolution curve in Figure 6 using the Chan-vese segmentation model;

[0030] Figure 8 schematically shows the velocity curves obtained by unsupervised clustering and interpolation of the energy clusters in Figure 7 at their maximum energy locations;

[0031] Figure 9 schematically shows the velocity curves predicted using the final velocity prediction model;

[0032] Figure 10 schematically illustrates a time-domain superimposed velocity volume according to an embodiment of this application;

[0033] Figure 11 schematically illustrates a time-domain offset velocity volume according to an embodiment of this application;

[0034] Figure 12 schematically illustrates the time-domain residual velocity volume according to an embodiment of this application;

[0035] Figure 13 schematically shows the time-domain offset velocity volume obtained after updating using the time-domain residual velocity volume of Figure 12;

[0036] Figure 14 schematically shows a cross-section of a time-domain velocity volume migration imaging method as shown in Figure 13;

[0037] Figure 15 schematically shows the offset imaging profile of the corresponding manually created time-domain offset velocity volume;

[0038] Figure 16 schematically illustrates the depth-domain velocity volume obtained by transforming Figure 13;

[0039] Figure 17 schematically illustrates a block diagram of a seismic data velocity modeling apparatus according to an embodiment of this application;

[0040] Figure 18 schematically illustrates a structural block diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0041] The specific embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the embodiments of this application.

[0042] If the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0043] In seismic data processing, high-precision velocity models are often required for migration imaging. Accordingly, this application provides a method for seismic data velocity modeling.

[0044] The method provided in this application performs velocity modeling in the following manner: for any seismic trace in the target work area, acquire or construct multimodal data of the seismic trace, use the multimodal data as input to a pre-trained velocity prediction model, and use the velocity prediction model to generate the velocity curve of the seismic trace, so as to form the velocity volume of the target work area from the velocity curves of all seismic traces in the target work area.

[0045] Specifically, the multimodal data includes the velocity spectrum of the seismic trace and seismic trace data extracted from the seismic trace set of the target work area. The velocity spectrum of the seismic trace is calculated based on the seismic trace set of the target work area. The velocity prediction model is obtained by training an initial machine learning model. The velocity prediction model is used to extract features from each modality of the multimodal data, fuse all extracted features, and then predict velocity curves based on the fused features. The features extracted from each modality include energy cluster features extracted from the velocity spectrum and hyperbolic features extracted from the seismic trace data.

[0046] It should be understood that in this application, the machine learning model can be constructed based on a multimodal neural network. When calculating the velocity spectrum based on the seismic gathers of the target work area, the calculation methods in the conventional embodiments can be combined; however, this application will not describe this part in detail.

[0047] The above embodiments of this application have at least the following beneficial effects:

[0048] 1) Compared with time-domain velocity analysis methods based on unsupervised clustering algorithms, unsupervised clustering methods only consider the tracking ability of velocity spectrum energy clusters without the constraint of prior geological background knowledge. This makes the velocity analysis process susceptible to multiple waves and other interference signals, making it difficult to improve the accuracy of velocity analysis. In contrast, the above embodiments of this application adopt a data and prior geological knowledge-driven method, which can fully consider the features related to velocity curves in seismic data and geological understanding. Based on the extraction of these features, the velocity curve is predicted, and the accuracy of velocity analysis can be significantly improved.

[0049] 2) In one comparative embodiment, local images of the CMP seismic gather requiring NMO correction are used as input to a convolutional neural network (CNN), and velocity is used as the output. This approach considers a single data source, resulting in poor accuracy of the predicted velocity curve. In another comparative embodiment, the velocity spectrum calculated from the seismic gather is used as input to the CNN, and velocity is used as the output. This approach also suffers from the problem of a single data source, leading to poor accuracy of the predicted velocity curve. In the embodiments described above in this application, to fully consider the characteristics of the velocity spectrum and seismic gather data, a machine learning method driven by multimodal data is used to predict the velocity curve. The multimodal data includes seismic trace data extracted from the seismic gather of the target work area and the velocity spectrum calculated from the seismic gather. This fully considers the geological and energy characteristics in the seismic data, thereby obtaining a more accurate velocity curve through richer feature extraction.

[0050] In summary, the embodiments described above overcome the influence of multiple waves and other interference signals, and achieve accurate velocity modeling of seismic data using multimodal data.

[0051] Figure 1 schematically illustrates a flowchart of a seismic data velocity modeling method according to an embodiment of this application. As shown in Figure 1, the seismic data velocity modeling method provided in this embodiment includes steps 102 to 106. It is understood that the above-described seismic data velocity modeling method may include all steps 102 to 106, or may include only some of them. As an optional embodiment of this application, a seismic data velocity modeling method may include only steps 102 and 104; in another optional embodiment, the seismic data velocity modeling method may include steps 102, 104, and 106.

[0052] Step 102: For any seismic trace in the target work area, construct multimodal data of the seismic trace. The multimodal data includes the velocity spectrum of the seismic trace and the seismic trace data of the seismic trace extracted from the seismic trace set of the target work area.

[0053] As an example, the multimodal data of the seismic trace can be constructed by: extracting the seismic trace data from the seismic trace set of the target work area; and calculating the velocity spectrum of the seismic trace based on the seismic trace set of the target work area.

[0054] As an example, seismic gathers can be any of the CMP gathers, CRP gathers, and residual gathers, and these gathers can be obtained after processing conventional seismic data. However, it should be understood that high-quality velocity spectra can improve the accuracy of subsequent velocity modeling, so these gathers can also be of improved quality. The improvement of seismic gather quality has been widely disclosed in the prior art, and this application will not describe this part in detail.

[0055] Step 104: Use the multimodal data of the seismic trace as input to the pre-trained velocity prediction model, and use the velocity prediction model to generate the velocity curve of the seismic trace.

[0056] Specifically, in this application, the velocity prediction model is obtained by training an initial machine learning model. The velocity prediction model is used to extract features from each modality of multimodal data, fuse all extracted features, and then use the fused features to predict the velocity curve.

[0057] It is important to understand that predicting the velocity curve of the fused features specifically means predicting the "time-velocity" sequence of the fused features to obtain the velocity curve.

[0058] Step 106: The time-domain velocity volume of the target work area is constructed from the velocity curves of all seismic traces in the target work area.

[0059] As an example, the velocity curve generated by the velocity prediction model can be any one of the time-domain superimposed velocity curve, the time-domain offset velocity curve, and the time-domain residual velocity curve. Correspondingly, the time-domain velocity volume of the target work area can be any one of the time-domain superimposed velocity volume, the time-domain offset velocity volume, and the time-domain residual velocity volume.

[0060] As an improvement to the above embodiments of this application, the following steps are included before step 104:

[0061] Step 103: Perform median filtering on the velocity spectrum in the multimodal data to remove salt-and-pepper noise, and perform two-dimensional Sobel filtering on the velocity spectrum in the multimodal data to characterize the energy trend band.

[0062] As an optional embodiment of this application, the velocity prediction model includes a first feature extraction network, a second feature extraction network, a fusion network, and a sequence prediction network, wherein: the first feature extraction network is used to extract energy cluster features from the velocity spectrum; the second feature extraction network is used to extract hyperbolic features from the seismic trace data; the fusion network is used to fuse the energy cluster features and the hyperbolic features; and the sequence prediction network is used to predict the velocity sequence based on the feature map obtained by the fusion network and generate the velocity curve of the seismic trace.

[0063] As an example, the first feature extraction network includes a first CNN (Convolutional Neural Networks) unit, a second CNN unit, a first CNN-Transformer unit, and a second CNN-Transformer unit, all connected in sequence with increasing network depth. The second feature extraction network includes a third CNN unit, a fourth CNN unit, a fifth CNN unit, a sixth CNN unit, a seventh CNN unit, an eighth CNN unit, and a pooling unit, all connected in sequence with increasing network depth. The fusion network includes a ninth CNN unit and a third CNN-Transformer unit, all connected in sequence with increasing network depth. As an example, the pooling unit is a max-pooling unit.

[0064] Figure 2 schematically illustrates the network architecture of the speed prediction model used in the above example. As shown in Figure 2, in the velocity prediction model: the first feature extraction network at the top receives the velocity spectrum of the three channels. From left to right, the first CNN refers to the first CNN unit, the second CNN refers to the second CNN unit, the first CNNTrans refers to the first CNN-Transformer unit, and the second CNNTrans refers to the second CNN-Transformer unit. The second feature extraction network at the bottom receives seismic trace data from the three-channel trace gather data. From left to right, the first CNN refers to the third CNN unit, the second CNN refers to the fourth CNN unit, the third CNN refers to the fifth CNN unit, the fourth CNN refers to the sixth CNN unit, the fifth CNN refers to the seventh CNN unit, and the sixth CNN refers to the eighth CNN unit. Max pooling represents maximum pooling. In the fusion network, from left to right, the first CNN refers to the ninth CNN unit, and the first CNNTrans refers to the third CNN-Transformer unit. In the sequence prediction network, GRU (Gated Recurrent Unit) refers to the GRU unit. Vrms refers to the output of the velocity prediction model, i.e., the velocity curve.

[0065] The principle of the aforementioned velocity prediction model is as follows: by using feature extraction and backpropagation algorithms of a certain depth, an optimal nonlinear mapping relationship is established between the velocity spectrum, trace gather data, and velocity labels. After obtaining the trace gather and velocity spectrum of the earthquake to be predicted, this relationship is used to predict the velocity curve. However, in existing technologies, supervised neural networks require manual collection of a large amount of rich label data for network training, which is time-consuming, labor-intensive, and results in weak model generalization performance.

[0066] To overcome the above-mentioned deficiencies, in a specific embodiment of this application, an improved training method is proposed. The improved training method trains an initial machine learning model to obtain a velocity prediction model in the following manner: using manually picked velocity curves as manual labels to train the initial machine learning model to obtain an initial velocity prediction model; and using the current velocity prediction model to predict the velocity curves of each seismic trace in the historical seismic trace set.

[0067] For any seismic trace in the historical seismic trace set, the velocity curve generated by the current velocity prediction model is used as a priori constraint. Candidate regions containing energy clusters are segmented from the velocity spectrum of the seismic trace. The position of the energy cluster with the maximum energy in the candidate region is tracked by an unsupervised clustering method to obtain the corrected velocity curve of the seismic trace. The current velocity prediction model is trained using the corrected velocity curves of each seismic trace and artificial labels. The velocity prediction model that meets the preset requirements is used as the final velocity prediction model.

[0068] As in the above embodiment, a small number of labels are manually picked up for initial training of the machine learning model to obtain an initial velocity prediction model. Then, a combination of energy cluster candidate region segmentation based on prior constraints and unsupervised clustering is used. Specifically, the velocity curve predicted by the initially trained velocity prediction model is used as a prior constraint to segment candidate regions where energy clusters are located in the velocity spectrum. An unsupervised clustering method is then used to track the location of the energy cluster with the maximum energy, thus obtaining a new velocity curve. This new velocity curve is more accurate than the velocity curve predicted by the initially trained model; therefore, it is a corrected velocity curve. This corrected velocity curve is used as a new velocity label, expanding the label sample. The velocity prediction model is then trained again, resulting in the final velocity prediction model. It is evident that the improved training method proposed in the above embodiment, based on expert geological knowledge, saves manpower and improves model training efficiency. Simultaneously, the combination of energy cluster candidate region segmentation based on prior constraints and unsupervised clustering achieves better picking efficiency and accuracy than existing velocity curve picking methods based on adaptive thresholds and unsupervised clustering. The superior label samples also improve the accuracy of the trained velocity prediction model.

[0069] As an example, the machine learning model employs a multimodal neural network. Figure 3 schematically illustrates a flowchart of the speed prediction model construction process according to an embodiment of this application. As shown in Figure 3, training the initial multimodal neural network to obtain the speed prediction model may include steps 202 to 212. It is understood that the above-described speed prediction model construction process may include all steps 202 to 212, or may include only some of them.

[0070] Step 202: Obtain historical earthquake gathers and calculate the velocity spectrum of each earthquake trace in the historical earthquake gathers.

[0071] Step 204: Manually pick the velocity curves of the first preset number of seismic traces in the velocity spectrum.

[0072] Step 206: Use the manually picked speed curves as manual labels to train the initial multimodal neural network to obtain the initial speed prediction model.

[0073] Step 208: Use the current velocity prediction model to predict the velocity curves of each seismic trace in the historical seismic trace set.

[0074] Step 210: For any seismic trace in the historical seismic trace set, the velocity curve generated by the current velocity prediction model is used as a priori constraint. Candidate regions where energy clusters are located are segmented from the velocity spectrum of the seismic trace. The position of the energy cluster with the maximum energy in the candidate region is tracked by an unsupervised clustering method to obtain the corrected velocity curve of the seismic trace.

[0075] Step 212: Train the current velocity prediction model using the corrected velocity curves of each seismic trace and artificial labels, and use the velocity prediction model that meets the preset requirements as the final velocity prediction model.

[0076] In this application, the candidate region where the energy cluster is located refers to the velocity candidate region.

[0077] It should be understood that the value of the first preset number is much smaller than the total number of historical seismic traces. It can be set and adjusted according to the specific application.

[0078] As an example, the current velocity prediction model is trained using the corrected velocity curves of each seismic trace and artificial labels. The velocity prediction model that meets the preset requirements is taken as the final velocity prediction model. Specifically, steps 208 to 210 are repeated iteratively until the velocity prediction model's velocity curve prediction results for each seismic trace meet the preset accuracy requirements.

[0079] As an improvement to the above embodiments of this application, the following steps are included before step 204:

[0080] Step 203: Perform median filtering on the velocity spectrum to remove salt-and-pepper noise, and perform two-dimensional Sobel filtering on the velocity spectrum to characterize the energy trend band.

[0081] As an improvement to the above embodiments of this application, in step 212, before using the speed prediction model when the predicted speed curve meets the preset requirements as the final speed prediction model, manual labels are used to adjust the speed prediction model when the predicted speed curve meets the preset requirements.

[0082] As a preferred embodiment of the present application, the velocity curve generated by the current velocity prediction model is used as a priori constraint to segment candidate regions for energy cluster picking from the velocity spectrum of the seismic trace. Specifically, this includes the following steps: constructing the initial evolution curve of the Chan-vese segmentation model using the velocity curve generated by the current velocity prediction model; segmenting the velocity spectrum of the seismic trace using the Chan-vese segmentation model to identify the candidate regions where the energy clusters are located.

[0083] Specifically, in this application, the candidate region where the energy cluster is located refers to: using the Chan-vese segmentation model to segment the velocity spectrum, identifying the energy cluster of the effective signal, and then outputting the velocity candidate region based on the location of the energy cluster.

[0084] The Chan-vese segmentation algorithm is used to segment objects without well-defined boundaries. This algorithm is based on iterative evolution to minimize a set of energy levels, defined by the sum of differences in intensity corresponding to the average values ​​from outside the segmentation region, the sum of differences in intensity corresponding to the average values ​​from inside the segmentation region, and a term dependent on the length of the segmentation region's boundary. The above embodiments construct a Chan-vese segmentation model based on the Chan-vese segmentation algorithm and utilize the velocity curve generated by the current velocity prediction model to construct the initial evolution curve, avoiding interference from noise and multiple wave energy clusters, ensuring the segmentation accuracy of the Chan-vese segmentation model, and achieving accurate identification of effective signal energy clusters and precise segmentation of candidate regions. It will be understood by those skilled in the art that other known and disclosed image segmentation methods can also be applied to this application, and the above embodiments are not the only limitation on the image segmentation methods that can be used in this application.

[0085] As an example, the initial evolution curve of the Chan-vese segmentation model is constructed as follows: the velocity curve generated by the current velocity prediction model is shifted along the negative direction of the velocity coordinate axis by a first preset distance, and the shifted velocity curve is used as the first initial evolution curve.

[0086] The velocity curve generated by the current velocity prediction model is translated along the positive direction of the velocity coordinate axis by a second preset distance, and the translated velocity curve is used as the second initial evolution curve; the first initial evolution curve and the second initial evolution curve constitute the initial evolution curve of the Chan-vese segmentation model; wherein, the rules for determining the first preset distance and the second preset distance include at least ensuring that the initial evolution curve includes all energy clusters.

[0087] As a preferred embodiment of the present application, the location of the maximum energy of the energy cluster in the candidate region is tracked in the following manner to obtain the corrected velocity curve of the seismic trace: the location of the maximum energy of the energy cluster in the candidate region is tracked using the mean-shift clustering method; the velocity curve of the seismic trace is obtained by interpolation based on each energy point of the location of the maximum energy of the energy cluster, and the velocity curve is used as the corrected velocity curve.

[0088] In the above embodiments, mean-shift clustering improves the efficiency and accuracy of energy cluster picking, with superior picking accuracy compared to K-means picking under the same threshold screening conditions. This improves the accuracy of the label data and leads to a more accurate velocity prediction model. It will be understood by those skilled in the art that other known and disclosed unsupervised clustering methods can also be applied to this application, and the above embodiments are not the only limitation on the unsupervised clustering methods that can be used in this application.

[0089] As an example, after tracing the energy points at the locations of maximum energy in an energy cluster, linear spline interpolation can be used to obtain the velocity curve of the seismic trace.

[0090] As an example, the principle of the unsupervised velocity picking method based on mean-shift clustering is explained below:

[0091] The velocity spectrum energy clusters identified by the Chan-Vese segmentation model can be viewed as regions with the highest density of velocity spectrum energy points. The mean-shift algorithm is a hill-climbing algorithm based on kernel density estimation. By continuously updating the pick-up points, it eventually reaches the region with the highest energy point density, which is the corresponding energy cluster pick-up position. The basic form of the mean-shift algorithm can be expressed as:

[0092] In formula (1), x represents the velocity spectrum energy point (th, vh), Dh is a circle with x as the center and h as the radius, xi is the energy point (ti, vi), i = 1, 2, ..., k, k is the number of energy points within the range of Dh, and Mh represents the drift amount. Then the formula for updating the center x is:

[0093] x * =x+M h (2);

[0094] Formulas (1) and (2) allow the center x to continuously move towards areas with high energy point density in the velocity spectrum. Updates stop when the drift Mh is less than the cutoff threshold. Formula (1) shows that the contribution of all energy points within Dh is the same when calculating the drift, but the sizes of energy points in the velocity spectrum are often different, resulting in different contributions to the drift. Higher energy contributes more, while energy points farther away contribute less. To improve the accuracy of velocity picking, this application uses an improved drift calculation method:

[0095] In formula (3), G is the Gaussian kernel function; h represents the bandwidth, which is a diagonal matrix. In this application, it is simplified to a scalar h, representing the radius of the computational region Dh; ω(x i ) represents the weight, i.e., the energy point x. i The magnitude of energy. Substituting formula (3) into formula (2) yields a new updated formula:

[0096] The center x is iteratively updated using formula (4) until it moves to the position with the highest local energy point density in the velocity spectrum, which is the current energy cluster picking position. Mean drift is continued to be used for energy points that are not utilized during the iteration until all energy points in the velocity spectrum are clustered, and the energy cluster picking ends.

[0097] As an improvement to the above embodiments of this application, this application proposes a method for assisting iterative training of a velocity prediction model. Specifically, when training the current velocity prediction model using the corrected velocity curves of each seismic trace and artificial labels,

[0098] The energy cluster features extracted by the first feature extraction network are superimposed with historical seismic traces for imaging, and the seismic trace data of specific seismic traces in the imaging profile obtained by superimposing the images are extracted. Then, the extracted seismic trace data of specific seismic traces are fused with the feature map obtained by the fusion network and input into the sequence prediction network. The specific seismic trace is the seismic trace where the current seismic trace data input into the second feature extraction network during training is located.

[0099] In a specific application, as shown in Figure 2, the extracted seismic trace data of a specific seismic trace is fused with the feature map output by the third CNN-Transformer unit and then input into the GRU unit.

[0100] As in the above embodiment, during the training of the velocity prediction model, the extracted seismic trace data of a specific seismic trace is used as input data into the velocity prediction model. This is because the extracted seismic trace data of the specific seismic trace contains velocity information, which can correct the velocity curve output of the velocity prediction model. Through this kind of auxiliary training, the training of the velocity prediction model is accelerated and the accuracy of the trained velocity prediction model is improved.

[0101] In a specific application example of this application, the seismic data velocity modeling method provided in the above embodiments is used to perform velocity modeling of a three-dimensional land actual seismic data in the Sichuan Basin. The velocity volume obtained by modeling includes a time-domain stacked velocity volume, a time-domain migrated velocity volume, and a time-domain residual velocity volume.

[0102] As shown in Figures 4 to 16, the velocity modeling process in this application example includes steps A1 to A12.

[0103] Step A1: Calculate the velocity spectrum of each seismic trace based on the CMP / CRP / residual gather obtained from conventional seismic data processing.

[0104] Step A2: Perform median filtering on the velocity spectrum to remove random salt-and-pepper noise.

[0105] Step A3: Perform two-dimensional Sobel filtering on the velocity spectrum energy clusters to characterize the energy trend bands.

[0106] Step A4: Manually collect a small number of velocity labels to train the multimodal neural network.

[0107] Step A5: Use a multimodal neural network to make a preliminary prediction of the velocity spectrum velocity curve.

[0108] Step A6: Construct the initial evolution curve of the Chan-vese segmentation model using the predicted velocity curve.

[0109] Step A7: Use the Chan-vese segmentation model to segment the velocity spectrum and identify effective signal energy clusters.

[0110] Step A8: Use the mean-shift clustering method to track energy clusters and interpolate to obtain new velocity curves.

[0111] Step A9: The new velocity curve is retrained together with the artificial velocity label to form a multimodal neural network.

[0112] Step A10: Iterate and repeat steps A5 to A9 until the prediction results of the multimodal neural network meet the accuracy requirements.

[0113] Step A11: Fine-tune the multimodal neural network with manual labels to obtain the final speed prediction model.

[0114] Step A12: Use the velocity prediction model to predict the velocity field of the seismic data to be analyzed in the entire time domain.

[0115] Figure 4 shows the velocity spectrum of a CMP gather from a 3D land data source in the Sichuan Basin. This is a single-trace velocity spectrum obtained after denoising and other processing methods. The velocity spectra of the CRP gather and the remaining gathers are similar. Figure 5 shows the initial velocity curve predicted by a multimodal neural network trained with a small number of artificial labels. The neural network has learned expert geological background knowledge. Figure 6 shows the initial evolution curve of the Chan-vese segmentation model constructed based on the initial velocity curve predicted by the multimodal neural network in Figure 5. Regions outside the curve are not included in the velocity analysis, effectively avoiding interference from noise energy and improving the segmentation efficiency of the Chan-vese model. Figure 7 shows the result of iterative evolution of the initial evolution curve in Figure 6 using the Chan-vese segmentation model. The effective energy clusters in the velocity spectrum are accurately extracted. Figure 8 shows the velocity curve obtained by unsupervised clustering and interpolation of the energy clusters with maximum energy in Figure 7. Figure 9 shows the velocity curve predicted using the final velocity prediction model. Figure 10 is a schematic diagram of the time-domain superimposed velocity volume according to an embodiment of this application. Figure 11 is a time-domain offset velocity volume according to an embodiment of this application. Figure 12 is a time-domain remaining velocity volume according to an embodiment of this application. Figure 13 shows the time-domain migrated velocity volume obtained after updating using the time-domain residual velocity volume of Figure 12, further improving the accuracy of the velocity volume. Figure 14 shows the cross-section of the migration imaging using the time-domain migrated velocity volume of Figure 13. Figure 15 shows the migration imaging cross-section of the corresponding manually established time-domain migrated velocity volume. Comparing Figures 14 and 15, it can be seen that the imaging results of this application are generally consistent with the results of detailed manual analysis, with improved local imaging accuracy, especially in the clearer depiction of shallow fractures and the more reasonable morphology of medium-deep buried hill structures. Figure 16 shows the depth-domain velocity volume obtained by converting from Figure 13.

[0116] Corresponding to the seismic data velocity modeling method in the above embodiments, Figure 17 schematically shows a block diagram of the seismic data velocity modeling apparatus of this application embodiment. For ease of explanation, only the parts related to the embodiments of this application are shown. As shown in Figure 17, the seismic data velocity modeling device 400 includes: a data acquisition module 410, used to acquire or construct multimodal data of any seismic trace in the target work area, the multimodal data including the velocity spectrum of the seismic trace and seismic trace data extracted from the seismic trace set of the target work area, the velocity spectrum of the seismic trace being calculated based on the seismic trace set; and a velocity curve prediction module 420, used to take the multimodal data as input to a pre-trained velocity prediction model, and use the velocity prediction model to generate the velocity curve of the seismic trace, so that the velocity curves of all seismic traces in the target work area constitute the velocity volume of the target work area; wherein, the velocity prediction model is obtained by training an initial machine learning model, the velocity prediction model is used to extract features from each modality of the multimodal data, fuse all the extracted features, and then predict the velocity curve from the fused features, the features extracted from each modality of the data including energy cluster features extracted from the velocity spectrum and hyperbolic features extracted from the seismic trace data.

[0117] As one embodiment of this application, the seismic data velocity modeling device 400 can implement the embodiment shown in FIG1 and other related method embodiments.

[0118] The process by which each module of the seismic data velocity modeling device 400 provided in this application implements its respective function can be specifically referred to in the description of the embodiment shown in Figure 1 above and other related method embodiments, and will not be repeated here.

[0119] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here.

[0120] Figure 18 schematically illustrates a structural block diagram of a computer device according to an embodiment of this application. In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown in Figure 18. The computer device includes a processor A01, a network interface A02, a display screen A04, an input device A05, and a memory (not shown) connected via a system bus. The processor A01 of the computer device provides computing and control capabilities. The memory of the computer device includes internal memory A03 and a non-volatile storage medium A06. The non-volatile storage medium A06 stores an operating system B01 and a computer program B02. The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 in the non-volatile storage medium A06. The network interface A02 of the computer device is used for communication with an external terminal via a network connection. When the computer program is executed by the processor A01, it implements a seismic data velocity modeling method. The display screen A04 of the computer device can be an LCD screen or an e-ink screen. The input device A05 of the computer device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the casing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0121] Those skilled in the art will understand that the structure shown in Figure 18 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0122] In one embodiment, the seismic data velocity modeling apparatus 400 provided in this application can be implemented as a computer program, which can run on a computer device as shown in FIG18. The memory of the computer device can store various program modules constituting the seismic data velocity modeling apparatus 400. The computer program, composed of the various program modules, causes a processor to execute the steps in the seismic data velocity modeling methods of the various embodiments of this application described in this specification.

[0123] In one embodiment, this application also provides a machine-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the seismic data velocity modeling method in the above embodiments.

[0124] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

Claims

1. A method for seismic data velocity modeling, characterized in that, include: For any seismic trace in the target work area, acquire or construct multimodal data of the seismic trace. The multimodal data includes the velocity spectrum of the seismic trace and seismic trace data of the seismic trace extracted from the seismic trace set of the target work area. The velocity spectrum of the seismic trace is calculated based on the seismic trace set. The multimodal data is used as input to a pre-trained velocity prediction model, which is then used to generate velocity curves for the seismic traces, so that the velocity volume of the target work area can be formed from the velocity curves of all seismic traces in the target work area. The velocity prediction model is obtained by training an initial machine learning model. The velocity prediction model is used to extract features from each modal data in the multimodal data, fuse all the extracted features, and then predict the velocity curve based on the fused features. The features extracted from each modal data include energy cluster features extracted from the velocity spectrum and hyperbolic features extracted from the seismic trace data.

2. The seismic data velocity modeling method according to claim 1, characterized in that, The speed prediction model is obtained by training an initial machine learning model, including: The manually collected speed curves are used as manual labels to train the initial machine learning model, resulting in an initial speed prediction model. Predict velocity curves for each seismic trace in the historical seismic trace set using the current velocity prediction model; For any seismic trace in the historical seismic trace set, the velocity curve generated by the current velocity prediction model is used as a priori constraint. Candidate regions where energy clusters are located are segmented from the velocity spectrum of the seismic trace. The position of the energy cluster with the maximum energy in the candidate region is tracked by an unsupervised clustering method to obtain the corrected velocity curve of the seismic trace. The current velocity prediction model is trained using the corrected velocity curves and artificial labels of each seismic trace. The velocity prediction model that meets the preset requirements is taken as the final velocity prediction model.

3. The seismic data velocity modeling method according to claim 2, characterized in that, The velocity prediction model includes: A first feature extraction network is used to extract energy cluster features from the velocity spectrum; A second feature extraction network is used to extract hyperbolic features from the seismic trace data; A fusion network is used to fuse the energy cluster features with the hyperbolic features; A sequence prediction network is used to predict velocity sequences based on feature maps obtained by fusion networks, and to generate velocity curves for the seismic traces.

4. The seismic data velocity modeling method according to claim 2, characterized in that, Using the velocity curve generated by the current velocity prediction model as a priori constraint, candidate regions containing energy clusters are segmented from the velocity spectrum of the seismic trace, including: The initial evolution curve of the Chan-vese segmentation model is constructed using the velocity curve generated by the current velocity prediction model; The velocity spectrum of the seismic trace was segmented using the Chan-vese segmentation model to identify candidate regions where energy clusters are located.

5. The seismic data velocity modeling method according to claim 4, characterized in that, The construction of the initial evolution curve of the Chan-vese segmentation model using the velocity curve generated by the current velocity prediction model includes: The velocity curve generated by the current velocity prediction model is shifted along the negative direction of the velocity coordinate axis by a first preset distance, and the shifted velocity curve is used as the first initial evolution curve. The velocity curve generated by the current velocity prediction model is translated along the positive direction of the velocity coordinate axis by a second preset distance, and the translated velocity curve is used as the second initial evolution curve. The initial evolution curves of the Chan-vese segmentation model are composed of the first initial evolution curve and the second initial evolution curve. The rules for determining the first preset distance and the second preset distance include at least ensuring that the initial evolution curve includes all energy clusters.

6. The seismic data velocity modeling method according to any one of claims 2 to 5, characterized in that, The location of the maximum energy cluster in the candidate region is tracked using an unsupervised clustering method to obtain the corrected velocity curve of the seismic trace, including: The mean-shift clustering method is used to track the location of the maximum energy of the energy cluster within the candidate region; Based on the energy points at the location of maximum energy in the energy cluster, the velocity curve of the seismic trace is obtained by interpolation, and this velocity curve is used as the corrected velocity curve.

7. The seismic data velocity modeling method according to any one of claims 2 to 5, characterized in that, Before using the speed prediction model that yields a predicted speed curve that meets preset requirements as the final speed prediction model, the following steps are also included: The speed prediction model is adjusted using the artificial labels to ensure that the predicted speed curve meets the preset requirements.

8. The seismic data velocity modeling method according to claim 1, characterized in that, Also includes: Median filtering and two-dimensional Sobel filtering are applied to the velocity spectrum in the multimodal data to trigger the execution of using the multimodal data as input to a pre-trained velocity prediction model, and using the velocity prediction model to generate the velocity curve of the seismic trace.

9. The seismic data velocity modeling method according to claim 3, characterized in that, When training the current velocity prediction model using the corrected velocity curves and artificial labels of each seismic trace, the energy cluster features are superimposed on the historical seismic trace set for imaging, and the seismic trace data of a specific seismic trace in the imaging profile obtained by superimposing the imaging is extracted. Then, the extracted seismic trace data of the specific seismic trace is fused with the feature map obtained by the fusion network and input into the sequence prediction network. The specific seismic trace is the seismic trace where the current seismic trace data input into the second feature extraction network during training is located.

10. The seismic data velocity modeling method according to claim 3, characterized in that, The first feature extraction network includes a first CNN unit, a second CNN unit, a first CNN-Transformer unit, and a second CNN-Transformer unit that are connected sequentially and whose network depths increase sequentially.

11. The seismic data velocity modeling method according to claim 3, characterized in that, The second feature extraction network includes a third CNN unit, a fourth CNN unit, a fifth CNN unit, a sixth CNN unit, a seventh CNN unit, an eighth CNN unit, and a pooling unit, which are connected sequentially and whose network depth increases sequentially.

12. The seismic data velocity modeling method according to claim 3, characterized in that, The fusion network includes a ninth CNN unit and a third CNN-Transformer unit that are connected sequentially and whose network depth increases sequentially.

13. The seismic data velocity modeling method according to claim 1, characterized in that, The time-domain velocity volume is a time-domain superimposed velocity volume, a time-domain offset velocity volume, or a time-domain remaining velocity volume.

14. The seismic data velocity modeling method according to claim 1, characterized in that, The seismic gathers are CMP gathers, CRP gathers, or residual gathers.

15. A seismic data velocity modeling device, characterized in that, include: The data acquisition module is used to acquire or construct multimodal data of any seismic trace in the target work area. The multimodal data includes the velocity spectrum of the seismic trace and seismic trace data of the seismic trace extracted from the seismic trace set of the target work area. The velocity spectrum of the seismic trace is calculated based on the seismic trace set. The velocity curve prediction module is used to take the multimodal data as input to a pre-trained velocity prediction model and use the velocity prediction model to generate the velocity curves of the seismic traces, so that the velocity curves of all seismic traces in the target work area can constitute the time domain velocity volume of the target work area. The velocity prediction model is obtained by training an initial machine learning model. The velocity prediction model is used to extract features from each modal data in the multimodal data, fuse all the extracted features, and then predict the velocity curve based on the fused features. The features extracted from each modal data include energy cluster features extracted from the velocity spectrum and hyperbolic features extracted from the seismic trace data.

16. A computer device, characterized in that, include: The memory is configured to store instructions; as well as The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the seismic data velocity modeling method according to any one of claims 1 to 14.

17. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to perform the seismic data velocity modeling method according to any one of claims 1 to 14.

Images