Model, method, and model training method for high-resolution processing of seismic data

Abstract

This disclosure proposes a model, a method, and a model training method for high-resolution processing of seismic data. The model includes: a scale factor determination unit configured to determine, based on an analysis of frequency characteristics of a dominant frequency band and a high-frequency band of seismic data, a stretching scale factor corresponding to the high-frequency band; a dominant frequency processing unit configured to perform high-resolution processing on the seismic data at an original scale to obtain initial data; a high-frequency processing unit configured to perform stretching processing on the seismic data in the time domain using the stretching scale factor, and to perform high-resolution processing on the seismic data after the stretching processing to obtain second data;and a fusion unit configured to fuse the first and second data in the time domain to obtain super-resolution broadband seismic data. By replanning the learning direction of the high-resolution seismic data processing model based on the scale factor, the model matches high-frequency information at an early training iteration stage, thus enabling higher-resolution seismic data to be obtained in fewer iterations. Figure for abstract: Fig. 7;

Description

Title of the invention: Model, method, and model training method for high-resolution processing of seismic data. Technical field

[0001] This disclosure relates to the technical field of geophysical exploration and development, and in particular to a model, a method and a model training method for high-resolution processing of seismic data.

[0002] Background

[0003] As oil and gas exploration and development gradually penetrates deep strata, deep waters, unconventional and other complex terrains, subtle reservoirs with thinner walls, deeper burials, and more complex surface conditions and structural characteristics have progressively become the central focus and challenge for increasing oil and gas reserves and production. To meet the requirements of fine exploration of complex oil and gas reservoirs, it is necessary to improve the resolution of seismic data, i.e., to enhance the definition of seismic images, so as to provide higher-quality seismic data for subsequent structural interpretation, parameter inversion, reservoir prediction, drilling engineering, etc.

[0004] Due to factors such as stratum absorption, frequency loss in seismic data, and wavelet interference, the actually collected narrowband seismic data are often blurry and have low resolution. Generally, a reflectivity has a wider frequency band than a seismic recording and is capable of representing the geometric structure features of subsurface strata with high resolution. High-resolution processing of the seismic recording (i.e., seismic data) can be understood as the calculation of a reflectivity when the actual seismic recording is known, and this processing is also called deconvolution. Deconvolution is a seismic wavelet compression process that compresses a long-duration seismic wavelet (broad pulse) into a narrow pulse.The resolution of the seismic recording is determined by the frequency band of the seismic wavelet. In the time domain, the process from seismic recording to reflectivity can be considered an inverse filtering process. In the frequency domain, the wavelet compression process results in the expansion of the amplitude spectrum and the phase shift of the phase spectrum.

[0005] The very first high-resolution processing method is impulse deconvolution The most classic and sophisticated method is the deconvolution method. This method assumes that underground reflectivity is white noise and that the seismic wavelet has a minimum phase. Reflectivity is estimated using a method such as least squares, thus allowing the effective increase in the seismic frequency band and the phase shift of the wavelets to be obtained. However, this method cannot be extended to the low signal-to-noise ratio zone or to information outside the seismic frequency band. Therefore, a deconvolution method based on compressed detection constraints and sparse representation is developed.This method primarily assumes that signals both within and outside the seismic frequency band satisfy the rule of superposition of a limited number of simple harmonic waves, and reduces the multiplicity of an inversion reflectivity by adding sparse prior information or regularization constraints to an objective function. Currently, regularization constraints commonly used to characterize the sparse nature of reflectivity include: Ll norm minimization, Cauchy criterion, Huber criterion, Lp norm minimization (0 < p < 1), Sech constraint, etc.

[0006] Research on high-resolution seismic data processing based on artificial intelligence dates back to the 1990s. Recently, a new cycle of high-resolution intelligent processing based on deep learning technology has been widely addressed and has developed rapidly. Depending on the uses of the samples and labels, high-resolution processing methods based on artificial intelligence can be divided into three categories: supervised learning, unsupervised learning, and semi-supervised learning.

[0007] The high-resolution method based on supervised learning learns an implicit correlation between seismic data and actual reflectivity under the constraint of well log labels, and a data-guided inverse operator is obtained, performing direct mapping from narrowband seismic data with high-resolution reflectivity. This method is primarily configured with two network architectures, namely a data-guided network and a physically guided network. On the one hand, the data-guided network can use the actual reflectivity derived from the well log data as a label to participate directly in the network training process, and build a neural network based on a wellbore.The trained model can describe a nonlinear relationship between the original seismic data and the actual reflectivity, and can obtain high-resolution seismic data consistent with the actual well distribution after being applied to the entire seismic dataset. The well-pipe-based neural network was used in the high-resolution data processing. Seismic processing has long been used, for example, in high-resolution processing based on Hopfield neural networks. After a multilayer perceptron was implemented, a deep neural network with deeper layers and more complex nonlinear units was progressively applied to high-resolution processing, demonstrating the advantages of excellent noise reduction performance and flexible, efficient computation. Furthermore, the data-driven network can generate a synthetic dataset from the actual seismic wavelet and a model reflectivity, and build a neural network based on wavelet conduction. When a synthetic dataset needs to be built, several reflectivity sequences are first randomly generated and then convolved with the actual seismic wavelets to synthesize the seismic data.The wavelet-driven neural network learns mapping relationships from synthetic seismic data to random reflectivity sequences, thereby mitigating or even eliminating the effects of seismic wavelet band limitation and sidelobe interference, and performing a deconvolution-like function. Finally, high-resolution seismic data processing results can be obtained by applying the wavelet-driven neural network to ground data. This concept is further applied to high-dimensional data such as 2D and 3D data. By analogy with model-driven processing, the physically guided neural network provides a network architecture that can be interpreted for intelligent, high-resolution processing.For example, by analogy with classical optimization algorithms such as the semiquadratic splitting method (HQS) and the alternating directions of multipliers method (ADMM), hidden layers are used to simulate the optimization algorithm corresponding to the iteration process, and reflectivity is iteratively updated and optimized by stacking the hidden layers.

[0008] The high-resolution processing method based on unsupervised learning is primarily intended to improve upon, by analogy with, the artificial intelligence process based on classical optimization algorithms, such as the introduction of Bayesian learning or backpropagation. This unsupervised learning method can perform reflectivity inversion without log data and obtain the result at high resolution. However, this method still requires certain prior information constraints.

[0009] The high-resolution method based on semi-supervised learning takes into account the constraint effect of seismic data on the high-resolution result, and introduces a non-recurrent network to construct the mapping from the estimated reflectivity to the actual seismic data, so as to balance the influences of the seismic data and the logging data on the high-resolution result, and to avoid overfitting caused by well constraints, thereby improving the stability of the neural network model in high-resolution processing.

[0010] High-resolution seismic data processing based on artificial intelligence eliminates, to some extent, the original assumption of the sparse nature of reflectivity and is no longer completely limited by the geological model framework, thus allowing adaptation to more complex geological conditions. However, it is difficult for existing artificial intelligence processes to achieve good high-resolution processing effects on field data. The main reasons are as follows: 1.The classic intelligent high-resolution process lacks guidance in terms of geophysical knowledge, and only builds the direct end-to-end mapping of seismic data to high-resolution data; however, the actually measured logging data is extremely sparse, and training the neural network with limited data will easily fall into a local optimum and cause overfitting of the training ensemble data; consequently, the learned model exhibits very insufficient generalization when applied to test data, and the prediction results demonstrate poor spatial continuity, the resulting model therefore has virtually no application value for a larger amount of field data, which greatly limits the scope of application; 2.Existing artificial intelligence processes exhibit a "bias" for learning seismic frequency, and high-frequency information is often not learned well, leading to "imprecision" or "uncertainty" in extended-frequency information.

[0011] Summary

[0012] One objective of this disclosure is to propose a model, a method and a model training method for high-resolution seismic data processing, so as to solve the problems of insufficient generalization and low accuracy of existing high-resolution seismic data processing methods based on artificial intelligence models.

[0013] To address the above technical issues, a first aspect of this disclosure proposes a high-resolution seismic data processing model, comprising: a scaling factor determination unit configured to receive seismic data from a target terrain, determine a dominant frequency band and a high-frequency band of the seismic data, and determine a stretching scaling factor corresponding to the high-frequency band. The stretching scaling factor is applied to the high-frequency data in a time domain to compress the high-frequency data into a frequency range dominant; a dominant frequency processing unit configured to perform high-resolution processing on the seismic data at an original scale to obtain first high-resolution seismic data; a high-frequency processing unit configured to perform stretching processing on the seismic data in the time domain using the stretching scale factor, and perform high-resolution processing on the stretched seismic data to obtain second high-resolution seismic data; and a fusion unit configured to fuse the first high-resolution seismic data and the second high-resolution seismic data in the time domain to obtain super-resolution broadband seismic data.

[0014] In certain embodiments, the scaling factor determination unit is further configured to determine a low-frequency band and a compression scaling factor corresponding to the low-frequency band; correspondingly, the high-resolution seismic data processing model further comprises a low-frequency processing unit configured to perform compression on the seismic data in the time domain with the compression scaling factor, and to perform high-resolution processing on the stretched seismic data to obtain third high-resolution seismic data; and the fusion unit merges the first high-resolution seismic data, the second high-resolution seismic data and the third high-resolution seismic data in the time domain to obtain the super-resolution broadband seismic data.

[0015] A second aspect of this disclosure proposes a method for training a high-resolution processing model for training the high-resolution seismic data processing model according to any element of the first aspect, the method comprising: obtaining seismic data samples from a target location in a target terrain; obtaining log data from an actual well at the target location in the target terrain, and determining high-resolution seismic data samples based on the log data; and performing iterative training of the high-resolution seismic data processing model by taking the seismic data samples as input to the high-resolution seismic data processing model and the high-resolution seismic data samples as the target output of the high-resolution seismic data processing model.

[0016] In certain embodiments, before performing iterative training of the high-resolution seismic data processing model by taking the seismic data samples as input to the high-resolution seismic data processing model and the high-resolution seismic data samples solution as the output target of the high-resolution seismic data processing model, the method further includes a pre-training method for the high-resolution seismic data processing model comprising: performing statistical wavelet extraction on seismic data samples to obtain a wavelet sample database of seismic data samples; convolving a geological model data reflectivity set with seismic wavelets in the wavelet sample database to obtain a synthetic seismic data sample database;and the pre-training of the high-resolution seismic data processing model by taking the synthetic seismic data sample database as input to the high-resolution seismic data processing model and the geological model data reflectivity set as the output target of the high-resolution seismic data processing model.

[0017] In certain embodiments, during the iterative training of the high-resolution seismic data processing model, an iteration direction is determined on the basis of an objective function which includes a difference term and a constraint term, and the difference term includes a difference between a target output value and an actual output value of the high-resolution seismic data processing model, and the constraint term includes a sum of a difference corresponding to each of the processing units; each of the processing units includes a high-frequency processing unit and / or a low-frequency processing unit, and the difference corresponding to a processing unit is a difference between the target output value and the actual output value of the processing unit.

[0018] A third aspect of this disclosure proposes a method for high-resolution seismic data processing, comprising: inputting seismic data of a target terrain into the high-resolution seismic data processing model according to any element of the first aspect, and taking output data from the high-resolution seismic data processing model as high-resolution seismic data.

[0019] A fourth aspect of this disclosure proposes a processing model training apparatus for training the high-resolution seismic data processing model according to any element of the first aspect; the apparatus comprising: a first acquisition unit configured to obtain seismic data samples from a target location in a target terrain; a second acquisition unit configured to obtain logging data from an actual well at the target location in the target terrain, and to determine high-resolution seismic data samples based on the logging data; and a unit iterative training configured to perform iterative training of the high-resolution seismic data processing model by taking seismic data samples as input to the processing model and high-resolution seismic data samples as the output target of the high-resolution seismic data processing model.

[0020] In certain embodiments, the apparatus further comprises: an extraction unit configured to perform statistical wavelet extraction on seismic data samples to obtain a wavelet sample database of seismic data samples; a synthesis unit configured to convolve a geological model data reflectivity set with seismic wavelets in the wavelet sample database to obtain a synthetic seismic data sample database;and a pre-training unit configured to pre-train the high-resolution seismic data processing model by taking the synthetic seismic data sample database as input to the high-resolution seismic data processing model and the geological model data reflectivity set as the output target to the high-resolution seismic data processing model.

[0021] In certain embodiments, during the iterative training of the processing model, an iteration direction is determined according to an objective function which includes a difference term and a constraint term, and the difference term includes a difference between a target output value and an actual output value of the high-resolution seismic data processing model, and the constraint term includes a sum of a difference corresponding to each of the processing units; each of the processing units includes a high-frequency processing unit and / or a low-frequency processing unit, and the difference corresponding to a processing unit is a difference between the target output value and the actual output value of the processing unit.

[0022] A fifth aspect of this disclosure proposes a high-resolution seismic data processing device, comprising an application unit configured to input seismic data from a target terrain into the high-resolution seismic data processing model according to any element of the first aspect, and to take output data from the high-resolution seismic data processing model as high-resolution seismic data.

[0023] A sixth aspect of this disclosure proposes an electronic device comprising: a memory and a processor in communication connection with each other, and the memory stores computer instructions, and the processor implements the steps of the process according to any element of the second or third aspect aspect by executing computer instructions.

[0024] A seventh aspect of this disclosure proposes a computer storage medium, and the computer storage medium stores computer program instructions; when executed, the computer program instructions implement the steps of the process according to any element of the second or third aspect.

[0025] In the model, method, and model training method for high-resolution seismic data processing proposed by the invention, a principle analysis and an exploration of the essence of high-resolution processing by the intelligent model are performed. A 'frequency learning preference' of the intelligent model is summarized, and the frequency change of the intelligent model in the iteration process is clarified, as a guide to explaining the difficulty of generalizing classical intelligent models. A processing model incorporating a high-resolution intelligent nonlinear operator based on multi-scale frequency characteristics is proposed.A scaling factor is designed based on internal correlations between the low frequency, dominant frequency and high frequency bands, and a seismic signal is proportionally stretched in the time domain, so that the high frequency component in the frequency domain approaches the dominant frequency component, thus reducing the difference between different multi-scale frequency characteristics.In addition, multi-frequency multi-scale learning and cooperative fusion are adopted, while being integrated into the process of solving a high-resolution nonlinear processing operator to build constraints of multi-scale frequency features, and by replanning a learning direction of the high-resolution seismic data processing model, the constructed high-resolution processing model rapidly descends in a solution space to an area of ​​strong generalization and high stability.When the mapping of low-resolution data to high-resolution data is learned, based on the matching of dominant frequency information, the high-frequency information is matched at an early stage of the network iteration. This improves the accuracy and certainty of the high-frequency component in the prediction result, thus enabling the acquisition of seismic images with higher resolution. High-resolution and high-definition results, cooperatively extended by multi-scale multi-frequency, can be obtained by applying the solution to non-visible seismic data. This not only avoids the transverse discontinuity or "horizontal discontinuity" phenomenon caused by unstable high-frequency extension after single-trace processing, but also achieves a high correlation with the actual log measurement data. The thin-layer response, which is finely represented on the well, can be improved transversely as stratum features in the profile. Brief description of the drawings

[0026] In order to more clearly describe the technical solutions in the embodiments of this disclosure or of the prior art, the drawings to be used in the descriptions of the embodiments or of the prior art will be briefly presented as follows. Of course, the drawings in the following descriptions only illustrate certain embodiments of this disclosure, and a person skilled in the art can obtain other drawings from them without resorting to any creative work.

[0027] Fig. 1 illustrates a diagram comparing amplitude spectrum curves of the results of high-resolution processing of an existing intelligent model with the change in the number of iterations;

[0028] Figure [Fig. 2] illustrates a relationship curve for high-resolution processing results and the number of iterations when an existing intelligent model converges;

[0029] Figure 3 illustrates a high-resolution results resolution diagram of an existing intelligent model after different numbers of iterations;

[0030] Figure 4 illustrates a schematic diagram of a structure of a high-resolution seismic data processing model according to the present disclosure;

[0031] Fig. 5 illustrates a schematic diagram of results of multi-scale frequency analysis of high-resolution single trace data in the time domain;

[0032] Fig. 6 illustrates a schematic diagram of multi-scale frequency analysis results of high-resolution single trace data in the frequency domain;

[0033] Figure 7 illustrates a schematic diagram of another structure of a high-resolution seismic data processing model according to the present disclosure;

[0034] Fig. 8 illustrates a schematic diagram of an internal structure of a high-frequency processing unit or a low-frequency processing unit;

[0035] Figure 9 illustrates a schematic diagram of a method for training a high-resolution seismic data processing model according to this disclosure;

[0036] Fig. 10 illustrates a schematic diagram of high-resolution field data processing results;

[0037] Fig. 11 illustrates a schematic diagram of comparison between amplitude spectra of real log data participating in training in the frequency domain;

[0038] Fig. 12 illustrates a schematic diagram of comparison between amplitude spectra of profile data in the frequency domain;

[0039] Figure 13 illustrates a schematic diagram of high-res treatment results. field data solution;

[0040] Figure 14 illustrates a schematic diagram comparing synthetic seismic data obtained by convolution of high-resolution data with seismic wavelets; and

[0041] Figure 15 illustrates a schematic diagram of a structure of an electronic device according to the present disclosure. Detailed description

[0042] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and fully described with reference to the drawings for the embodiments of this disclosure. Of course, the described embodiments are only parts, and not all, of the embodiments of this disclosure. Any other embodiment obtained by those skilled in the art based on the embodiments of this disclosure, without any creative work, falls within the scope of protection of this disclosure.

[0043] During high-resolution processing using an intelligent model, it is observed that there is a certain rule during iterative learning, which is mainly manifested by the priority matching of dominant frequency information, followed by extensions to low and high frequency parts.

[0044] Fig. 1 illustrates a diagram comparing amplitude spectrum curves of high-resolution processing results of an existing intelligent model with a change in the number of iterations, on which a solid black line with a narrow frequency band next to the letter A represents an amplitude spectrum of seismic data, a solid black line with a wide frequency band next to the letter B represents an amplitude spectrum of high-resolution data, a dashed gray line represents an amplitude spectrum of a high-resolution result after 10 iterations, a dashed gray line represents an amplitude spectrum of a high-resolution result after 30 iterations, a dashed black line represents an amplitude spectrum of a high-resolution result after 50 iterations, and a light gray plus sign line represents an amplitude spectrum of a high-resolution result after 100 iterations.As shown in [Fig.1], the high-resolution result after 10 iterations is similar to the seismic data, and is only extended in the vicinity of a seismic frequency band; when 30 iterations are reached, the amplitude spectrum is clearly extended, but the amplitude is not increased efficiently; after 50 iterations, only the high-frequency part is deflected, and after 100 iterations, the loss function finally converges and both the high part. The frequency and low-frequency parts are well adjusted.

[0045] When the iteration process is refined into an early stage, an intermediate stage and a late stage, it can be found that the frequency rule also corresponds to an obvious change rule in the time profile data.

[0046] In the early stage of iteration, the predicted high-resolution data closely match the seismic data, and only the amplitudes of certain events present in the seismic data are improved, primarily due to the preference for the dominant frequency. At this stage, the intelligent model's learning occurs essentially only within the dominant frequency range.

[0047] At the intermediate stage of the iteration, as the iteration continues, a new fine-layer event appears in the predicted high-resolution data. These events are originally hidden in thick layers of the seismic data, and original fine-layer information in the thick layers is obscured by wavelet interference. At this point, the model can learn the wavelet compression capability similar to that in model-driving methods. After compression, the wavelet interference is further reduced, some fine layers appear, and the amplitude spectrum has some extension at high frequencies. However, the amplitude is not increased, and a significant absence of high-frequency information remains.

[0048] In the late stage of T iteration, when the objective function converges almost, a major-event decomposition appears in the prediction result. That is, a major event that has a significant influence on the dominant frequency in the original seismic data is decomposed into two fine events with some separation under T high-frequency information intervention. However, due to the incomplete inversion and the large multiplicity of high-frequency information, the amplitude spectrum exhibits strong oscillations and perturbations in the high-frequency region. The fine event obtained from the major-event decomposition exhibits the time-shift dislocation gap, and the profile with transverse continuity reduction characteristics is represented.

[0049] When the iteration process finally converges, in the ideal state of the model data, the intelligent model fully learns the low-frequency and high-frequency information in the log data, and eliminates the disturbances in the high-frequency part, so that the major event is decomposed into two new events with good continuity and obvious boundary characteristics.

[0050] Figure 2 illustrates a high-resolution treatment outcome relationship curve and the number of iterations required for an existing intelligent model to converge, and [Fig. 3] illustrates a high-resolution output diagram of an existing intelligent model after different numbers of iterations. The intelligent model's learning rule can be viewed more intuitively by combining [Fig. 2] and [Fig. 3].

[0051] In [Fig.2], the positions of the 10th, 30th, 50th and 100th iterations are marked as principal analysis objects. As shown in [Fig.2], after 10 iterations, the decline of the convergence curve clearly slows down, then falls again; in 30 iterations, the decline of the convergence curve falls rapidly; after 50 iterations, the convergence curve almost converges, and after 100 iterations, the model converges almost completely and the convergence curve oscillates only slightly.

[0052] Fig. 3 (a) illustrates a high-resolution result after 10 iterations; at this point, the events are thick, the resolution is comparable to seismic data, and there are still thick events, each indicated by black arrows, in two areas. Fig. 3 (b) illustrates a high-resolution result after 30 iterations; at this point, a wavelet is compressed, and the interference is weakened. A thin event appears in the area indicated by the upper black arrows, but a major event still exists in the area indicated by the lower black arrows. Fig. 3 (c) illustrates a high-resolution result after 50 iterations, in which a major event in an area indicated by the lower arrows is decomposed, but there is noise perturbation and horizon dislocation.3] (d) illustrates a high-resolution result after 100 iterations, in which a major event in an area indicated by the lower arrows is effectively decomposed and precisely matched, and good transverse continuity is represented in the profile.

[0053] In summary, the frequency learning rule of the intelligent model can be understood as a process in which a strong constraint is imposed by the dominant frequency at the beginning, the influence of the dominant frequency is progressively weakened after the dominant frequency exhibits a high match, and then a high-frequency constraint and a low-frequency constraint are introduced. In other words, in the high-resolution intelligent model learning process, a dominant frequency preference emerges, meaning that information of the dominant frequency exhibits a preferential match, resulting in insufficient learning of high-frequency information and insufficient spectrum extension; that is, there is a problem of "frequency bias".

[0054] Moreover, due to the imprecision and instability of the high-frequency extension, the transverse continuity of the high-resolution treatment results is poor.

[0055] Based on the frequency-domain learning rule of the intelligent model, with an increasing number of iterations, the high-resolution prediction effect improves more and more in the model data. Theoretically, regardless of the complexity of the high-resolution information, it is possible to converge and eventually learn the corresponding high-frequency features, provided there is an infinite number of iterations.

[0056] However, in practical data applications, due to limitations such as the quantity of well log data and the quality of seismic well mapping, the assumption of the theoretical geological model is not satisfied by the seismic and high-resolution data, and it is impossible to construct a highly generalized mapping relationship. High-resolution processing by artificial intelligence encounters the problem of small samples in real-world situations.When the classical intelligent model is used to directly process high-resolution data, based on 'frequency bias', although a very good fitting effect in the training set is obtained progressively with the progression of the iteration process, the model inevitably falls into a state of overfitting due to the limited amount of real log data, such that the model proves useless for application to seismic datasets.

[0057] However, by simply reducing the number of iterations in the learning process, the model will only reach the dominant frequency matching state, and it is impossible to efficiently learn the high-frequency characteristic. The final prediction result of the model is substantially similar to the dominant frequency characteristic, and only a few traces of high-frequency perturbation can be found in the profile.

[0058] Thus, improving the learning direction of the intelligent model has become an urgent problem to be solved by intelligent high-resolution processing in real-world data applications. This means that high-frequency information must be rapidly introduced into the model iteration process and impose effective constraints.

[0059] The inventor analyzes the multi-scale frequency characteristics of high-resolution data and implements a scaling factor with a frequency-domain stretching and compression function. After the scaling factor is applied to the high-resolution data, the high-frequency information is transformed into a frequency-dominant range in the frequency domain to obtain scaling characteristics similar to the dominant frequency. A high-resolution nonlinear operator constrained by at least two types of frequency-scale information in the update process is constructed by merging the factor scaling in the intelligent model building process. The scaling factor stretches the seismic signal longitudinally in the time domain, causing the frequency domain to be compressed into a low frequency range based on the scaling factor, so as to compress the high frequency information into the dominant frequency range in the frequency domain, while protecting the time domain from information loss.

[0060] As illustrated in [Fig.4], the present disclosure proposes a high-resolution seismic data processing model, which includes a scale factor determination unit, a dominant frequency processing unit, a high-frequency processing unit, and a fusion unit.

[0061] The scaling factor determination unit is configured to receive seismic data from a target terrain, determine a dominant frequency band and a high frequency band of the seismic data, and determine a stretching scaling factor corresponding to the high frequency band, and the stretching scaling factor is applied to the high frequency data in the time domain to compress a frequency band of high frequency data into a dominant frequency band range.

[0062] It should be noted that there may be one or more high-frequency bands defined here, and each of the high-frequency bands corresponds to a stretching scale factor and a high-frequency processing unit. There may be one or more low-frequency bands defined here, and each of the low-frequency bands corresponds to a compression scale factor and a low-frequency processing unit.

[0063] The dominant frequency processing unit is configured to perform high-resolution processing on seismic data at an original scale to obtain first high-resolution seismic data. When the seismic data in the dominant frequency band is not compressed or stretched, the scale factor corresponding to the dominant frequency can be considered equal to 1.

[0064] The scale is represented by the number of sampling points (or data volume) in the seismic data. The original scale means that the number of sampling points (or data volume) in the seismic data remains unchanged. After stretching the seismic data, the number of sampling points (or data volume) in the seismic data increases, i.e., the scale of the seismic data increases. After compressing the seismic data, the number of sampling points (or data volume) in the seismic data decreases, i.e., the scale of the seismic data decreases.

[0065] High-resolution processing of seismic data is a classic operation in the field of seismic data processing, and will not be described here.

[0066] The high-frequency processing unit is configured to perform stretching processing on seismic data in the time domain using the stretching scale factor to perform high-resolution processing on the seismic data to obtain high-resolution seismic data.

[0067] The fusion unit is configured to fuse the first high-resolution seismic data and the second high-resolution seismic data in the time domain to obtain super-resolution broadband seismic data.

[0068] The determination of the dominant frequency band, the high-frequency band, and the low-frequency band of seismic data and the corresponding scale factor is also called multi-frequency scale analysis. Figure 5 illustrates a schematic diagram of multi-frequency scale analysis results of high-resolution single-trace data in the time domain, in which one curve in Figure 5 (a) represents the high-resolution single-trace data itself; three curves in Figure 5 (b) represent results after frequency division, a solid line representing a low-frequency portion from 0 to 10 Hz, a dashed line representing a dominant frequency portion from 10 to 50 Hz, and a dashed line representing a high-frequency portion above 50 Hz; three curves in Figure 5 (b) represent the high-frequency portion from 0 to 10 Hz.5] (c) are results from different frequency bands under the action of a scaling factor; at this point, the low-frequency portion is compressed in the time domain and the high-frequency portion is stretched in the time domain. [Fig. 6] illustrates a schematic diagram of multiscale frequency analysis results from high-resolution single-trace data in the frequency domain, in which [Fig. 6] (a) illustrates a full-frequency amplitude spectrum of high-resolution single-trace data; [Fig. 6] (b) illustrates amplitude spectrum information from different frequency bands after frequency division, and the curve designations are the same as those in [Fig. 5] (b); and [Fig. 6] (c) illustrates amplitude spectrum results under the action of a scaling factor.We can see that both the low and high frequency parts are transformed into the dominant frequency range, and have frequency characteristics similar to those of the dominant frequency.

[0069] The scaling factor determination unit can determine the stretching scaling factor by taking the lowest frequency of the dominant frequency band and the highest frequency of the high-frequency band as the first frequency range after the high-frequency band stretching process in the time domain, and by taking a ratio of the first frequency range to an original range of the high-frequency band before the stretching process as stretch scaling factor.

[0070] The lowest frequency in the dominant frequency band can refer to the lowest frequency with an amplitude meeting a predetermined condition in the dominant frequency band, and the highest frequency in the high-frequency band can refer to the highest frequency with an amplitude meeting a predetermined condition in the high-frequency band. The predetermined condition can mean that it is above a predetermined threshold.

[0071] The method for the high-frequency processing unit to perform a stretching process on full-frequency seismic data in the time domain can consist of: multiplying the number of sampling points of the full-frequency seismic data in the time domain by the stretching scale factor to obtain the first number of sampling points after the stretching process, and performing a linear interpolation on the full-frequency seismic data in the time domain taking the first number of sampling points as the sampling object; at that time, the amplitude spectrum of the seismic data in the corresponding frequency domain is compressed in the dominant frequency range, and the high-frequency part is compressed in the dominant frequency range.

[0072] The dominant frequency processing unit, the high frequency processing unit and the low frequency processing unit can also be called multi-scale nonlinear high-resolution operators.

[0073] As illustrated in [Fig. 5], the difference in information between different frequency bands can be perceived more intuitively after frequency division of the high-resolution data. The low frequency changes slowly with the evolution of a large set of strata, the dominant frequency characteristics are similar to those of seismic data, and the high frequency changes rapidly with a large amount of information on the thin layers. The results of frequency division are transformed linearly on the basis of the scale factor. It can be seen intuitively from the time-domain data that the information in each frequency band has a similar rate of change, and it can also be seen in the frequency-domain amplitude spectrum curve that all the different frequency bands after frequency division are concentrated in the dominant frequency range.After multi-scale frequency analysis, this solution proportionally stretches the high-frequency components outside the dominant seismic frequency, so that the high frequency exhibits frequency characteristics close to those of the dominant frequency, and after acting on the seismic data in a sample database during training, the high-resolution nonlinear multi-scale operator can simultaneously perform cor. The dominant and high-frequency information from the high-resolution label is retrieved. This means that the dominant and high-frequency information is learned in the early stages of the iterative training of the processing model. Due to the function of high-frequency information, high-resolution seismic data can be obtained in the early stages of the iterative training of the processing model, so high-resolution seismic data can still be obtained by reducing the number of iterations. For example, high-resolution seismic data is obtained after 500 iterative trainings in the prior art, as opposed to only 200 iterative trainings in the solution proposed here.

[0074] The high-frequency processing unit normalizes high-frequency information with respect to the dominant frequency band in the frequency domain, so that the processing model exhibits a similar sensitivity to high-frequency information as to dominant frequency information and has the ability to preferentially learn the target frequency band. 'Normalization of high-frequency information to the dominant frequency band in the frequency domain' can be understood as the frequency domain normalization process, which is primarily implemented by a mathematical transformation between the time domain and the frequency domain.

[0075] In some embodiments, the fusion unit merges the first high-resolution seismic data and the second high-resolution seismic data in the time domain, meaning that the first high-resolution seismic data and the second high-resolution seismic data are processed at a target number of sampling points, and then the processed first high-resolution seismic data and the processed second high-resolution seismic data are added together. In other words, the fusion operation can be a summation operation.

[0076] In other embodiments, the fusion unit merges the first high-resolution seismic data and the second high-resolution seismic data in the time domain, which also means that the first high-resolution seismic data and the second high-resolution seismic data are processed to the target number of sampling points, and that the first (for example, the original scale of the seismic data can be restored) processed seismic data and the second (for example, the original scale of the seismic data can be restored) processed seismic data are subjected to a convolution operation. In other words, the fusion operation can be a convolution operation.

[0077] The low-frequency processing unit, the dominant-frequency processing unit and The high-frequency processing units mentioned below can be considered as three channels for full-frequency seismic data, respectively. That is, the input data for the dominant-frequency processing unit, the high-frequency processing unit, and the low-frequency processing unit are full-frequency seismic data. After processing by the dominant-frequency processing unit, the high-frequency processing unit, and the low-frequency processing unit, the number of sampling points in the output data for the different channels varies. Therefore, summation or convolution cannot be performed directly; it is necessary to process the data from each channel to the same target number of sampling points initially.

[0078] The target number of sampling points can refer to the number of sampling points in the seismic data at the original scale. The second set of high-resolution seismic data can be compressed using the stretching scale factor to obtain the target number of sampling points. The third set of high-resolution seismic data mentioned below can be stretched using the compression scale factor to obtain the target number of sampling points. This method can be understood as a reverse frequency normalization process and can also be called frequency denormalization.

[0079] In certain embodiments, the scaling factor determination unit is further configured to determine the low-frequency band and the compression scaling factor corresponding to a low frequency. Thus, as illustrated in [Fig. 7], the high-resolution seismic data processing model further comprises a low-frequency processing unit configured to perform compression on the seismic data in the time domain with the compression scaling factor to obtain seismic data at a second scale, and to perform intelligent high-resolution processing on the seismic data at the second scale to obtain third high-resolution seismic data.In a case where the high-resolution seismic data processing model includes a low-frequency processing unit, the fusion unit merges the first high-resolution seismic data, the second high-resolution seismic data, and the third high-resolution seismic data in the time domain to obtain super-resolution broadband seismic data.

[0080] The processing of seismic data in the low frequency band can be understood by reference to the description of seismic data in the high frequency band and [Fig.5] and 6.

[0081] The scaling factor determination unit can determine the compression scaling factor by taking the highest frequency in the frequency band dominant and lowest frequency of the low frequency band as second frequency range after high frequency band compression processing in the time domain, and taking a ratio of an original low frequency band range before compression processing to the second frequency range as compression scaling factor.

[0082] The highest frequency in the dominant frequency band may refer to the highest frequency with an amplitude meeting a predetermined condition in the dominant frequency band, and the lowest frequency in the low-frequency band may refer to the lowest frequency with an amplitude meeting a predetermined condition in the low-frequency band. The predetermined condition may mean that it is above a predetermined threshold.

[0083] The method for the low-frequency processing unit to perform compression processing on high-frequency seismic data in the time domain can consist of: multiplying the number of sampling points of the high-frequency seismic data in the time domain by the compression scaling factor to obtain the second number of sampling points after the compression processing, and taking the second number of sampling points as the sampling target to perform linear sampling of the full-frequency seismic data in the time domain.

[0084] When the high-frequency processing unit and the low-frequency processing unit are used in the processing model, it is possible to directly establish an end-to-end mapping of seismic data to high-resolution data, and to obtain good prediction results.

[0085] Figure 8 illustrates a schematic diagram of a data processing procedure for the high-resolution seismic data processing model. In Figure 8, T in a lower frame represents a scale factor (comprising a stretch scale factor and a compression scale factor), and Multiplity T represents a seismic data scaling transformation using the scale factor; Conv 21x1 represents a data convolution operation with a 21x1 convolution kernel, and Conv 1x1 represents a data convolution operation with a 1x1 convolution kernel; Divide T represents a transformation from a data scale to an original scale using the scale factor (i.e., frequency domain denormalization); and ReLU represents data processing with a ReLU activation function.

[0086] As illustrated in [Fig. 8], the input data for the processing model are seismic data, and the output data are high-resolution seismic data. During training, a single-trace learning mode (seismic data) versus a single-trace learning mode (high-resolution data) is adopted. After that a trace of the seismic data is entered, the scale factor T is obtained from the multi-scale frequency analysis, and the seismic data is stretched or compressed with the scale factor T in the time domain to implement frequency domain normalization based on the dominant frequency band; then, through a submodule included by a one-dimensional convolution and a nonlinear ReLu function, each convolution layer is added with a grouping layer, so that the data retain the same frequency characteristics in the low-frequency, dominant-frequency, and high-frequency processing unit.The output of the processing units with different frequency bands is subjected to frequency domain denormalization based on the scale factor, so that the data is compressed to an original length in the time domain, the frequency characteristics identical to those of the seismic data are restored in the frequency domain, and the restored data are combined in the time domain in a multi-channel form. Finally, the data are merged by a 1x1 linear convolution kernel, in order to obtain the final predicted high-resolution data by taking into account globally the results of different processing units.

[0087] A multi-scale frequency characteristic analysis is performed, including frequency division and joint analysis of the frequency and time domains, so as to select the target frequency band to be extended and augmented. Corresponding scaling factors are designed for different target frequency bands based on a mathematical transformation relationship between the frequency and time domains, and the target frequency bands are compressed into the dominant frequency range of the seismic data in the frequency domain.Next, the corresponding processing units are constructed from different scaling factors, the time domain information is combined by connecting the output ends of the processing units in parallel via a linear convolution layer, so as to obtain a high-resolution processing model based on multi-scale frequency characteristics.

[0088] The high-resolution, nonlinear, multi-scale operator in the processing model is cooperatively constrained by frequency information from different scales, such as low frequency, dominant frequency, and high frequency. This not only improves the degree of utilization of low-frequency and high-frequency information but also utilizes the intrinsic correlation between different frequencies for mutual assistance, thereby enhancing the inversion effects of different frequency bands and particularly enriching the high-frequency information in the inversion results. Thus, the model of This processing method is particularly well-suited for addressing the inverse problem of high-resolution seismic data processing. The processing model exhibits good generalization and spatial continuity. It can perform high-resolution seismic data processing and generate detailed representations of geological structures for various geological targets, providing high-quality baseline data for structural interpretation, seismic inversion, reservoir prediction, drilling engineering, and other applications.

[0089] The processing model is an artificial intelligence processing model. This disclosure further proposes a training method for the processing model, which includes the following steps 10, 20, and 30.

[0090] S10: Obtain seismic data samples from a target location in a target area.

[0091] S20: Obtain log data from an actual well at the target location in the target terrain, and determine high-resolution seismic data samples based on logging data.

[0092] An impedance curve is calculated based on a primary wave velocity curve and a density curve in the log data, and a reflectivity curve is calculated based on the impedance curve. The reflectivity obtained at that time is an extremely sparse full-frequency reflectivity, and high-resolution data can be obtained by applying a high-pass filter to the full-frequency reflectivity.

[0093] S30: perform iterative training of the high-resolution processing model of seismic data by taking seismic data samples as input to the high-resolution seismic data processing model and high-resolution seismic data samples as the output target of the high-resolution seismic data processing model.

[0094] Before being entered into the processing model, seismic data samples and high-resolution seismic data samples can be subjected to normalization processing.

[0095] During the iterative training of the processing model in S30, an iteration direction is generally determined based on the objective function. In some embodiments, the objective function is the difference between the target output and the input data of the entire processing model.

[0096] In S10, S20, and S30, multi-scale frequency characteristics are used to guide the construction of the spatial structure of the intelligent model and the design of the objective function, and the geophysical rule in high-resolution processing is integrated into the design of the artificial intelligence process. Under such a learning mode, the processing model can have a better effect at high resolution and an application value in the fields and scenarios.

[0097] In other embodiments, the objective function may include a difference term and a constraint term, and the difference term includes a difference between the output target and the input data of the high-resolution seismic data processing model, and the constraint term includes a sum of a difference corresponding to each of the processing units; each of the processing units includes a high-frequency processing unit and / or a low-frequency processing unit, and the difference corresponding to the processing unit is a difference between the output data and the input data of the processing unit. treatment.

[0098] For example, the objective function is £ v , and the function The objective consists of a difference term and a constraint term, in which the difference term f represents a difference between a target output value (i.e., label data) and an actual output value of the high-resolution seismic data processing model, and the constraint term is jointly determined by the high-frequency processing units and the low-frequency processing units in a total number of n; the constraint term ç $ includes a sum of a difference corresponding to each of the Each processing unit comprises a high-frequency processing unit and / or a low-frequency processing unit, and the difference corresponding to a processing unit is the difference between a target output value and an actual output value of that processing unit. For example, represents a difference corresponding to the high-frequency processing unit or the low-frequency processing unit, and the coefficient is a regularization parameter, ïl S(0,1 The difference above can be an absolute value of a difference between numerical values ​​or a square of a difference between numerical values.

[0099] It should be noted that the input of the model or unit in this disclosure is the characteristic data in the sample data, and the output data of the model or unit is the label data in the sample data.

[0100] On the one hand, a downward direction of objective function values ​​of the training process of the processing model is constrained by the multiscale characteristics of different frequencies, and exhibits a smoother and more efficient downward trend. On the other hand, the training process of the processing model is supervised by the difference between the target output and the input data of the processing model, and is constrained by the multiscale frequency information of the high-frequency processing unit and / or the low-frequency processing unit. From the beginning, the iteration is constrained by the characteristics of different frequency bands jointly, which makes it possible to overcome the strong constraint of dominant frequency information to some extent, greatly reducing the possibility of overfitting, and further improving the application value of the trained processing model.

[0101] The high-resolution nonlinear multi-scale operator is cooperatively constrained by frequency information from different scales, which not only avoids the problem of limited use of high and low frequency information with a low signal-to-noise ratio, but also uses the intrinsic correlation between information from different frequencies for mutual assistance, which is particularly suitable for solving the inverse problem of high-resolution seismic data processing.

[0102] Due to the high cost of drilling, the amount of well log data is limited. When processing field data, transfer learning is generally adopted to reduce the influence of overfitting on the high-resolution result, but the effect is still limited. According to the present disclosure, a high-resolution nonlinear operator based on multiscale frequency characteristics is adopted to optimize a downward gradient direction under the multiscale frequency constraint, so that the high-resolution operator with transfer learning can obtain a high-resolution result with good spatial continuity in a limited number of iterations.

[0103] In certain embodiments, as illustrated in [Fig.9], prior to S30, the method further comprises a pre-training method for the high-resolution seismic data processing model comprising the following steps 40, 50 and 60.

[0104] S40: perform statistical wavelet extraction on the samples of seismic data to obtain a wavelet sample database of seismic data samples.

[0105] S50: convolve a set of geological model data reflectivity with seismic wavelets in the wavelet sample database to obtain a synthetic seismic data sample database.

[0106] Reflectivity is an inherent property of the stratum and will not change with a change in the detection mode. The reflectivity set of geological model data can be a set of reflectivities calculated by a geological model constructed in any way. In general, the reflectivity set of geological model data is known in the art and may be readily available for use.

[0107] S60: Pre-train the high-resolution seismic data processing model taking the synthetic seismic data sample database as input to the high-resolution seismic data processing model and the geological model data reflectivity set as the output target of the high-resolution seismic data processing model.

[0108] The processing model obtained by pre-training using the model training method described in S40, S50, and S60 exhibits the high-resolution capability of wavelet advance guidance. After pre-training, well logging data (small-sample data) from a real well is fed in for transfer learning and monitoring of the network iteration process. Since the real well logging data is accurate, the processing model obtained by transfer learning can achieve a good high-resolution effect.

[0109] In the training process, it is permissible to iterate the model many times until convergence. Finally, the nonlinear operator can learn the wavelet compression capability similar to that in model driving (a process where seismic data achieve high resolution is also called wavelet compression). The pre-trained model can be applied to field data to obtain high-resolution results with evident stratification, but at that point, the correlation of the high-resolution result with the well is limited, and some major events with a lot of dominant frequency information are still difficult to decompose.

[0110] Based on the pre-trained model, when the model parameters are further modified by transfer learning based on real well data, in order to prevent the model from overfitting, the number of iterations in S30 can be controlled below a predetermined number (for example, 200). At that point, under the constraint of multi-frequency scale information, the downward gradient direction is jointly constrained by information from three scales: low frequency, dominant frequency, and high frequency. This further improves the descent efficiency and balances the influences of seismic and well log data on the high-resolution result in the short-term iterations, obtaining the high-resolution result with a high signal-to-noise ratio and good spatial continuity.

[0111] This disclosure further proposes a method for high-resolution seismic data processing, which inputs seismic data from a target terrain into the high-resolution seismic data processing model, and takes the output data from the high-resolution seismic data processing model as high-resolution data.

[0112] For example, in general, the number of log data traces in the target terrain is low, but the number of seismic data traces is high, Thus, high-resolution seismic data is calculated using only the small number of traces from the logging data, and a first part of seismic data (the seismic data) is entered into a pre-trained model to obtain a second part of seismic data; then, the second part of seismic data is entered into the trained processing model to obtain target high-resolution data, thus enabling the acquisition of high-resolution data corresponding to all the seismic data of a target terrain profile.

[0113] Based on big data analysis and seismic wave propagation theory, a multi-band, multi-scale signal analysis is performed on seismic and well logging data to guide the construction of a high-resolution, multi-frequency, nonlinear model. The central point is to implement the interactive and cooperative fusion of multi-frequency, multi-scale features, and the high-resolution well logging data is used as a constraint to optimize the high-resolution, multi-scale, multi-frequency processing apparatus and process. This solution can be directly applied to any real-world work site in different regions with the same distribution as the database, and allows for the rapid acquisition of high-definition seismic images, even in real time.This solution can also be applied to any real working terrain in different regions of different distributions, and it is sufficient to appropriately modify the model of the new process and the new apparatus provided by this disclosure using broadband logging data of the working terrain, and finally to apply the modified high-resolution model to real pre-stacking or post-stacking seismic data to achieve high-resolution nonlinear multi-scale processing, and obtain high-resolution data with a wider frequency band than the original seismic data, so that the processed images are clearer and more detailed.

[0114] Fig. 10 illustrates a schematic diagram of high-resolution processing results of field data. Fig. 10 (a) illustrates field data, in which a black curve represents high-resolution data extracted from an actual well at that location, and an area indicated by the black arrows is a target stratum, which existed in log interpretations and seismic attributes but cannot be displayed in the seismic data. Fig. 10 (b) illustrates high-resolution processing results based on a pre-trained model; as can be seen in the profile, the resolution is clearly improved and the stratification is good, but the correlation with the well is weak, and no target stratum appears.10] (c) illustrates high-resolution processing results obtained by transfer learning on a pre-trained model using an ordinary depth convolutional neural network; at that time, components. The low-frequency response of the profile is clearly improved, but it is not continuous, and the profile loses much of its transverse continuity. Figure 10(d) illustrates high-resolution processing results obtained by transfer learning on a pre-trained model using a high-resolution, nonlinear, multi-frequency-scale operator; in a target stratum indicated by the black arrows, a new event with strong continuity appears, while the profile maintains a high signal-to-noise ratio and transverse continuity. In Figure 10, CDP is a common depth collection point.

[0115] Figure 11 illustrates a schematic diagram comparing amplitude spectra of real log data used for training in the frequency domain. In this diagram, a black curve represents an amplitude spectrum of field data, a light gray curve represents an amplitude spectrum of an output from a pre-trained model without well intervention, a black dashed line represents an amplitude spectrum of an output from a transfer learning process of an ordinary deep convolutional neural network, and a black dashed line represents an amplitude spectrum of an output from a transfer learning process of a multi-frequency-scale network. As can be seen from an area indicated by a black circle in the drawing, only a high-resolution, nonlinear, multi-frequency-scale operator obtains the numerous high-frequency inversion features in a limited number of iterations.

[0116] Figure 12 illustrates a schematic diagram of a comparison between amplitude spectra of profile data in the frequency domain, in which the curve names are the same as those in Figure 11. As can be seen from an area indicated by a black circle in Figure 12, a high-resolution, nonlinear, multi-frequency scaled operator favors the high-frequency inversion capability over the entire profile, but the ordinary convolutional neural network has only a weak high-frequency response over the entire profile.

[0117] Figure 13 illustrates a schematic diagram of high-resolution processing results (enlarged target terrains) of field data. Figure 13(a) illustrates field data, on which a white circle indicates a target stratum, the upper and lower bounds of the target stratum are marked with short black lines, but the target stratum does not appear in the seismic data. Figure 13(b) illustrates high-resolution processing results obtained by transfer learning on a pre-trained model using an ordinary deep convolution neural network; at this time, high energy concentrations appear near a well at the level of a target stratum, but they are not efficiently extended to the well's outflow, and there is a possibility of overfitting. Figure 13(c) illustrates a high-resolution processing result obtained by transfer learning on a pre-trained model using a high-resolution nonlinear multi-frequency scale operator, on which a continuous layer is generated in the target layer, which is consistent with the layer on the well, thus showing that the process of the present disclosure has certain advantages.

[0118] Figure 14 illustrates a schematic diagram for comparing seismic data synthesis obtained by convolution of high-resolution data with seismic wavelets, on which a leftmost curve is a primary wave velocity curve, a dark curve in the middle is composed of five identical data traces after log data are convolved with seismic wavelets to synthesize seismic data; a light-colored curve illustrates high-resolution processing results obtained by a high-resolution nonlinear multi-frequency-scale operator, and is composed of five well bypass traces.It can be seen that a thin layer in the seismic data synthesized from the log data shows a good match on the seismic profile inverted by the high-resolution nonlinear multi-frequency scale operator, and particularly in a target terrain in a black dashed line frame, reflects that the process proposed by this disclosure can be used for high-resolution processing to obtain a high-fidelity result.

[0119] This disclosure proposes a processing model training apparatus for training the high-resolution processing model of seismic data. The apparatus comprises a first acquisition unit, an acquisition unit, and an iterative training unit.

[0120] The first acquisition unit is configured to obtain seismic data samples from a target location in a target terrain.

[0121] The second acquisition unit is configured to obtain log data from an actual well at the target location in the target terrain, and to determine high-resolution seismic data samples based on the log data.

[0122] The iterative training unit is configured to perform iterative training of the high-resolution seismic data processing model by taking the seismic data samples as input to the high-resolution seismic data processing model and the high-resolution seismic data samples as the output target to the high-resolution seismic data processing model.

[0123] In certain embodiments, the apparatus further comprises: an extraction unit configured to perform statistical wavelet extraction on seismic data samples to obtain a wavelet sample database with seismic data characteristics; a synthesis unit configured to convolve a reflectivity dataset with seismic wavelets in the wavelet sample database to obtain a synthetic seismic sample database with a very large volume of data; and a pre-training unit configured to pre-train the processing model by taking the synthetic seismic sample database as the input to the processing model and the synthetic data reflectivity dataset as the output target to the processing model.

[0124] In certain embodiments, when iterative training of the processing model is performed, an iteration direction is determined on the basis of an objective function comprising a difference term and a constraint term, and the difference term comprises a difference between a target output value and an actual output value of the processing model, and the constraint term comprises a sum of a difference corresponding to each of the processing units; each of the processing units comprises a high-frequency processing unit and / or a low-frequency processing unit, and the difference corresponding to a processing unit is a difference between the target output value and the actual output value of the processing unit.

[0125] The present disclosure proposes a high-resolution seismic data processing device, comprising an application unit configured to input seismic data from the target terrain into the high-resolution seismic data processing model, and to take output data from the high-resolution seismic data processing model as high-resolution data.

[0126] The descriptions and functions of the above devices can be understood by reference to the high-resolution seismic data processing model, the drive method or the high-resolution seismic data processing method, and will not be repeated here.

[0127] One embodiment of the present disclosure further proposes an electronic device, as illustrated in [Fig. 15], which may include a processor 1501 and a memory 1502, and the processor 1501 and the memory 1502 may be connected by a bus or other means, and a connection by a bus is taken as an example in [Fig. 15],

[0128] The 1501 processor may be a central processing unit (CPU). The 1501 processor may also be any other general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a user-programmable pre-diffused array (FPGA) or any other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component or other chips, or combinations of the various chips above.

[0129] As a computer-readable, non-transient storage medium, memory 1502 can be used to store non-transient software programs, non-transient computer-executable programs and modules, such as program instructions / modules corresponding to the high-resolution seismic data processing method in the embodiments of this disclosure. The processor 1501 performs various processor-related functional applications and data processing by running the non-transient software programs, instructions, and modules stored in memory 1502, i.e., to implement the high-resolution seismic data processing model, the drive method, or the high-resolution seismic data processing method in the above-mentioned process embodiments.

[0130] The memory 1502 may include a program storage area and a data storage area, and the program storage area may store an operating system and an application program required by at least one function; the data storage area may store data created by the processor 1501 and similar devices. Furthermore, the memory 1502 may include high-speed random access memory and non-transient memory such as at least one disk memory device, a flash memory device, or any other non-transient solid-state memory device. In some embodiments, the memory 1502 may optionally include memories located remotely from the processor 1501, and these remote memories may be connected to the processor 1501 via a network.Examples of the networks mentioned above include, but are not limited to, the Internet, an intranet, a local area network, a mobile communications network, and combinations thereof.

[0131] The module(s) are stored in memory 1502 and, when executed by processor 1501, implement the high-resolution seismic data processing model, the drive method or the high-resolution seismic data processing method.

[0132] Details of the electronic device can be understood by reference to the corresponding descriptions and effects of the high-resolution seismic data processing model, the drive method or the high-resolution seismic data processing method, and will not be repeated here.

[0133] This disclosure further proposes a computer storage medium, which stores computer program instructions, when executed, implement steps of the high-resolution seismic data processing model, the training method or the high-resolution seismic data processing method.

[0134] This disclosure further offers a computer program product comprising a computer program, and the computer program, when executed by a processor, implements steps of the high-resolution seismic data processing model, the training process, or the high-resolution seismic data processing process.

[0135] As a person skilled in the art can understand, all or part of the procedures in the processes of the above embodiments can be implemented by giving instructions to the related hardware by means of a computer program stored on a computer-readable storage medium, and when executed, the program includes the procedures of the processes of the above embodiments. In this embodiment, the storage medium can be a magnetic disk, an optical disk, read-only memory (ROM), random-access memory (RAM), flash memory, a hard disk drive (HDD), a solid-state drive (SSD), etc. The storage medium can further include combinations of the above-mentioned memory types.

[0136] The embodiments here are described progressively, and identical or similar parts of the embodiments may refer to one another. Each embodiment emphasizes its differences from the other embodiments.

[0137] Any system, device, module or unit shown in the embodiments can specifically be implemented by a computer chip or entity, or by a product having a certain function.

[0138] For ease of description, the device is described by dividing it into different units based on its functions and describing them respectively. Of course, the functions of the different units can be implemented in the same software and / or hardware when this disclosure is implemented.

[0139] From the description of the embodiments above, a person skilled in the art will understand that this disclosure can be implemented using software and a general hardware platform. Based on this understanding, the technical solution of this disclosure can be substantially, or include a portion contributing to the prior art, implemented in the form of a software product, which can be stored on a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include several instructions to enable a computing device (a personal computer, a server, or a network device, etc.) to implement the methods described in various embodiments or parts thereof in this disclosure.

[0140] This disclosure can be applied in many general-purpose or specific-purpose computer system environments or configurations, for example, a personal computer, a server computer, a handheld or portable device, a tablet device, a multiprocessor system, a system with microprocessor base, a decoder, a programmable consumer electronic device, a networked PC, a minicomputer, a mainframe computer, a distributed computing environment comprising any of the aforementioned systems or devices, etc.

[0141] This disclosure can be described in the general context of computer-executable instructions executed by the computer, for example, a program module. Generally, a program module comprises a routine, program, object, component, data structure, etc., performing a particular task or realizing a particular type of abstract data. This disclosure can also be implemented in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can be located in both local and remote computing storage media, including the storage device.

[0142] Although the present disclosure has been described through embodiments, those skilled in the art know that there are many modifications and variations in the present disclosure without departing from its spirit, and it is anticipated that the attached claims will incorporate these modifications and variations without departing from the spirit of the present disclosure.

Claims

Claims

1. A high-resolution seismic data processing model, comprising: a scale factor determination unit configured to receive seismic data of a target terrain, determine a dominant frequency band and a high frequency band of the seismic data, and determine a stretching scale factor corresponding to the high frequency band, wherein the stretching scale factor is applied to the high frequency data in a time domain to compress the high frequency data into a dominant frequency band range; a dominant frequency processing unit configured to perform high-resolution processing on the original-scale seismic data to obtain first high-resolution seismic data; a high-frequency processing unit configured to perform stretching processing on the time-domain seismic data using the stretching scale factor to obtain seismic data at a first scale, and perform high-resolution processing on the first-scale seismic data to obtain second high-resolution seismic data; and a fusion unit configured to fuse the first high-resolution seismic data and the second high-resolution seismic data in the time domain to obtain super-resolution broadband seismic data.

2. The processing model of claim 1, wherein the scale factor determination unit is further configured to determine a low-frequency band and a compression scale factor corresponding to the low-frequency band; correspondingly, the high-resolution seismic data processing model further comprises a low-frequency processing unit configured to perform compression on the seismic data in the time domain with the compression scale factor to obtain seismic data at a second scale, and perform high-resolution processing on the seismic data at the second scale to obtain third high-resolution seismic data; and the fusion unit fuses the first high-resolu- solution, the second high-resolution seismic data and the third high-resolution seismic data in the time domain to obtain the super-resolution broadband seismic data.

3. A high-resolution processing model training method for training the high-resolution seismic data processing model according to claim 1 or 2, the method comprising: obtaining seismic data samples of a target location in a target terrain; obtaining logging data of an actual well at the target location in the target terrain, and determining high-resolution seismic data samples based on the logging data; and performing iterative training of the high-resolution seismic data processing model by taking the seismic data samples as an input of the high-resolution seismic data processing model and the high-resolution seismic data samples as an output target of the high-resolution seismic data processing model.

4. The method of claim 3, wherein before performing the iterative training of the high-resolution seismic data processing model by taking the seismic data samples as an input of the high-resolution seismic data processing model and the high-resolution seismic data samples as an output target of the high-resolution seismic data processing model, the method further comprises a pre-training method for the high-resolution seismic data processing model comprising: performing statistical wavelet mining on the seismic data samples to obtain a wavelet sample database of the seismic data samples;convolving a geological model data reflectivity set with seismic wavelets in the wavelet sample database to obtain a synthetic seismic data sample database; and pre-training the high-resolution seismic data processing model by taking the synthetic seismic data sample database as an input of the model;

5.

6.

7. high-resolution seismic data processing and the geological model data reflectivity set as the output target of the high-resolution seismic data processing model. The method of claim 3, wherein during iterative training of the high-resolution seismic data processing model, an iteration direction is determined based on an objective function comprising a difference term and a constraint term, wherein the difference term comprises a difference between an output target value and an actual output value of the high-resolution seismic data processing model, and the constraint term comprises a sum of a difference corresponding to each of the processing units; each of the processing units comprises a high-frequency processing unit and / or a low-frequency processing unit, and the difference corresponding to a processing unit is a difference between the output target value and the actual output value of the processing unit. Method for high-resolution processing of seismic data, comprising: inputting seismic data of a target terrain into the high-resolution seismic data processing model according to claim 1 or 2, and taking output data from the high-resolution seismic data processing model as high-resolution seismic data. A processing model training apparatus for training the high-resolution seismic data processing model according to claim 1 or 2; the apparatus comprising: a first obtaining unit configured to obtain seismic data samples of a target location in a target terrain; a second obtaining unit configured to obtain logging data of an actual well at the target position in the target terrain, and determine high-resolution seismic data samples on the basis of the logging data; and an iterative training unit configured to perform iterative training of the high-resolution seismic data processing model by taking the seismic data samples as an input of the high-resolution seismic data processing model and the high-resolution seismic data samples as an output target of the high-resolution seismic data processing model. seismic data.

8. A high-resolution seismic data processing apparatus, comprising: an application unit configured to input seismic data of a target terrain into the high-resolution seismic data processing model according to claim 1 or 2, and take output data from the high-resolution seismic data processing model as high-resolution seismic data.

9. 9 An electronic device, comprising: a memory and a processor in communicative connection with each other, wherein the memory stores computer instructions, and the processor implements the steps of the method according to any one of claims 3 to 6 by executing the computer instructions.

10. 10 A computer storage medium, wherein the computer storage medium stores computer program instructions, wherein when executed, the computer program instructions implement the steps of the method according to any one of claims 3 to 6.

Images