Depth compressed sensing reconstruction method for large contiguous depressed seismic data

Abstract

The invention provides a depth compressed sensing reconstruction method for depressed large contiguous seismic data, and the method comprises the steps: 1, carrying out the performance analysis of an original observation system, and selecting observation parameters; 2, calculating an azimuth angle and an offset distance; step 3, mapping a new observation system and reconstructing compressed sensing data; 4, deep learning attribute matching is carried out; and step 5, when an expected condition is satisfied, outputting mapping data under a new observation system. According to the depth compressed sensing reconstruction method for the depression-level large continuous seismic data, the seismic data between different observation systems is mapped and reconstructed by using a depth compressed sensing method, the seismic data difference of different blocks of a large continuous region can be eliminated, and the efficiency is higher.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploration seismic data processing, and in particular to a method for deep compression sensing reconstruction of depression-level large-scale seismic data. Background Art

[0002] Foreign experts and scholars have taken the lead in conducting research on the combination of compressed sensing theory and seismic data acquisition and processing. Offshore seismic data is collected by random observation at sea, and the seismic wave field is recorded by measuring and arranging the random positions of adjacent geophones. This random undersampling method is convenient for field construction. On this basis, new observation systems can be designed to select the optimal shot points and geophone points under constraints, and the non-uniform sampling observation system can be applied to the acquisition of marine and land seismic data. Land seismic data acquisition is stimulated by controllable seismic sources, and irregular three-dimensional observation systems are designed, and restored and reconstructed to obtain better seismic data. Compressed sensing technology can realize the reconstruction of seismic data in the later stage, and many scholars have conducted theoretical analysis and discussion on the feasibility of this scheme. For example, by establishing a local area of ​​an irregular observation system of compressed sensing and using interpolation technology to reconstruct data. Or by designing a sparse observation system and applying it to the reconstruction of theoretical models. At present, most foreign countries have carried out certain observation system designs and compressed sensing combined schemes in deserts and offshore areas; while China mainly stays at the reconstruction of model data, and there is no observation system design combined with compressed sensing.

[0003] For the reconstruction of seismic data, the wave equation can be used. This method mainly restores and reconstructs based on the propagation principle of seismic waves, and the amount of calculation is relatively large. Seismic data is restored and reconstructed based on predictive filtering, such as the FX domain reconstruction method. After some mathematical transformations, the seismic data is restored and reconstructed through threshold processing. This method is greatly affected by the transformation method. The Nyquist-Shannon sampling theorem has always limited the quality of data reconstruction in signal processing, and the introduction of compressed sensing theory has fundamentally changed the status quo of signal reconstruction. Under the theoretical framework of compressed sensing, real data can often be reconstructed better with a relatively small amount of data. On the basis of the theory of compressed sensing, sparse representation of data has become a prerequisite for the reconstruction problem. According to the source of sparse representation, it can be divided into: seismic data recovery and reconstruction methods using a single mathematical transformation for processing, seismic data recovery and reconstruction by superposition of different types of mathematical changes, and seismic data recovery and reconstruction methods using learning dictionaries. Seismic data reconstruction with a single mathematical transformation can be understood as using a single fixed mathematical transformation to sparsely represent data. On this basis, more scholars have tried to bring more superior mathematical transformations into the compressed sensing system and continuously improve it. By further improving the fast iterative shrinkage threshold algorithm, a fast convex set projection method was developed, which has greatly improved the effect compared with the traditional Fourier transform. The unsampled wavelet transform, which is sparser than the Fourier transform in mathematical transformation, is used to restore and reconstruct seismic data, thereby obtaining a better coefficient representation, and its reconstruction effect is also significantly improved. The shear wave transform with scale transformation and more completeness is introduced into the restoration and reconstruction of seismic data, and the effect obtained is better than the reconstruction effect obtained by the curvelet transform, and the signal-to-noise ratio is higher.

[0004] Since the seismic data collected in practice often do not have a single feature, a single mathematical transformation method may not be able to accurately sparsely represent the seismic data. In view of this, the morphological component analysis method in the field of image processing is borrowed from the field of seismic exploration to restore and reconstruct seismic data, achieving better restoration and reconstruction results than a single sparse representation. The principle of the morphological component analysis method is to use a divide-and-conquer strategy for complex seismic data with multiple properties. The best sparse representation method corresponding to each characteristic is used for sparse representation, and finally the processed results are superimposed. It is for this reason that when the morphological component analysis method selects a sparse representation basis, different sparse representation bases have different reconstruction effects. Combining different types of sparse representation methods, the combination of the final curvelet dictionary and the discrete cosine transform dictionary has better sparse representation capabilities than other dictionary combinations, and can achieve better restoration and reconstruction effects. Dictionary learning, which is to learn the dictionary to obtain the ideal sparse representation capability, is another effective sparse representation method. The dictionary learning algorithm is a process of sparsely representing and re-characterizing seismic data based on the extracted data features as samples. The resulting sparse representation dictionary can also be used to restore and reconstruct the data well. The dictionary learning method of K-singular value decomposition has also been introduced into seismic data recovery and reconstruction, and has been developed into a more efficient dictionary learning method, namely the data-driven tight frame algorithm as a new sparse dictionary training method. Under the premise of the same effect, its efficiency is higher than the K-singular value decomposition method.

[0005] Looking at the existing compressed sensing data reconstruction technologies at home and abroad, only the reconstruction of local regular missing or irregular missing pre-stack seismic data is achieved. In field seismic exploration, due to the existence of obstacle areas or irregular surfaces, seismic data are inevitably partially missing. Existing compressed sensing technology can be used to achieve data reconstruction and repair in this case. However, there is no feasible and practical implementation strategy for large-scale / ultra-large-scale seismic data reconstruction involving data mapping of different observation systems for depression-level large-scale seismic data.

[0006] In the Chinese patent application with application number: CN202010235260.2, a method for seismic data reconstruction using a compressed sensing algorithm is involved, including: using an overcomplete dictionary to replace the basis function to perform sparse transformation on the original data; selecting an initialized random unit matrix as the observation matrix; combining the sampling matrix with regularization and then using the orthogonal matching tracking algorithm to restore and reconstruct the missing seismic data. This invention proposes improvements to the processes of signal sparse representation, observation matrix design, signal reconstruction, and data reconstruction in the compressed sensing theory, further enhancing weak signals, and significantly improving seismic resolution.

[0007] In the Chinese patent application with application number: CN201811434030.8, a method for seismic data reconstruction based on spatially constrained compressed sensing is involved, including: using a portion of the data as training data, using K-SVD dictionary learning to train an over-complete dictionary to reconstruct the original seismic data; using the joint sparse decomposition method to extract shared spatial information, and transforming the perception matrix in the compressed sensing algorithm; improving the sparsity adaptive matching pursuit algorithm, introducing the initial sparsity estimation method, and using a variable step size strategy to reconstruct the data. The reconstructed results are not only clearer in details, the operation time is greatly reduced compared to IRLS and SAMP, and the lateral transition is smoother, indicating that the algorithm designed by the present invention utilizes the relevant information of the space, and the reconstruction result is more realistic.

[0008] In the Chinese patent application with application number: CN202011105556.9, a method, device, electronic device and medium based on compressed sensing seismic data reconstruction are involved. The method may include: establishing an initial relationship between unknown seismic data and linear measurement vectors; converting the initial relationship into a sparse domain relationship; calculating the sparse solution of the sparse domain relationship, which is the unknown seismic data. The invention effectively suppresses the interference of multiple waves, improves the attributes of facets, and improves the quality of offset imaging. The research significance of the reconstruction of missing seismic information for the subsequent interpretation and processing of seismic data, and the reconstruction of seismic data will help further seismic data processing and serve the exploration and development of conventional oil and gas and coalbed methane.

[0009] In the Chinese patent application with application number: CN201710181390.0, a method for real-time compression and high-precision reconstruction of wavelet domain seismic data based on compressed sensing is involved, which includes the following steps: first, sparsely represent the microseismic signal in the wavelet domain; then, construct the chaotic Bernoulli measurement matrix (CBMM) using the Logistic chaotic sequence, and use the measurement matrix to compress the observation of the sparsely represented microseismic signal; finally, use the Bayesian wavelet tree structure compressed sensing reconstruction algorithm (BTSWCS) to restore the complete original data. The actual comparative experimental results of the invention show that: using this method to compress data with a total of 28 sampling points, the compression time can be shortened to 10-5s, that is, if the seismograph sampling rate is 1KSPS, the CBMM measurement matrix can basically achieve real-time compression of the 0.25s data collected by it. Under low signal-to-noise ratio conditions, the reconstruction algorithm of the invention improves the PSNR value by at least 5dB, which significantly improves the peak signal-to-noise ratio and reduces the reconstruction error compared to the greedy algorithm.

[0010] The above existing technologies are all significantly different from the present invention and fail to solve the technical problem we want to solve. For this reason, we have invented a new deep compressed sensing reconstruction method for depression-level large-scale seismic data. Summary of the invention

[0011] The purpose of the present invention is to provide a method for mapping and reconstructing seismic data between different observation systems using a deep compressed sensing method, which can eliminate the differences in seismic data between different blocks in the Dalian area and has a higher efficiency of deep compressed sensing reconstruction of depression-level Dalian seismic data.

[0012] The purpose of the present invention can be achieved by the following technical measures: a method for reconstructing deep compressed sensing of depression-level Dalian seismic data, the method comprising:

[0013] Step 1: Analyze the performance of the original observation system and select observation parameters;

[0014] Step 2, calculate the azimuth and offset;

[0015] Step 3: Map the new observation system and reconstruct the compressed sensing data;

[0016] Step 4: Perform deep learning attribute matching;

[0017] Step 5: When the expected conditions are met, the mapping data under the new observation system is output.

[0018] The purpose of the present invention can also be achieved by the following technical measures:

[0019] In step 1, based on the acquisition parameters of the original observation system, the performance indicators of the observation system are calculated, including the minimum shot offset, the maximum shot offset, the minimum track distance and the maximum track distance. After determining the minimum and maximum shot offsets and the minimum and maximum track distances, the seismic data of the original observation system can be mapped to the new observation system under the premise of ensuring that the performance of the observation system remains unchanged.

[0020] In step 1, the minimum offset is calculated as:

[0021]

[0022] Where t0 represents the two-way normal travel time of the target layer; v represents the average velocity of the formation; D represents the rate of change of the wavelength of the seismic wavelet in the time dimension; x min Represents the minimum offset of the original observation system.

[0023] In step 1, the maximum offset is calculated as:

[0024]

[0025] Where t0 represents the two-way normal travel time of the target layer; f is the main frequency of the seismic wavelet; v1 represents the velocity of the first reflection wave of the seismic wave; v2 represents the velocity of the multiple reflection waves of the seismic wave; x max Represents the maximum offset of the original observation system.

[0026] In step 1, the minimum track distance is calculated as:

[0027]

[0028] Where Δx min represents the minimum track distance; v int Represents the layer velocity of the underground strata; f max is the peak frequency of the seismic wavelet; φ represents the formation inclination.

[0029] In step 1, the maximum track distance is calculated as:

[0030]

[0031] Where Δx max represents the maximum track distance; v int Represents the layer velocity of the underground strata; f dom It is the center frequency value of the dominant frequency band of the seismic wavelet.

[0032] In step 2, the azimuth and offset distribution of the seismic data are calculated using the designed new observation system parameters, and compared with the original observation system to establish an offset mapping relationship.

[0033] In step 3, the compressed sensing technology based on the morphological component analysis method is used to realize the data mapping and reconstruction between the two sets of new and old observation systems; in the process of compressed sensing data reconstruction, a divide-and-conquer strategy is adopted, that is, various characteristics are sparsely represented using the corresponding optimal sparse representation method, and finally the processed results are superimposed.

[0034] In step 3, for the seismic signal s, if it has k different morphological features, as shown in formula (5):

[0035]

[0036] In the formula, the seismic signal is represented as a linear combination of k morphological components, which can be represented by a sparse dictionary D k To express, a k is the corresponding sparse coefficient;

[0037] In the framework of morphological component analysis, the equation can be rewritten as:

[0038]

[0039] In the formula, λ is the L1 modulus coefficient of the sparse coefficient; y is the seismic signal under the new observation system; R is the amplitude matching factor of the data mapping; γ is the L1 modulus coefficient of the sparse dictionary represented as the objective function; the addition of this term has little effect on the reconstruction of locally missing seismic data, but it is indispensable for the large-scale reconstruction of seismic data from different observation systems at the depression level, otherwise it will seriously affect the matching degree of the reconstructed big data with the seismic data of the original observation system in terms of amplitude and energy; the other variables have the same meaning as formula (5).

[0040] In step 3, the reconstruction process of compressed sensing data is carried out in the order of common shot point gathers, common detection point gathers, and finally common CMP gathers / common azimuth gathers. This can better maintain the consistency characteristics of the actual seismic wave field and is also beneficial to ensure the accuracy of seismic data reconstruction.

[0041] In step 3, the steps of reconstructing the compressed sensing data specifically include:

[0042] Step 1: Map the data in each shot gather according to the offset distribution of the new observation system to obtain the seismic records at the corresponding geophones;

[0043] Step 2: After each shot gather is completed, it is re-extracted into a common detection point gather. In each common detection point gather, data is reconstructed and mapped according to the shot point distribution of the new observation system to obtain the seismic record of each shot point;

[0044] Step 3: After the first two steps are completed, the preliminary seismic data under the new observation system are obtained, and they are re-extracted into common CMP gathers and common azimuth gathers, and their distribution such as offset and azimuth is analyzed to ensure that the performance of the new and old observation systems remains unchanged.

[0045] In step 4, after completing the large-scale data reconstruction of the new observation system, the power spectrum, amplitude spectrum and phase spectrum of the reconstructed seismic data are calculated; based on the compressed sensing data reconstruction, a deep learning strategy is used to match the pre-stack attributes of the new and old data to achieve consistency in amplitude, phase and frequency of the large-scale data.

[0046] In step 5, it is determined that the expected conditions are met, that is, the degree of agreement between the reconstructed data and the original data in terms of pre-stack attributes (such as amplitude, phase, frequency, etc.) reaches more than 95%.

[0047] When the expected condition is not met, the process returns to step 2.

[0048] The deep compressed sensing reconstruction method of depression-level Dalian piece seismic data in the present invention realizes seismic data reconstruction under the premise of ensuring that the performance indicators of the observation system do not change. The observation system performance indicators are calculated with the help of the observation system parameters, and the mapping relationship of the offset distance is calculated on this basis. The compressed sensing technology based on the morphological component analysis method is used to achieve data reconstruction. The L1 modulus coefficient constraint of the sparse dictionary is added in the reconstruction process to better adapt to the reconstruction of depression-level Dalian piece data. Specifically, it is carried out in the order of common shot point gathers, common detection point gathers and common CMP / azimuth gathers. Finally, the deep learning strategy is used to match the dynamic characteristics before and after data mapping to ensure the consistency of different blocks of Dalian piece data in terms of amplitude, phase and frequency, laying the necessary data foundation for subsequent splicing processing and interpretation.

[0049] The advantages of the present invention are: the seismic data of different three-dimensional blocks are very different, the data volume is large, and the processing flow is complex. The adaptive weighted sparse promotion reconstruction method based on morphological component analysis theory is a data-driven method for improving the ability to identify sparse signals. The method uses the characteristics of the signal itself, designs an adaptive nonlinear measurement matrix, constructs a least squares objective function constrained by the L1 norm, adds assumptions to the inversion equation, and realizes high-precision identification and data reconstruction of effective signals through an iterative threshold algorithm, thereby improving the quality of reconstruction of effective signals.

[0050] In addition, since deep learning does not involve missing and down-sampling positions during the training process, the present invention is very flexible and applicable to different types of seismic data channel missing, has a plug-and-play nature, and has high versatility. Therefore, the present invention can flexibly handle the problems existing in seismic data processing of different three-dimensional blocks. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flowchart of a specific embodiment of the method for deep compression sensing reconstruction of depression-level large-scale seismic data of the present invention;

[0052] Figure 2 It is a single shot gather and shot inspection distribution diagram of the original observation system in a specific embodiment of the present invention;

[0053] Figure 3 It is a single shot gather and shot inspection distribution diagram of a new observation system in a specific embodiment of the present invention;

[0054] Figure 4 A schematic diagram of three commonly used observation systems in the industry in a specific embodiment of the present invention, and performing observation system mapping and data reconstruction based on actual data;

[0055] Figure 5 It is a schematic diagram of the superimposed cross-section of reconstructed data under three observation systems in a specific embodiment of the present invention;

[0056] Figure 6 It is a schematic diagram of energy spectra of reconstructed data under three observation systems in a specific embodiment of the present invention;

[0057] Figure 7 It is a schematic diagram of the frequency spectrum of reconstructed data under three observation systems in a specific embodiment of the present invention;

[0058] Figure 8 Schematic diagram of variance spectra of reconstructed data under three observation systems in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0059] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0060] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations and / or combinations thereof.

[0061] The deep compressed sensing reconstruction method for depression-level large-scale contiguous seismic data of the present invention adopts the deep compressed sensing contiguous data efficient reconstruction technology based on morphological component analysis theory to match and integrate the pre-stack data of different blocks, thereby improving the accuracy of data reconstruction; the deep compressed sensing (DCS) technology combining deep learning and compressed sensing is used to process large-scale three-dimensional block contiguous processing, thereby improving the efficiency of data reconstruction.

[0062] like Figure 1 As shown, the method for reconstructing depression-level large-scale seismic data based on deep compressed sensing algorithm of the present invention includes:

[0063] Step 1: Analyze the performance of the original observation system and select observation parameters;

[0064] Step 2, calculate the azimuth and offset;

[0065] Step 3: Map the new observation system and reconstruct the compressed sensing data;

[0066] Step 4: Perform deep learning attribute matching;

[0067] Step 5: determine whether the expected conditions are met, that is, the degree of agreement between the reconstructed data and the original data in terms of pre-stack attributes (such as amplitude, phase, frequency, etc.) reaches more than 95%.

[0068] When the expected condition is not met, the process returns to step 2; when the expected condition is met, the process proceeds to step 6.

[0069] Step 6: Output the mapping data under the new observation system.

[0070] The following are several specific embodiments of the present invention.

[0071] Example 1

[0072] In a specific embodiment 1 of the present invention, after repeated tests and effect analysis, the preferred solution is as follows:

[0073] Since there are different types of seismic data in the Dalian area of ​​the depression level, they are affected by many factors such as acquisition system parameters, source type and surface conditions, resulting in huge differences in amplitude, phase and frequency between different seismic data. This mismatch directly leads to differences in structure and lithology in the later stacking or migration sections, misleading the interpretation of seismic data.

[0074] (1) First, the performance indicators of the observation system are calculated based on the acquisition parameters of the original observation system.

[0075] The calculation of the minimum offset is shown in formula (1):

[0076]

[0077] Where t0 represents the two-way normal travel time of the target layer; v represents the average velocity of the formation; and D represents the rate of change of the wavelength of the seismic wavelet in the time dimension. min Represents the minimum offset of the original observation system.

[0078] The calculation of the maximum offset is shown in formula (2):

[0079]

[0080] Where t0 represents the two-way normal travel time of the target layer; f is the main frequency of the seismic wavelet; v1 represents the velocity of the first reflection wave of the seismic wave; v2 represents the velocity of the multiple reflection waves of the seismic wave; x max Represents the maximum offset of the original observation system.

[0081] From formulas (1) and (2), the minimum and maximum offsets of the original observation system can be determined.

[0082] The calculation of the minimum track distance is shown in formula (3):

[0083]

[0084] Where Δx min represents the minimum track distance; v int Represents the layer velocity of the underground strata; f max is the peak frequency of the seismic wavelet; φ represents the formation inclination.

[0085] The calculation of the maximum track distance is shown in formula (4):

[0086]

[0087] Where Δx max represents the maximum track distance; v int Represents the layer velocity of the underground strata; f dom It is the center frequency value of the dominant frequency band of the seismic wavelet.

[0088] After determining the minimum and maximum offsets, and the minimum and maximum trace distances, the seismic data of the original observation system can be mapped to the new observation system while ensuring that the performance of the observation system remains unchanged.

[0089] (2) Secondly, the azimuth and offset distribution of the seismic data are calculated using the parameters of the designed new observation system, and compared with the original observation system to establish an offset mapping relationship.

[0090] (3) Again, the compressed sensing technology based on the morphological component analysis method is used to realize the data mapping and reconstruction between the two sets of new and old observation systems. In the process of compressed sensing data reconstruction, a divide-and-conquer strategy is adopted, that is, the best sparse representation method corresponding to each characteristic is used for sparse representation, and finally the processed results are superimposed.

[0091] For a seismic signal s, if it has k different morphological features, as shown in formula (5):

[0092]

[0093] In the formula, the seismic signal is represented as a linear combination of k morphological components, which can be represented by a sparse dictionary D k To express, a k is the corresponding sparse coefficient.

[0094] In the framework of morphological component analysis, the equation can be rewritten as:

[0095]

[0096] In the formula, λ is the L1 modulus coefficient of the sparse coefficient; y is the seismic signal under the new observation system; R is the amplitude matching factor of the data mapping; γ is the L1 modulus coefficient of the sparse dictionary represented as the objective function. The addition of this term has little effect on the reconstruction of locally missing seismic data, but it is indispensable for the large-scale reconstruction of seismic data from different observation systems at the depression level, otherwise it will seriously affect the matching degree of the reconstructed big data with the seismic data of the original observation system in terms of amplitude and energy. The meanings of other variables are the same as those in formula (5). The specific reconstruction process is carried out in the order of common shot point gathers, common detection point gathers, and finally common CMP gathers / common azimuth gathers, which can better maintain the consistency characteristics of the actual seismic wave field and is also conducive to ensuring the accuracy of seismic data reconstruction.

[0097] (4) Finally, after completing the large-scale data reconstruction of the new observation system, the power spectrum, amplitude spectrum, and phase spectrum of the reconstructed seismic data are calculated. Based on the compressed sensing data reconstruction, a deep learning strategy is used to match the pre-stack attributes of the new and old data to achieve consistency in amplitude, phase, and frequency of the large-scale data.

[0098] Example 2

[0099] In a specific embodiment 2 of the present invention, the parameters of the newly mapped observation system are determined by analyzing the performance of the observation systems of different blocks of the large-scale data, and compressed sensing data reconstruction is performed under the premise of ensuring that the performance of the observation system remains unchanged. The reconstructed data can ensure that the imaging capability and accuracy of underground structures and lithologies are not affected by the mapping of the observation system. Deep learning attribute matching is crucial to maintaining the dynamic characteristics of the original seismic data, such as amplitude, phase, and frequency. The design of this link can achieve the fidelity and reliability of seismic data reconstruction. After the seismic data reconstruction meets the expected conditions (for example, the reconstruction error is controlled within 5%), it is terminated. Otherwise, the iteration continues.

[0100] Firstly, the key performance indicators under the acquisition parameters of the original observation system are calculated based on formulas (1), (2), (3) and (4), including the minimum offset, maximum offset, minimum track distance and maximum track distance.

[0101] Secondly, under the premise that the performance indicators of the observation system remain unchanged, an offset mapping relationship is established, that is, the specific distribution of the original offset is mapped to the new observation system to obtain the offset distribution in the new system.

[0102] The seismic data reconstruction is realized by using the compressed sensing technology based on the morphological component analysis method, and formulas (5) and (6) are used. In this step, because it is necessary to process the 3D seismic data of the large contiguous area at the depression level, the most efficient and economical way is to gradually reconstruct the data according to the principle of first common shot gathers, then common receiver gathers, and finally common CMP gathers and common azimuth gathers.

[0103] Step 1: Data mapping is performed in each shot gather according to the offset distribution of the new observation system to obtain the seismic records at the corresponding detectors.

[0104] Step 2: After each shot gather is completed, it is re-extracted into a common detection point gather. In each common detection point gather, data is reconstructed and mapped according to the shot point distribution of the new observation system to obtain the seismic record of each shot point.

[0105] Step 3: After the first two steps are completed, the preliminary seismic data under the new observation system are obtained, and they are re-extracted into common CMP gathers and common azimuth gathers, and their distribution such as offset and azimuth is analyzed to ensure that the performance of the new and old observation systems remains unchanged.

[0106] Finally, the power spectrum, amplitude spectrum and phase spectrum of the reconstructed sag-level Dalian seismic data are calculated. A deep learning strategy is used to match the prestack attributes of new and old data to ensure that the dynamic characteristics before and after data reconstruction do not change.

[0107] From a theoretical analysis, the technology for reconstructing depression-level large-scale seismic data based on a deep compressed sensing algorithm proposed in the present invention has higher reconstruction accuracy and reconstruction efficiency.

[0108] Compressed sensing (CS) theory is a process of signal information extraction and recovery, which can effectively overcome the spatial aliasing phenomenon. This theory breaks the traditional Nyquist sampling concept, uses the sparsity or compressibility of the signal, and achieves accurate reconstruction of the signal by solving the regularized inversion problem. A measurement matrix that meets the RIP criterion is used to perform low-dimensional projection on the noisy signal (when the observation dimension is sufficient to contain effective information, part of the noise information will be discarded after projection due to the lack of sparseness of the noise, and this part of the noise cannot be restored during signal reconstruction). Through a suitable sparse decomposition algorithm, the basis function that best matches the effective signal can be selected to achieve the extraction of the effective signal.

[0109] The deep compressed sensing (DCS) framework jointly trains the generator and optimizes the reconstruction process through meta-learning. It breaks through the problem of insufficient training labels required by deep learning methods. The technology has high flexibility and sampling efficiency, and significantly improves the signal recovery performance and speed. It mainly includes two parts:

[0110] First, a set of convolutional neural network (CNN) denoisers are trained to learn denoising from a given training dataset. The CNN is trained using synthetic seismic data modelled by wave equations, which allows the network to extract feature information and generate feature memory. The network parameters are optimized using the Adam optimizer, and data augmentation based on rotation and flipping is used during mini-batch learning. The learning rate is set to 0.001 at the beginning of training and then fixed to 0.0001 when the training loss stops decreasing. Training is terminated if the training loss is fixed for five consecutive time periods. To reduce the overall training time, adjacent denoisers are initialized during training using the model obtained at the previous noise level.

[0111] Secondly, the trained CNN denoiser is integrated into the compressed sensing framework for seismic data reconstruction. In the compressed sensing reconstruction algorithm, the noise variance of the POCS algorithm continues to decrease in each iteration, so the recovered seismic data gradually approaches noise-free. And the POCS framework allows nested denoisers, because it helps improve the performance of the POCS method on seismic interpolation. Therefore, we integrate the trained CNN denoiser into the POCS framework to maximize the performance of the CNN denoiser and POCS.

[0112] In order to prove the application effect of this method, actual tests are carried out using actual single shot data. Figure 2 (a) shows the single-shot gather of the original observation system. Figure 2 (b) shows the shot detection distribution map of the original observation system shot gather. Figure 3 (a) shows the single-shot gather of the new observation system. Figure 3 (b) shows the shot distribution map of the new observation system shot gather.

[0113] From the comparison of the two figures, it can be seen that the data reconstruction technology of the sag-level Dalian piece seismic data based on the deep compressed sensing algorithm realizes the mapping of data from different observation systems and ensures the dynamic characteristics before and after data reconstruction. This technology can be used to match and reconstruct the sag-level Dalian piece data, laying a solid preliminary foundation for the consistent processing and interpretation of the Dalian piece three-dimensional block.

[0114] Example 3

[0115] In the specific embodiment 3 of the present invention, the aforementioned deep compressed sensing reconstruction method is used to reconstruct the actual data seismic data of three different observation systems in the industry, and its actual application effect is analyzed. Unlike the single-shot test based on embodiment 2, this embodiment 3 is tested based on all pre-stack seismic data, so the amount of data and the amount of calculation are greatly increased, reaching the scale of industrial data processing. It has been proved through testing that the deep compressed sensing reconstruction method proposed in the present invention can still be applied efficiently and robustly and obtain better results.

[0116] Figure 4 Three observation systems commonly used in the industry were adopted, and observation system mapping and data reconstruction were carried out based on actual data. Figure 5 (a), (b), and (c) are the stacked sections of the reconstructed data under the three observation systems. It can be seen that the seismic data under different observation systems have been well stacked and imaged, and the imaging quality is similar. This reflects that the deep compressed sensing reconstruction method proposed in the present invention can be applied to the mapping and data reconstruction between any different observation systems.

[0117] Based on the deep compressed sensing reconstruction method of the Dalian depression-level seismic data, the attribute analysis of the data reconstructed by three different observation systems was carried out to test the reconstruction effect. Figure 6 (a), (b), and (c) are the energy spectra of the reconstructed data under three observation systems respectively; Figure 7 (a), (b), and (c) are the spectra of the reconstructed data under three observation systems respectively; Figure 8 (a), (b), and (c) are the variance spectra of the reconstructed data under three observation systems. It can be seen that although the observation systems are quite different, this method has achieved good results and has good adaptability.

[0118] It can be seen from the above embodiments that the depression-level Dalian seismic data reconstruction technology based on the deep compressed sensing algorithm proposed in this patent can effectively improve the accuracy and efficiency of seismic data reconstruction, thereby better providing the necessary data guarantee for the consistency processing of Dalian seismic data.

[0119] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0120] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. A deep compressed sensing reconstruction method for depression-level Dalian seismic data, characterized in that: The deep compressed sensing reconstruction method of the depression-level Dalian seismic data includes: Step 1: Analyze the performance of the original observation system and select observation parameters; Step 2, calculate the azimuth and offset; Step 3: Map the new observation system and reconstruct the compressed sensing data; Step 4: Perform deep learning attribute matching; Step 5: When the expected conditions are met, the mapping data under the new observation system is output.

2. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 1 is characterized in that: In step 1, based on the acquisition parameters of the original observation system, the performance indicators of the observation system are calculated, including the minimum shot offset, the maximum shot offset, the minimum track distance and the maximum track distance. After determining the minimum and maximum shot offsets and the minimum and maximum track distances, the seismic data of the original observation system can be mapped to the new observation system while ensuring that the performance of the observation system remains unchanged.

3. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 2 is characterized in that: In step 1, the minimum offset is calculated as: Where t0 represents the two-way normal travel time of the target layer; v represents the average velocity of the formation; D represents the rate of change of the wavelength of the seismic wavelet in the time dimension; x min Represents the minimum offset of the original observation system.

4. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 2 is characterized in that: In step 1, the maximum offset is calculated as: Where t0 represents the two-way normal travel time of the target layer; f is the main frequency of the seismic wavelet; v1 represents the velocity of the first reflection wave of the seismic wave; v2 represents the velocity of the multiple reflection waves of the seismic wave; x max Represents the maximum offset of the original observation system.

5. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 2 is characterized in that: In step 1, the minimum track distance is calculated as: Where Δx min represents the minimum track distance; v int Represents the layer velocity of the underground strata; f max is the peak frequency of the seismic wavelet; φ represents the formation inclination.

6. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 2 is characterized in that: In step 1, the maximum track distance is calculated as: Where Δx max represents the maximum track distance; v int Represents the layer velocity of the underground strata; f dom It is the center frequency value of the dominant frequency band of the seismic wavelet.

7. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 1 is characterized in that: In step 2, the azimuth and offset distribution of the seismic data are calculated using the designed new observation system parameters, and compared with the original observation system to establish an offset mapping relationship.

8. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 1 is characterized in that: In step 3, the compressed sensing technology based on the morphological component analysis method is used to realize the data mapping and reconstruction between the two sets of new and old observation systems; in the process of compressed sensing data reconstruction, a divide-and-conquer strategy is adopted, that is, various characteristics are sparsely represented using the corresponding optimal sparse representation method, and finally the processed results are superimposed.

9. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 8 is characterized in that: In step 3, for the seismic signal s, if it has k different morphological features, as shown in formula (5): In the formula, the seismic signal is represented as a linear combination of k morphological components, which can be represented by a sparse dictionary D k To express, a k is the corresponding sparse coefficient; In the framework of morphological component analysis, the equation can be rewritten as: In the formula, λ is the L1 modulus coefficient of the sparse coefficient; y is the seismic signal under the new observation system; R is the amplitude matching factor of the data mapping; γ is the L1 modulus coefficient of the sparse dictionary represented as the objective function; the addition of this term has little effect on the reconstruction of locally missing seismic data, but it is indispensable for the large-scale reconstruction of seismic data from different observation systems at the depression level, otherwise it will seriously affect the matching degree of the reconstructed big data with the seismic data of the original observation system in terms of amplitude and energy; The other variables have the same meaning as formula (5).

10. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 9 is characterized in that: In step 3, the reconstruction process of compressed sensing data is carried out in the order of common shot point gathers, common detection point gathers, and finally common CMP gathers / common azimuth gathers. This can better maintain the consistency characteristics of the actual seismic wave field and is also beneficial to ensure the accuracy of seismic data reconstruction.

11. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 10 is characterized in that: In step 3, the steps of reconstructing the compressed sensing data specifically include: Step 1: Map the data in each shot gather according to the offset distribution of the new observation system to obtain the seismic records at the corresponding geophones; Step 2: After each shot gather is completed, it is re-extracted into a common detection point gather. In each common detection point gather, data is reconstructed and mapped according to the shot point distribution of the new observation system to obtain the seismic record of each shot point; Step 3: After the first two steps are completed, the preliminary seismic data under the new observation system are obtained, and they are re-extracted into common CMP gathers and common azimuth gathers, and their distribution such as offset and azimuth is analyzed to ensure that the performance of the new and old observation systems remains unchanged.

12. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 1 is characterized in that: In step 4, after completing the large-scale data reconstruction of the new observation system, the power spectrum, amplitude spectrum and phase spectrum of the reconstructed seismic data are calculated; based on the compressed sensing data reconstruction, a deep learning strategy is used to match the pre-stack attributes of the new and old data to achieve consistency in amplitude, phase and frequency of the large-scale data.

13. The method for deep compressed sensing reconstruction of depression-level large-scale seismic data according to claim 1 is characterized in that: In step 5, it is determined whether the expected conditions are met, that is, whether the degree of agreement between the reconstructed data and the pre-stack attributes of the original data, including amplitude, phase and frequency, reaches more than 95%. If the expected conditions are not met, the process returns to step 2.

Images