Non-ideal illumination passive source seismic wave field intelligent reconstruction method based on sparse active source constraint

Abstract

The invention belongs to the technical field of intelligent seismic exploration, and particularly relates to a sparse active source constrained non-ideal illumination passive source seismic wave field intelligent reconstruction method, so as to solve the problems of false events and coherent noise under the condition of non-ideal distribution of underground passive seismic sources. In order to improve the reconstruction effect of a passive source seismic wave field under a non-ideal illumination condition, the method builds an improved U-Net network, so that the U-Net network can intelligently extract data features and derive an attention mechanism along two independent dimensions of a channel and a space in the network, and multiplies the attention mechanism by an input feature map, so that the reconstruction effect of the passive source seismic wave field is improved. Self-adaptive feature refinement is carried out, so that the purpose of carrying out intelligent reconstruction on passive source data is achieved; by learning the trained deep learning network, the sparse active source seismic record constrains the passive source seismic data prediction network, passive source data reconstruction which shares a measuring line with the active source under the sparse active source condition is realized, and the multi-dimensional deconvolution reconstruction effect and the calculation stability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent seismic exploration, and specifically relates to a method for intelligently optimizing the reconstruction effect of multi-dimensional deconvolution passive source wave field by using a deep learning neural convolution network combined with active source data constraints. Background Art

[0002] Passive source seismic interferometry mainly uses the noise seismic signals observed in the field to obtain the geological characteristics of the medium through which the seismic waves pass, and infer the characteristics of the underground geological structure. Compared with pure active source seismic exploration, it has the advantages of low cost, green environmental protection, and flexible data reconstruction.

[0003] The traditional cross-correlation seismic interferometry method requires that the seismic sources are evenly and densely distributed underground, and it is difficult to have such an ideal illumination state in actual data. The multidimensional deconvolution method proposed later can overcome the influence of non-ideal illumination of underground seismic sources on the reconstructed records, and has achieved good application results in transient source seismic data. The dynamic and kinematic information of the reconstructed records are closer to active source seismic records. However, in noisy passive source seismic data, since wave field separation is difficult to achieve, there are errors in the extraction of point spread functions. Therefore, there are non-physical events in the reconstructed virtual records, which are quite different from active sources. In addition, the computational stability of multidimensional deconvolution is poor and it has a large memory overhead.

[0004] Based on this, in order to improve the computational stability of multidimensional deconvolution and overcome the error of point spread function extraction, it is urgent to develop an improved method to solve the problem of interferometric reconstruction of passive source seismic data with non-ideal illumination noise. The method can use deep learning neural convolutional network combined with active source data constraints to intelligently optimize the multidimensional deconvolution reconstruction effect, so that the reconstruction effect is closer to the active source record, thereby achieving more advantages than traditional cross-correlation and multidimensional deconvolution interferometric methods. Summary of the invention

[0005] The purpose of the present invention is to provide a method for intelligent reconstruction of non-ideal illumination passive source seismic wave fields constrained by sparse active sources, so as to solve the problems of blurred imaging effect under sparse active source conditions and the existence of false event axes and coherent noise under non-ideal distribution of underground passive seismic sources.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] An intelligent reconstruction method for non-ideal illumination passive source seismic wave field with sparse active source constraints, comprising the following steps:

[0008] A. Using the finite difference forward modeling method, evenly spaced detectors are set on the surface to collect noisy passive source seismic records and active source seismic records with uneven source distribution under different velocity models;

[0009] B. Multi-dimensional deconvolution reconstruction of passive source seismic data;

[0010] C. The virtual source seismic records and active source seismic records obtained after multi-dimensional deconvolution reconstruction are intercepted to the number of sampling points suitable for the network input size and normalized:

[0011]

[0012] Among them, x represents a pixel point in the earthquake reconstruction record, x min Represents the value of the minimum pixel in the earthquake reconstruction record where x is located. max Represents the value of the maximum pixel in the earthquake reconstruction record where x is located;

[0013] D. Virtual shot records are used as samples and corresponding active source seismic records are used as labels. An improved U-Net deep learning network is constructed to extract features from label seismic records by adding an attention-guided multi-scale feature fusion mechanism. Reasonable sampling rates of different scales are set to extract multi-scale features from input seismic records. A dual-channel attention mechanism module is then added to U-Net to learn the nonlinear mapping relationship between seismic data.

[0014] E. Based on the trained deep learning network, the test set that did not participate in the training and the sparse active source information are labeled and input into the set network to predict the reconstruction information of the passive source data that is in the same line as the active source.

[0015] Furthermore, step A is specifically as follows: performing data forward simulation, using the finite difference method to solve the wave equation, simulating the underground propagation process of seismic waves, and collecting earthquake records with uneven source distribution and active source earthquake records under different velocity models.

[0016] Furthermore, in step B, the expression for reconstructing passive source seismic records using the multidimensional deconvolution method is:

[0017]

[0018] in,

[0019]

[0020] C(x B ,x A ,t) is the scattered wave field response; G d (x B ,x,t) represents x B The total field or outward propagation field is also called the dipole Green's function, Γ(x,x A,t) represents the underground reference state underground response, also known as the point spread function; S rec is the absorption interface where the detector is located, t is the time series, -t means the time series is reversed; * is the convolution symbol. A With x B are two detectors deployed on the ground, x and x A 、x B They are located on the integral surface S rec Both sides of u in (x A ,σ (m) ,-t) and u out (x B ,σ (m) ,t) are σ (m) The up-going and down-going wave fields excited by the passive source at σ (m) represents the upgoing wavefield generated by all combinations of passive source excitations, and m represents the passive source location.

[0021] Furthermore, step D is specifically as follows: before the U-Net network input, an attention-guided multi-scale feature fusion module ASPP is first added, by sampling the input features at different sampling rates, that is, extracting input features from different scales, and then fusing the acquired features to obtain the final feature extraction result; an attention mechanism CBAM module is added to the convolutional layer of U-Net, and the attention mechanism CBAM module is composed of two sub-modules, CAM and SAM. The two sub-modules can add attention mechanisms in channels and spaces, and the two modules are fused and embedded in U-Net, and then reasonable hyperparameters are set to learn the nonlinear mapping relationship between seismic data.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The present invention performs intelligent optimization of active source data constraints on the multidimensional deconvolution reconstruction effect of non-ideally distributed noise-type passive source data, avoids the error of point spread function extraction and improves the computational stability of multidimensional deconvolution; the deep learning method is used to fully consider the strong nonlinear mapping relationship between seismic data, and has high computational efficiency while improving the accuracy of reconstructed records.

[0024] Specifically, the present invention has the following advantages:

[0025] 1. This method is based on passive source seismic exploration with non-ideal distribution, which effectively suppresses the coherent noise generated in the virtual shot record due to uneven source distribution or seismic interference calculation process, and improves the passive source seismic wave field reconstruction effect under non-ideal lighting conditions;

[0026] 2. This method uses an improved U-Net network to train and learn passive source seismic data reconstructed by multidimensional deconvolution. On the one hand, it has more advantages than multidimensional deconvolution in terms of reconstruction accuracy and computational stability. On the other hand, it overcomes the errors in point spread function extraction and removes non-physical events in virtual records.

[0027] 3. This method uses only a small amount of active source information in the prediction stage to predict the re-record of the complete survey line, which can improve the reconstruction effect, and the reconstructed information has a clear physical meaning. At the same time, it greatly reduces the cost of active source data acquisition;

[0028] 4. The predicted virtual source data obtained by this method can be used in the subsequent inversion imaging of seismic data, and high-quality seismic data inversion imaging results can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 Flow chart of the steps of the intelligent reconstruction method of non-ideal illumination passive source seismic wavefield with sparse active source constraints;

[0031] Figure 2a Forward velocity training model;

[0032] Figure 2b Forward velocity test model;

[0033] Figure 3 Non-ideal distribution of earthquake source locations;

[0034] Figure 4a Virtual source records of non-ideally distributed earthquake sources;

[0035] Figure 4b Active source earthquake records;

[0036] Figure 5 Improve U-Net network;

[0037] Figure 6 The prediction results output by the network;

[0038] Figure 7a Sparse active source data imaging results;

[0039] Figure 7b Traditional multi-dimensional deconvolution reconstructs the recorded imaging results;

[0040] Figure 7c Predict virtual sources and reconstruct recorded imaging results. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with embodiments:

[0042] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0043] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0044] The present invention aims at the problem that in practical applications, noise-type passive sources are often unevenly distributed and wave fields are difficult to separate, which makes it difficult for traditional interference reconstruction methods to achieve good reconstruction results. A non-ideal illumination passive source seismic wave field intelligent reconstruction method with sparse active source constraints is proposed. First, a training model data set is constructed, and an underground non-ideally distributed seismic source is set for the velocity model. Then, a forward simulation is performed, and the passive source seismic data of the forward simulation is subjected to multi-dimensional deconvolution processing; at the same time, a forward simulation is performed on the active source; a training data set is created, and the passive source reconstruction records and active source records are cropped and normalized, and samples and labels are made respectively; then, a neural network model is established: an improved U-Net deep learning network is built, and an attention-guided multi-scale feature fusion mechanism and a self-attention mechanism are added to learn the nonlinear mapping relationship between seismic data; training model: the created training set is input into the neural network, and through the training data set, the parameters are iteratively updated to learn effective signal features and discard false waveform information; finally, the model is tested: using the trained network model, the passive source reconstruction data to be tested and the sparse active source data of the common survey line with the passive source are input into the network for prediction, and the passive source seismic data reconstruction of the entire survey line is completed, and a high-quality sparse active source-constrained passive source seismic wave field intelligent reconstruction shot set is obtained.

[0045] The non-ideal illumination passive source seismic wave field intelligent reconstruction method of the present invention is realized through the MATLAB platform and the Spyder platform, and comprises the following steps:

[0046] 1. Install the MATLAB software platform under Linux system, which requires MATLAB R2016a or above, and is equipped with the Parallel Computing Toolbox. Install the Spyder software platform under Linux system, which requires Python 3.8 or above.

[0047] 2. Use the finite difference forward modeling method to set up evenly spaced detectors on the surface to collect noisy passive source seismic records and active source seismic records with uneven source distribution under different velocity models.

[0048] Specifically, data forward modeling is performed. The finite difference method is used to solve the wave equation to simulate the process of seismic wave propagation underground. Seismic records of uneven source distribution and active source seismic records are collected under different velocity models.

[0049] 3. Perform multidimensional deconvolution reconstruction on passive source seismic data. The expression for reconstructing passive source seismic records using the multidimensional deconvolution method is:

[0050]

[0051] in,

[0052]

[0053] C(x B ,x A ,t) is the scattered wave field response; G d (x B ,x,t) represents x B The total field or outward propagation field is also called the dipole Green's function, Γ(x,x A ,t) represents the underground reference state underground response, also known as the point spread function. rec is the absorption interface where the detector is located, t is the time series, -t means the time series is reversed; * is the convolution symbol. A With x B are two detectors deployed on the ground, x and x A 、x B They are located on the integral surface S rec On both sides. in (x A ,σ (m) ,-t) and u out (x B ,σ (m) ,t) are σ (m) The up-going and down-going wave fields excited by the passive source at σ (m)represents the upgoing wavefield generated by all combinations of passive source excitations, and m represents the passive source location.

[0054] 4. The virtual source seismic records and active source seismic records obtained after multi-dimensional deconvolution reconstruction are intercepted to the number of sampling points suitable for the network input size and normalized:

[0055]

[0056] Among them, x represents a pixel point in the earthquake reconstruction record, x min Represents the value of the minimum pixel in the earthquake reconstruction record where x is located. max Represents the value of the maximum pixel in the earthquake reconstruction record where x is located.

[0057] 5. Virtual gun records are used as samples and the corresponding active source seismic records are used as labels. An improved U-Net deep learning network is constructed, and the features in the labeled seismic records are extracted by adding an attention-guided multi-scale feature fusion mechanism. Reasonable sampling rates of different scales are set, and multi-scale features are extracted from the input seismic records. Then a dual-channel attention mechanism module is added to U-Net to learn the nonlinear mapping relationship between seismic data.

[0058] Specifically, an attention-guided multi-scale feature fusion module ASPP (Atrous Spatial Pyramid Pooling) is added before the input of the traditional U-Net network. The input features are sampled at different sampling rates, that is, the input features are extracted from different scales, and then the acquired features are fused to obtain the final feature extraction result. In addition, an attention mechanism CBAM module is added to the convolutional layer of U-Net. It consists of two modules - CAM (Channel Attention Module) and SAM (Spartial Attention Module). These two sub-modules can add attention mechanisms on the channel and space. The two modules are fused and embedded in U-Net, and then reasonable hyperparameters are set to learn the nonlinear mapping relationship between seismic data.

[0059] 6. According to the deep learning network that has been trained, the test set that has not participated in the training and the sparse active source information are labeled and input into the set network to predict the reconstruction information of the passive source data in the same line as the active source. The waveform information recorded by the virtual gun is clear and accurate, and the coherent noise is well suppressed.

[0060] Example 1

[0061] A flowchart of the steps of an intelligent reconstruction method for non-ideal illumination passive source seismic wavefields constrained by sparse active sources, such as Figure 1shown.

[0062] The real velocity model used is shown in Figure 2, which is a layered velocity model with different thickness (84×128); the network spacing is 10m, the velocity range is between 1500m / s and 4000m / s, and the seismic detectors are set on the model surface. Each grid point is set with a detector, that is, the detector interval is 10m. Set the source distribution position as follows Figure 3 As shown in the figure, the passive source wavelet is set with a random noise sequence, with a frequency between 1 Hz and 30 Hz. The sampling interval is 0.001 s. The real source function used by the active source wavelet is the 15 Hz main frequency Ricker wavelet.

[0063] After multi-dimensional deconvolution of the noise record, the virtual source record is obtained as follows: Figure 4a As shown in the figure, it can be seen that in the virtual shot record with uneven source distribution, the reconstructed virtual source seismic record not only has coherent noise and false phase axis, but also has straight line intersection waveform information that does not conform to the laws of physics. Figure 4b As shown, not only is the waveform continuous without coherent noise interference, but it also presents the reflected wave information of the correct terrain position.

[0064] In this embodiment, the unevenly distributed source virtual shot records under 80% of the velocity model are used as training data, and the corresponding active source source records are used as training labels; the unevenly distributed source virtual shot records under 20% of the velocity model are used as test data, and the sparse active source earthquake records are used as test labels.

[0065] Use the prepared training data to design Figure 5 The improved U-Net network model shown in the figure is trained. After 1200 iterations, the trained network is used to process the test data, and the results are as follows.

[0066] Figure 6 The virtual source seismic record reconstructed by the inhomogeneous source is shown in Figure 1. It is predicted by the neural network training. It can be observed that the coherent noise and false event axes are effectively suppressed, and the waveform is effectively repaired. Compared with the test label, the predicted seismic record can well restore the formation reflection wave information.

[0067] In order to verify the imaging effect of the predicted virtual shot record, the present invention adopts the PSPI prestack depth migration imaging method of phase shift plus interpolation. Under the condition of sparse active sources, the imaging effect is relatively fuzzy, and some reflection layer information is not reflected. Figure 7a As shown. However, the conventional multi-dimensional deconvolution passive source reconstruction is used for imaging. Figure 7bAs shown in the figure, the effect is better on the left side of the model, that is, the passive source dense area. For the underground space without passive sources, although multidimensional deconvolution can overcome part of the impact caused by the non-ideal distribution of seismic sources, there is a significant difference in vertical resolution compared with the left side. The imaging of the layers on the right side of the model is blurred and reliable layers cannot be identified. After deep learning reconstruction of the passive source reconstruction records using the method proposed in the present invention, the imaging effect is as follows: Figure 7c Compared with the sparse active source, the imaging result shows that the deep reflection layer information is clearer, and compared with the pure passive source reconstruction imaging, the right layer information is clearer and more accurate, which proves the accuracy of the training network.

[0068] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. An intelligent reconstruction method for non-ideal illumination passive source seismic wavefields constrained by sparse active sources, characterized in that: The following steps are involved: A. Using the finite difference forward modeling method, evenly spaced detectors are set on the surface to collect noisy passive source seismic records and active source seismic records with uneven source distribution under different velocity models; B. Multi-dimensional deconvolution reconstruction of passive source seismic data; C. The virtual source seismic records and active source seismic records obtained after multi-dimensional deconvolution reconstruction are intercepted to the number of sampling points suitable for the network input size and normalized: Among them, x represents a pixel point in the earthquake reconstruction record, x min Represents the value of the minimum pixel in the earthquake reconstruction record where x is located. max Represents the value of the maximum pixel in the earthquake reconstruction record where x is located; D. Virtual shot records are used as samples and corresponding active source seismic records are used as labels. An improved U-Net deep learning network is constructed to extract features from label seismic records by adding an attention-guided multi-scale feature fusion mechanism. Reasonable sampling rates of different scales are set to extract multi-scale features from input seismic records. A dual-channel attention mechanism module is then added to U-Net to learn the nonlinear mapping relationship between seismic data. E. Based on the trained deep learning network, the test set that did not participate in the training and the sparse active source information are labeled and input into the set network to predict the reconstruction information of the passive source data that is in the same line as the active source.

2. According to claim 1, a sparse active source constrained non-ideal illumination passive source seismic wave field intelligent reconstruction method is characterized in that: Step A is as follows: perform data forward simulation, use the finite difference method to solve the wave equation, simulate the underground propagation process of seismic waves, and collect earthquake records with uneven source distribution and active source earthquake records under different velocity models.

3. The method for intelligent reconstruction of non-ideal illumination passive source seismic wavefields with sparse active source constraints according to claim 1 is characterized in that: Step B, the expression for passive source seismic record reconstruction using multidimensional deconvolution method is: in, C(x B ,x A ,t) is the scattered wave field response; G d (x B ,x,t) represents x B The total field or outward propagation field is also called the dipole Green's function, Γ(x,x A ,t) represents the underground reference state underground response, also known as the point spread function; Srec is the absorption interface where the detector is located, t is the time series, -t means the time series is reversed; * is the convolution symbol. A With x B are two detectors deployed on the ground, x and x A 、x B Located on both sides of the integral surface Srec; u in (x A ,σ (m) ,-t) and u out (x B ,σ (m) ,t) are σ (m) The up-going and down-going wave fields excited by the passive source at σ (m) represents the upgoing wavefield generated by all combinations of passive source excitations, and m represents the passive source location.

4. The method for intelligent reconstruction of non-ideal illumination passive source seismic wavefields with sparse active source constraints according to claim 1, characterized in that: Step D, specifically, is as follows: before the U-Net network input, an attention-guided multi-scale feature fusion module ASPP is first added, by sampling the input features at different sampling rates, that is, extracting input features from different scales, and then fusing the acquired features to obtain the final feature extraction result; An attention mechanism CBAM module is added to the convolutional layer of U-Net. The attention mechanism CBAM module consists of two sub-modules, CAM and SAM. The two sub-modules can add attention mechanisms in channels and spaces. The two modules are fused and embedded in U-Net, and then reasonable hyperparameters are set to learn the nonlinear mapping relationship between seismic data.

Images