Reflection seismic data cross reconstruction method based on UNET + + network

Abstract

The invention discloses a reflection seismic data cross reconstruction method based on a UNET + + network, and aims to solve the problem that seismic data are irregular or under-sampled in space. A traditional seismic data reconstruction method is weak in robustness, strong in parameter selection dependency, low in automation degree and the like. In order to solve the problems of poor stability and lack of specifications of a seismic data reconstruction method based on deep learning, the invention provides the following improvement schemes: an improved neural network model is used to enhance the extraction capability of data features; a selective slicing mechanism constructs a sample set to improve the capacity and richness of the sample set; a model generalization strategy of seismic trace cross reconstruction is used for increasing the reliability of a prediction result. In an actual shot gather test, the method has a good reconstruction effect on seismic records with random missing of 30%, 50% and 80% of gathers, and the reconstruction profile effectively suppresses random noise while recovering a reflected wave event, so that the data quality is improved.

Description

Technical Field

[0001] The present invention belongs to the field of seismic data processing, and relates to artificial neural network technology, 2D seismic data reconstruction technology, and cross-training strategy. The invention results serve for seismic missing trace interpolation, improving the signal-to-noise ratio of seismic data, and optimizing the field acquisition design plan, achieving the effects of reducing the exploration construction cost and improving the quality of seismic exploration data. Background Art

[0002] I. Seismic Data Reconstruction Method

[0003] The missing of seismic traces is often inevitable due to natural condition limitations or human factors, and usually data reconstruction needs to be carried out to make up for it to meet the needs of subsequent seismic data processing and interpretation. The traditional seismic data reconstruction methods are mainly divided into four categories:

[0004] (1) Simple interpolation between seismic traces, estimating the data points between two known seismic traces through statistical methods. The advantages are simple to implement and fast in calculation speed. The disadvantages are that only the information of local adjacent points is considered, and the effect is poor for areas with complex geological structures and uneven data distributions. For example: linear interpolation, that is, simply using the observed values of adjacent seismic traces for linear interpolation, suitable for simple filling in the case of sparse data; bilinear interpolation, that is, performing linear interpolation in the two-dimensional space, considering the changes in adjacent seismic traces and time; triangulation interpolation, that is, triangulating the two-dimensional space formed by seismic traces, and then using the data inside the triangle for interpolation, suitable for irregularly distributed seismic traces.

[0005] (2) Spatial domain interpolation, inferring the seismic data of unknown points by using the spatial distribution of known seismic data points, combining the physical properties of the underground medium and the interpolation algorithm. The advantages are suitable for irregular distribution and high-dimensional data. The disadvantages are high complexity and sensitive to data noise. For example: Kriging interpolation: based on statistical principles, interpolating by fitting the spatial variability between data points, which can adapt to irregular data distributions and noise; radial basis function interpolation (RBF), that is, using radial basis functions to perform weighted summation of data points to estimate the values at unknown positions, suitable for high-dimensional data interpolation; spline interpolation, that is, using spline curves to smoothly fit the data on seismic traces, and then performing interpolation in space, retaining the original features of the data.

[0006] (3) Time-frequency domain interpolation, based on the sparse constraint strategy of seismic signals, restores the complete signal. Its advantages are that it is applicable to data with different sampling rates and resolutions and has strong parameter controllability. Its disadvantages are that it has problems such as false frequency, noise sensitivity and large amount of calculation. For example: wavelet interpolation: using wavelet transform to decompose data into components of different frequencies, and then interpolating, which is suitable for processing irregular data in the time-frequency domain; time domain interpolation, that is, by analyzing the data distribution and change trend of seismic traces in the time domain, interpolation fills in missing data.

[0007] (4) Wave equation interpolation, which relies on prior information of underground velocity to derive the wave field data of the missing traces. The advantage is that it can use the characteristics of seismic waveforms for interpolation. The disadvantage is that it has high requirements for data quality, is greatly affected by noise, and requires more computing resources and time. For example: finite difference method, which is to perform differential calculation and interpolation of seismic data based on wave equations; reflection travel time interpolation, which is to infer the data at the missing location based on the propagation path and velocity information of seismic waves.

[0008] In summary, the application effect of traditional seismic data reconstruction methods is easily affected by data quality, prior conditions and hyperparameter selection, and has disadvantages such as large computational complexity, high storage consumption and strong manual intervention, which limits its application.

[0009] 2. Seismic data reconstruction method based on deep learning

[0010] In recent years, with the successful application of neural networks in various fields, the interpolation method of reflection seismic exploration based on neural networks has gradually become a research hotspot. Compared with traditional methods, its advantages are:

[0011] (1) Nonlinearity: Compared with traditional interpolation methods, neural networks can better adapt to the nonlinear characteristics of data and improve the accuracy and precision of interpolation.

[0012] (2) Adaptability: Neural networks are adaptive and flexible, and can automatically adjust network structure and parameters according to the characteristics of the data, thereby better adapting to seismic data of different types and sizes.

[0013] (3) Parallel processing: The parallel processing capability of neural networks enables them to efficiently process large-scale seismic data and improve the speed and efficiency of interpolation. This is of great significance for real-time or large-scale data processing.

[0014] (4) Automated learning: Neural networks are able to learn feature representations from data without the need to manually design features. This means that it can automatically extract the most informative features from seismic data for better interpolation.

[0015] Strong generalization ability: With proper training, neural networks can have strong generalization ability and can accurately interpolate and predict earthquake data that has never been seen. This makes the neural network interpolation method very versatile and scalable.

[0016] Although this type of deep learning-based seismic data reconstruction method has great potential, it still has some shortcomings in practical applications: first, traditional convolutional neural networks are insufficient in their ability to extract seismic data features; second, there is a lack of effective methods for constructing training data sets; third, in a single reconstruction of seismic data, the performance of convolutional neural networks is not stable enough; finally, there is currently a lack of a complete modeling, optimization, and evaluation process for the reconstruction of actual seismic data. Summary of the invention

[0017] In view of the problems of weak robustness, strong dependence on parameter selection, low degree of automation of traditional seismic data reconstruction methods, and weak stability and lack of standardization of seismic data reconstruction methods based on deep learning, the present invention intends to develop a reflection seismic data cross-reconstruction method and standard process based on the improved UNET network. The key technical breakthroughs include: improved neural network model to enhance the ability to extract data features; selective slicing mechanism to construct sample sets to improve their capacity and richness; model generalization strategy for cross-reconstruction of seismic traces to increase the reliability of prediction results. On this basis, a corresponding standard reconstruction process for actual data is proposed to obtain reflection seismic data with strong integrity and high signal-to-noise ratio.

[0018] Technical solution:

[0019] A cross-reconstruction method for reflection seismic data based on UNET++ network, the method comprising the following steps:

[0020] S1. Establish the initial UNET++ network architecture;

[0021] S2, using data samples, and using a selective random slicing mechanism to create a training set;

[0022] S3, use the training set to train the initial UNET++ network until the iteration number and error value are met and then the trained UNET++ network model is output;

[0023] S4. Regularly slice the missing data to be reconstructed, input each slice into the UNET++ network model to obtain reconstruction results of different depths, select the slice with the highest signal-to-noise ratio as the current reconstruction result, and use cross reconstruction to output reconstructed slices with high signal-to-noise ratio. Finally, splice the reconstructed slices into a complete single-shot record.

[0024] Preferably, in S1, the initial UNET++ network architecture adds a nested structure based on the original UNET architecture, and fuses low-level and high-level features by introducing feature fusion modules of multiple scales, thereby improving the feature representation capability and transmission efficiency of the traditional UNET structure.

[0025] Preferably, the initial UNET++ network architecture comprises 15 convolutional modules, each of which comprises two convolutional layers and two exponential linear units (ELUs), wherein the filter size of the convolutional layer is 3×3;

[0026] In the deepest convolutional block CB40 in the UNET++ network architecture, a dropout operation is added between the first ELU and the second convolutional layer;

[0027] The UNET++ network architecture uses a cross-connection operation, where the feature information in the downsampling process is directly connected to the corresponding upsampling feature information;

[0028] The decoders of the UNET++ network architecture are also connected to each other at the same depth;

[0029] During the downsampling process, the number of filters used in different network depths is 32, 64, 128, 256, and 512;

[0030] During upsampling, the number of filters used in different network depths is 512, 256, 128, 64, and 32.

[0031] Preferably, in S2, the selective random slicing mechanism is specifically as follows:

[0032] ① Select the training data window: set the hyperbolic window, that is, Where x is the offset, t 0 is the self-excited two-way travel time, v is the roughly estimated equivalent layer velocity, and t represents the time period of the training data window on each seismic data track; therefore, two sets of parameters need to be given to determine the window, namely

[0033] ② Random cutting: Within the window range determined in step ①, randomly select sampling points according to a certain ratio as the center points of the training slices; the deepest network performs a 16-fold downsampling operation relative to the original data, and the data input to the network must be an integer multiple of 16×16; a boundary judgment mechanism is added to the random sampling process to ensure that the data range of the slice does not exceed the range of the original single-shot data;

[0034] ③ Set the missing percentage probability and randomly extract the traces from the slice; after randomly determining the missing ratio, randomly extract the seismic traces in the slice and assign their amplitudes to 0.

[0035] Preferably, in S3, the overall loss function of the UNET++ network model is as follows:

[0036]

[0037] Where n represents the depth of the UNET++ network, ‖·‖ 1 Indicates L 1 The calculated loss value L(δ) is used to update the network parameters in the back-propagation process.

[0038] Preferably, in S4, the specific process of the cross-reconstruction is as follows:

[0039] ①Mark the missing tracks in the input slices;

[0040] ② Input the slice data into the trained UNET++ network model for reconstruction;

[0041] ③ Set the number of cross-reconstruction times n, and randomly divide the traces other than the marked missing traces in the input slice into n groups;

[0042] ④ Find the position of the missing track of the i-th group in step ③ in the existing reconstruction results and assign it to 0, i≤n; input it into the UNET++ network again for reconstruction, and repeat this operation until i=n.

[0043] Specifically, in ① of S4, the track number of the missing track is obtained by calculating the root mean square amplitude of each track or reading the track header information of the original single shot.

[0044] Specifically, in S4, when the slice grid exceeds the original data range, the edge is expanded by filling zeros.

[0045] Beneficial effects of the present invention

[0046] The present invention proposes a cross-reconstruction method for reflection seismic data based on the UNET++ network. When more than 30% of the gather data is missing, the layout of field detection points is severely limited, and the random noise interference is strong, the method can effectively recover most of the data and has a good suppression effect on random noise, thereby effectively improving the overall quality of the exploration data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of forward propagation of UNET++ network.

[0048] Figure 2 Schematic diagram of the cross-reconstruction strategy.

[0049] Figure 3 This is the flow chart of cross-reconstruction of reflection seismic data based on UNET++ network.

[0050] Figure 4 Reconstruct the renderings for actual data.

[0051] Figure 5 This is the reconstruction effect diagram of the measured data by cross-reconstruction of reflection seismic data based on the UNET++ network. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto:

[0053] The present invention proposes a theoretical algorithm for cross-reconstruction of reflection seismic data based on the UNET++ network, including three core contents: establishing an initial UNET++ network architecture, constructing a selective slicing training set, and a seismic trace cross-reconstruction strategy. On this basis, a standard process for actual data reconstruction is developed, which is as follows:

[0054] (1) Establishing the initial UNET++ network architecture

[0055] The present invention advocates the use of an improved UNET structure (i.e., UNET++) as an artificial neural network model for seismic data reconstruction. UNET++ adds a nested structure based on the original UNET architecture and introduces feature fusion modules of multiple scales to fuse low-level and high-level features, thereby improving the feature representation capability and transmission efficiency of the traditional UNET structure.

[0056] like Figure 1 As shown in (a), the network architecture used in the present invention includes 15 convolutional modules (CBs), and the structure of each CB is ( Figure 1 (b)) contains two convolutional layers with a filter size of 3×3 and two exponential linear units (ELUs):

[0057]

[0058] Where x is the input and the hyperparameters in ELU control the output. For the deepest convolutional block (CB40) in the UNET++ network, a dropout operation is added between the first ELU and the second convolutional layer to reduce overfitting. In the UNET++ network, the downsampling process is implemented by a maximum pooling layer with a span of 2×2, while the upsampling process is implemented by a deconvolution layer with a span of 2×2 and a convolution kernel of 3×3. The stride connection operation is used in U-Net++. The feature information in the downsampling process is directly connected to the corresponding upsampled feature information. Unlike the traditional U-Net, the decoder of U-Net++ is also interconnected at the same depth. In the downsampling process, the number of filters used in different network depths is 32, 64, 128, 256 and 512, while in the right upsampling path, the number of filters is 512, 256, 128, 64 and 32.

[0059] UNET++ networks can be trained using deep supervision strategies. All U-Net++ with different depths can be trained through CB after the entire network is trained. 01 ,CB 02 ,CB 03 and CB 04 Output data and select the optimal depth of the network by calculating the loss values ​​at different depths, thereby reducing the phenomenon of gradient disappearance or gradient explosion. In seismic data reconstruction, we mark the observation data to be reconstructed as d obs , and use it as the input of the UNET++ network. The reconstructed data is recorded as d. In the forward propagation of the UNET++ network, it is necessary to train Figure 1 (a) The whole network parameter u δ , where δ is the parameter of the UNET++ network. The network outputs of the four different depths are obtained through a convolutional layer with a filter size of 1×1. The network parameter u δ It is obtained by calculating the loss function L(δ). Since the output of each layer is accompanied by a loss function, the overall loss function of the method of the present invention is the sum of the loss functions of the four layers, which is expressed as:

[0060]

[0061] Where n represents the depth of the UNET++ network, ‖·‖ 1 Indicates L 1 The calculated loss value is used to update the network parameters during the back-propagation process.

[0062] (2) Constructing a selective slicing training set

[0063] The amount and quality of training set data greatly affect the effect of deep learning. The present invention proposes a method for randomly extracting missing traces and cutting pre-stack single shot data to construct a training set. By establishing a selective slicing training set construction mechanism, the trained network model can better identify the image features of pre-stack reflection waves. Therefore, in the data reconstruction process, it is easier to restore the reflection wave features of the missing traces, and it has a suppressive effect on the original noise. Taking into account the characteristics of seismic data, including: seismic data is usually missing traces, single shot data also contains spatiotemporal attributes related to seismic wave propagation, and the data quality in different regions is different, the present invention adopts a selective slicing training set construction mechanism, and the specific method is as follows:

[0064] ① Select the training data window: set the hyperbolic window, that is, Where x is the offset, t 0 is the self-excited two-way travel time, and v is the roughly estimated equivalent layer velocity. Therefore, determining this window requires two sets of parameters, namely The parameter selection is preferably such that the area between the two hyperbolas contains as much of the reflection data as possible with higher quality;

[0065] ② Random cutting: within the window range determined in step ①, randomly select sampling points in a certain ratio as the center points of training slices (patches). UNET and its improved network have strict requirements on the size of slices. For the network used in the present invention, the deepest network performs a 16-fold downsampling operation relative to the original data, so the data input to the network must be an integer multiple of 16×16. At the same time, a boundary judgment mechanism is added to the random sampling process to ensure that the data range of the slice does not exceed the range of the original single-shot data; compared with cutting according to a fixed step size, this random cutting method not only effectively increases the amount of training samples but also avoids the additional false edge features introduced by data expansion;

[0066] ③ Set the missing percentage probability and randomly extract the traces from the slice: For example, in the present invention, the probability of missing 10% to 80% traces is set at intervals of 10 percentage points in the training set, and the probability of each percentage missing situation is 0.125. After randomly determining the missing ratio, randomly extract the seismic traces in the slice and assign their amplitude to 0.

[0067] (3) Seismic trace cross reconstruction strategy

[0068] In the subsequent model generalization stage, the present invention proposes a cross-reconstruction strategy, which is to randomly cut off the originally non-missing parts of the slice after the first reconstruction, and then input them "in small quantities and multiple times" into the UNET++ model for secondary reconstruction, so as to improve the local feature recovery ability of the seismic profile and reduce the uncertainty of the single generalization result. Figure 2 The cross-reconstruction strategy proposed by the present invention has the following basic process:

[0069] ① Mark the missing traces in the input slice. The trace number of the missing trace can generally be determined by calculating the root mean square amplitude of each trace or reading the trace header information of the original single shot;

[0070] ② Input the slice data into the trained UNET++ network model for reconstruction;

[0071] ③ Set the number of cross-reconstruction times n, and randomly divide the traces other than the marked missing traces in the input slice into n groups;

[0072] ④ Find the position of the missing track of the i-th (i≤n) group in step ③ in the existing reconstruction results and assign it to 0, input it into the UNET++ network again for reconstruction, and repeat this operation until i=n.

[0073] (4) Standard process for data reconstruction:

[0074] like Figure 3 As shown, the application of the cross-reconstruction method of reflection seismic data based on the UNET++ network advocated by the present invention is mainly divided into two processes, namely the training process and the generalization process. In the training process, a selective random slicing mechanism is used to prepare a training set, and its data is added to the UNET++ neural network of the attention mechanism, and the sum of the four-layer loss function is used as the overall loss function to perform back-propagation iteration until the iteration conditions such as the iteration number and the error value are met, and then the model is output; in the reconstruction process, the missing data to be reconstructed is regularly sliced, and when the slice grid exceeds the part of the original data range, the edge is expanded by zero padding, and each slice is input into the trained UNET++ network to obtain the reconstruction results of four depths, and the slice with the highest signal-to-noise ratio is selected as the current reconstruction result, and cross-reconstruction is used to output the reconstructed slice with a high signal-to-noise ratio, and finally the reconstructed slices are spliced ​​into a complete single-shot record.

[0075] In order to test the application effect of the method of the present invention on actual reflection seismic data, a test was conducted on a certain measured data of the Subei Basin. Figure 4 As shown. The reconstruction effects of missing 30%, 50% and 80% are tested respectively, as shown Figure 5 As shown in the figure, 50% and 80% of the data were randomly removed, and a more obvious empty channel area appeared. Figure 5 (c) is a structure reconstructed based on the method of the present invention. It can be seen that the reconstruction results not only effectively supplement the blank channel, but also suppress a large amount of noise, highlight the deep reflection signal, and greatly enhance the quality of the data. Compared with the single reconstruction and cross reconstruction methods, the latter greatly enhances the lateral continuity of the effective reflection event axis.

[0076] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A cross-reconstruction method for reflection seismic data based on UNET++ network, characterized in that The method comprises the following steps: S1. Establish the initial UNET++ network architecture; S2, using data samples, and using a selective random slicing mechanism to create a training set; S3, use the training set to train the initial UNET++ network until the iteration number and error value are met and then the trained UNET++ network model is output; S4. Regularly slice the missing data to be reconstructed, input each slice into the UNET++ network model to obtain reconstruction results of different depths, select the slice with the highest signal-to-noise ratio as the current reconstruction result, and use cross reconstruction to output reconstructed slices with high signal-to-noise ratio. Finally, splice the reconstructed slices into a complete single-shot record.

2. The method according to claim 1, characterized in that In S1, the initial UNET++ network architecture adds a nested structure based on the original UNET architecture, and fuses low-level and high-level features by introducing feature fusion modules of multiple scales, thereby improving the feature representation capability and transmission efficiency of the traditional UNET structure.

3. The method according to claim 2, characterized in that The initial UNET++ network architecture includes 15 convolutional modules, each of which includes two convolutional layers and two exponential linear units (ELUs), where the filter size of the convolutional layer is 3×3; In the deepest convolutional block CB40 in the UNET++ network architecture, a dropout operation is added between the first ELU and the second convolutional layer; The UNET++ network architecture uses a cross-connection operation, where the feature information in the downsampling process is directly connected to the corresponding upsampling feature information; The decoders of the UNET++ network architecture are also connected to each other at the same depth; During the downsampling process, the number of filters used in different network depths is 32, 64, 128, 256, and 512; During upsampling, the number of filters used in different network depths is 512, 256, 128, 64, and 32.

4. The method according to claim 1, characterized in that In S2, the selective random slicing mechanism is as follows: ① Select the training data window: set the hyperbolic window, that is, Where x is the offset, t0 is the self-excited two-way travel time, v is the estimated equivalent layer velocity, and t represents the time period of the training data window on each seismic data track; therefore, two sets of parameters need to be given to determine the window, namely ② Random cutting: Within the window range determined in step ①, randomly select sampling points according to a certain ratio as the center points of the training slices; the deepest network performs a 16-fold downsampling operation relative to the original data, and the data input to the network must be an integer multiple of 16×16; a boundary judgment mechanism is added to the random sampling process to ensure that the data range of the slice does not exceed the range of the original single-shot data; ③ Set the missing percentage probability and randomly select the slices; After randomly determining the missing ratio, the seismic traces in the slice are randomly sampled and their amplitudes are assigned to 0.

5. The method according to claim 1, characterized in that In S3, the overall loss function of the UNET++ network model is as follows: Where n represents the depth of the UNET++ network, ‖·‖1 represents the L1 norm, and the calculated loss value L(δ) is used to update the network parameters in the back-propagation process.

6. The method according to claim 1, characterized in that In S4, the specific process of cross-reconstruction is as follows: ①Mark the missing tracks in the input slices; ② Input the slice data into the trained UNET++ network model for reconstruction; ③ Set the number of cross-reconstruction times n, and randomly divide the traces other than the marked missing traces in the input slice into n groups; ④ Find the position of the missing track of the i-th group in step ③ in the existing reconstruction results and assign it to 0, i≤n; input it into the UNET++ network again for reconstruction, and repeat this operation until i=n.

7. The method according to claim 6, characterized in that In ① of S4, the track number of the missing track is obtained by calculating the root mean square amplitude of each track or reading the track header information of the original single shot.

8. The method according to claim 1, characterized in that In S4, when the slice grid exceeds the original data range, zero padding is used to expand the edge.

Images