An Adaptive Blocking and Lossless Cutting Method for Seismic Data

Abstract

The present invention discloses a method for generating deep learning sample labels for new seismic data and a method for adaptively block-dividing and losslessly stitching seismic data. It belongs to the field of preprocessing and postprocessing of seismic signal data. Based on composite filtering, the present invention extracts the noise of actual seismic data, combines it with simulated seismic data to realize the construction of a sample set, and based on the adaptive block-dividing and dynamic weighted fusion algorithm, realizes the random block-dividing of seismic data and the lossless synthesis of data blocks into complete seismic data, which is a complete preprocessing and postprocessing process for seismic data training. The randomly block-divided training set and sample set generated by this method are more suitable for deep learning training. The training sample set contains actual noise, thereby improving the accuracy of the model prediction results, and further improving the quality of seismic data; the lossless stitching can eliminate the overall distortion and overfitting phenomena caused by the synthesis of data blocks, providing a new idea for seismic data noise suppression based on deep learning.

Description

Technical Field

[0001] The present invention relates to the field of signal processing, and particularly to a method for preprocessing and postprocessing seismic data for deep learning algorithms. Background Art

[0002] In the field of geophysical exploration, techniques such as full waveform inversion and prestack migration imaging are of great significance for reservoir distribution and formation lithology structure. The accuracy of these techniques depends on the accurate extraction of prestack seismic wavelets. The Tarim region in the southwest of China is rich in oil and gas resources, but the oil and gas resources are stored in deep and ultra-deep underground layers. The seismic signals received through seismic exploration are very weak, seriously affecting the accuracy of formation imaging and inversion.

[0003] To improve seismic data noise suppression, more and more research has started to adopt deep learning algorithms, but there are significant limitations in the preprocessing of deep learning data. Firstly, deep learning requires pure label data and noisy sample data. Ideal pure data without noise cannot be obtained in nature, so it is difficult to generate the training labels used. Secondly, the scale of seismic data is huge and cannot be directly input into the neural network model. Even if the seismic data is cut and divided into blocks, there are obvious edge effects during synthesis, resulting in a large number of synthetic traces at the edges of the blocks in the results, affecting the denoising effect of deep learning. Summary of the Invention

[0004] To solve the problems existing in the prior art, the present invention provides a new method for generating deep learning sample labels for seismic data and an adaptive block division and lossless fitting method for seismic data. By randomly and dynamically weighting and fusing the simulated and synthesized pure seismic data and the noise adaptively extracted from real seismic data, training data with real noise and corresponding pure samples without noise can be obtained, enabling the neural network model to better learn noise suppression. At the same time, through adaptive gradual weighting and fusion, the data blocks of deep learning can be synthesized without loss, realizing enhanced denoising and lossless synthesis of seismic data.

[0005] The specific technical solution adopted by the present invention is as follows:

[0006] Firstly, real seismic data is obtained, and low-amplitude noise containing a very small amount of effective signals is obtained through compound filtering, and local data enhancement is performed based on the real seismic data to obtain the real noise distribution existing in nature in the seismic data.

[0007] The simulated and generated pure seismic data and its distribution are obtained. According to the distribution, the low-amplitude noise and the pure seismic data are dynamically and adaptively superimposed. The superimposition process adopts block division and Gaussian random processing to achieve small-scale noise samples covering all large-scale pure samples, obtaining training samples containing only real noise and pure label data without noise.

[0008] When cutting blocks, set the data block size and stride input into the neural network model, cut the two-dimensional seismic data into blocks, and perform multiple dimension exchanges. Finally, convert the two-dimensional seismic data into three-dimensional training data, and then input it into the neural network model for training.

[0009] When synthesizing blocks, automatically generate the required size and stride for synthesizing blocks according to the parameters during block cutting. Gradually weight the overlapping parts of each block dynamically according to the opposite-side distance, and weight and fuse the contact edges of the two blocks with a customized weight value in total, so that the edge data of the overlapping blocks can be smoothly transitioned, eliminating the generated distorted data and solving the problem that the geological structures between data blocks cannot be connected.

[0010] Compared with the existing methods, the method for synthesizing training sample label data of the present invention enables the neural network model to learn more real noise suppression rules, avoids suppressing effective signals, and improves the signal-to-noise ratio of the results; the seismic data block division and fusion method proposed by the present invention enables the data blocks to generate seismic data files in the original format without loss, avoiding edge highlighting and overfitting phenomena, and making the data more valuable for practical applications; the fitting method can unify the input and output of the neural network, facilitating the gradient optimization and training of the network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 is the flowchart of the method of the present invention;

[0012] Figure 2 is the flowchart of the method for generating deep learning sample labels for new seismic data;

[0013] Figure 3 is the schematic diagram of the seismic data adaptive block division and lossless fitting method;

[0014] Figure 4 is the sectional gray scale diagram synthesized by the conventional method of the seismic data of the present invention in the embodiment, and the discontinuous phenomenon caused by splicing can be clearly seen.

[0015] Figure 5 is the sectional gray scale diagram synthesized by the seismic data synthesis method of the present invention in the embodiment, and the seamless splicing effect can be seen.

[0016] Figure 6 is the synthetic comparison spectrogram of the seismic data of the present invention in the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0017] The following further describes the specific implementation manners of the present invention in conjunction with specific embodiments:

[0018] The present invention discloses a method for generating deep learning sample labels for new seismic data and a method for adaptively dividing and losslessly fitting seismic data, belonging to the field of seismic data processing. Through the adaptive synthesis of real data and simulated pure data, the present invention obtains training label sample data close to the real situation. At the same time, the present invention divides and cuts seismic data, which can be directly imported into a neural network model for training and prediction. Through the block synthesis method of the present invention, the model result can obtain the complete seismic data result without loss. This method is also applicable to traditional algorithms for block processing.

[0019] The following further specifically describes the present invention by introducing the specific implementation steps of the present invention. In combination with the embodiments and the drawings, the content, features, advantages, and effects of the present invention will become clearer according to the following description.

[0020] Figure 1 The following is the flowchart of the method of the present invention, and the specific implementation steps are as follows:

[0021] Step 101: Construct deep learning training sample data based on real data and theoretical simulation data.

[0022] Establish a training set containing real noise and a corresponding sample set without pure noise. The overall process of this step is as Figure 2 shown. Seismic data can be regarded as composed of effective signals and noise:

[0023] f(n) = s(n) + v(n) (1)

[0024] Where d(n) represents a single seismic data value, s(n) represents the effective signal contained in a single seismic data value, and v(n) represents the real noise contained in a single seismic data value.

[0025] Real noise is different from random noise and is the actual noise interference received during actual geoscience seismic data acquisition. This method first obtains real seismic data and simulated synthetic seismic data, extracts the noise samples from the real seismic data through composite filtering, and obtains the actual noise distribution existing during seismic data acquisition; subsequently, the noise samples and the simulated synthetic seismic data are weighted and dynamically superimposed and fused. The fusion process adopts Gaussian random blockization, and the process can be expressed by formula 2:

[0026]

[0027] Where, F compose is the matrix block of the synthetic data, F simu is the matrix block of the simulated data, V is the matrix block of the extracted noise, and the representation form of V is as follows:

[0028]

[0029] This method can globally cover the original data with noise while preserving the distribution of the original data, enabling the deep learning network model to learn real noise suppression and weak signal retention, and improving the noise suppression effect and overall signal-to-noise ratio of seismic data.

[0030] Step 102: Generating the input set of the neural network model using the scalable block and lossless fitting method for seismic data

[0031] The seismic data is segmented to be suitable for the input and output of the deep learning network model. Seismic data is a two-dimensional double-precision array, usually with a length much greater than the width, and relatively large length and width values, such as 123600×2000; while deep learning network models usually require data matrices with the same length and width, and relatively small length and width, such as 256×256. The gap between the two is too large to be directly applicable. This method divides the seismic data into scalable data blocks suitable for deep learning models.

[0032] The overall process of this step is as Figure 3 shown. First, the two dimensions of the two-dimensional seismic data are cut according to size and step, and multiple one-dimensional matrices of the same size are generated for each dimension. The formula is as follows:

[0033]

[0034] Subsequently, through composite dimension fusion and dimension transformation, a data block set that conforms to the deep learning network model is obtained. The formula is as follows:

[0035]

[0036] In the above transformation, variable overlapping regions and random shuffling are set to ensure the comprehensiveness of learning and the randomness of samples, and to avoid overfitting in network training.

[0037] Step 103: Training / Prediction of the deep learning model

[0038] Using the data from step 102 as the input for deep learning, training can be carried out. In theory, it is applicable to any neural network model. At this time, steps 101, 102, and 103 have formed a training closed-loop, and step 104 is not required; if prediction is to be carried out, step 104 needs to be additionally adopted, and the input and output scales of the neural network model should be consistent, such as AutoEncoder, Generative Adversarial Network (GAN) and related variants such as Wasserstein GAN (WGAN), etc. This method is also applicable to traditional algorithms that process data blocks separately, such as Support Vector Machine (SVM).

[0039] Step 104: Synthesizing lossless seismic data using the scalable block and smooth fitting method for seismic data

[0040] The output of the network model is a set of data blocks. If directly pieced together, obvious block edges will be generated in the synthesized seismic data, resulting in serious data distortion. The geological structures between data blocks cannot be connected, and it has no practical application value. For example, Figure 4 the left seismic data map of

[0041] The reason for the above problem is that the data blocks after cutting are independent of each other (if randomly shuffled, the correlation loss is reduced to 0). The neural network model predicts data without temporal correlation before and after, so that the output data blocks are also independent of each other, and the edges are not coherent. To avoid the above problems, this method adopts a scalable overlapping cutting and dynamic gradient weighting algorithm. The overall process is as shown in Figure 3 shown.

[0042] First, preset the step size step of the block in the cutting and partitioning, so that there is overlapping data between each cut block. Thus, the neural network will generate multiple sets of prediction results for the same region during prediction, providing a prerequisite for the smooth fitting of multiple sets of data.

[0043] During the process of data block splicing, dynamic weighting is used for the merging of the overlapping regions of the blocks. The algorithm detects the size and step attributes of data block 1 and data block 2, and generates the weight matrix W 1 (n) of data block 1 according to the overlapping size, where the weight matrix value w 1 (n) ∈ [0, 1], as shown in the following formula:

[0044]

[0045] where m and n are the same as the size of the overlapping region of the two data blocks, and the value of w 1 (n) changes with the center distance of the blocks. The weight matrix is the smallest at the farthest distance, so that the closer the data of the data block is to the edge, the lower the influence on the fitting effect; the weight matrix W 2 (n) of data block 2, where the weight matrix value W 2 (n) ∈ [0, 1], and the relationship with w 1 (n) is as shown in the following formula:

[0046]

[0047] The process of fusing two data blocks is the sum of the products of each and the weight matrix, which can be expressed by formula 2:

[0048] O 1，2 (n) = I 1 (n) × W 1 (n) + I 2 (n) × W 2 (n) (8)

[0049] The process of fusing data blocks into the entire seismic data is composed of several sub-processes mentioned above. After multiple overlaps and dynamic weighting in the same area, the influence of overfitting data at the edge of a single block can be eliminated, which is equivalent to combining the prediction results of multiple neural network models into one, making the entire seismic data natural and real. The comparison of the cross-sectional diagram effects is as Figure 4 and Figure 5 shown. The comparison of the spectrogram is as Figure 6 shown, indicating that the effect of this method is significant.

[0050] From the above process, it can be seen that the present invention provides a new method for generating deep learning sample labels for seismic data and a method for lossless fitting of seismic data adaptive block segmentation. By constructing a training set containing real noise and simulated generated data without any noise, the neural network model is allowed to learn the suppression of real noise and improve the signal-to-noise ratio of the prediction results. By using the method of data block cutting with overlapping regions and dynamic weighting for the merging of block overlapping regions, while making the seismic data applicable to deep learning, the prediction results can be losslessly synthesized into seismic data, which has more practical engineering application value and realizes a complete set of seismic data training and prediction data preprocessing processes.

Claims

1. An adaptive block-based lossless splicing method for seismic data, characterized in that: Noise close to the real situation is obtained by extracting noise from real seismic data through a composite filtering algorithm. Subsequently, the noise samples are weighted and dynamically superposed and fused with simulated synthetic seismic data. The Gaussian random block-based method is adopted in the fusion process to obtain pure labels and sample data containing the real noise distribution, so that the network model can learn to effectively suppress seismic noise. By adaptively cutting the two-dimensional seismic data in two dimensions and dynamically gradually changing the weights at the edges, while deep learning the seismic data, adaptive cutting and lossless fusion are achieved, which is more suitable for training deep learning models for seismic data. The main steps of the method are as follows: First, a wide range of noise extraction is performed on the selected seismic data through a composite filtering algorithm to obtain low-amplitude real noise samples. Then, according to the distribution of the simulated seismic data, the low-amplitude real noise is weighted and dynamically superposed and fused to obtain sample data containing the real noise distribution, and the simulated seismic data is the pure label. Then, the generated two-dimensional data is adaptively cut according to a custom scale, and the single two-dimensional matrix seismic data is converted into multiple groups of two-dimensional seismic data blocks. Each block directly contains a certain amount of repeated data to ensure the comprehensiveness of model learning, and the blocks are randomly shuffled to prevent the model from overfitting. If a prediction result is required, the multiple groups of two-dimensional seismic data blocks obtained by the model are merged in pairs. When merging, the edge dynamic gradual change weighting is adopted to ensure the data connection and smooth transition between the data blocks, so as to generate a prediction result with less loss. The adaptive cutting and edge dynamic gradual change weighting are specifically as follows: First, a step size step of the block is preset in the cutting block, so that there is overlapping data between each cut block, so that the neural network will generate multiple groups of prediction results in the same area during prediction, providing a prerequisite for the smooth fitting of multiple groups of data. During the data block splicing process, the merging of the overlapping regions of the blocks adopts dynamic weighting processing. Detect the size and step attributes of data block 1 and data block 2, and generate the weight matrix W 1 (n) of data block 1 according to the overlapping size, where the weight matrix value w 1 (n) ∈ [0, 1], as shown in the following formula: where m, n are the same as the size of the overlapping region of the two data blocks, and w 1 (n) varies with the center distance of the blocks, and the weight matrix is the smallest at the farthest distance, so that the closer the data of the data block is to the edge, the lower the influence on the fitting effect; the weight matrix W of data block 2 2 (n), where the weight matrix value w 2 (n) ∈ [0, 1], and the relationship with w 1 (n) is as follows: The fusion process of the two data blocks is the sum of the products of each and the weight matrix, as shown in the following formula: O 1,2 (n) = I 1 (n) × W 1 (n) + I 2 (n) × W 2 (n).

2. The adaptive block-based lossless splicing method for seismic data according to claim 1, characterized in that: The real seismic data required to generate the sample label should be the seismic data actually measured in the project, not the theoretically synthesized data.

3. The adaptive block-based lossless splicing method for seismic data according to claim 1, characterized in that: The generation and synthesis method of seismic data blocks need to be consistent with the input and output specifications of the deep learning model.

Images