Earthquake denoising method based on improved generative adversarial network

Abstract

The invention provides an earthquake denoising method based on an improved generative adversarial network. The denoising method comprises the following steps: improving a kernel function in the generative adversarial network; training the adversarial network to obtain a trained adversarial network; and performing network denoising processing on the training adversarial network. The signal-to-noise ratio of the seismic signals is improved, and the influence of manually selected characteristic functions or threshold parameters is avoided.

Description

Technical Field

[0001] The present invention relates to the field of oil and gas exploration seismic data processing, and in particular to a seismic denoising method based on an improved generative adversarial network. Background Art

[0002] Deep learning is a new branch of machine learning. Its emergence makes machine learning closer to its original design goal - artificial intelligence. Deep learning has a structure with multiple processing layers. The features learned in the low-level processing layers are combined into more abstract features by the high-level processing layers. Therefore, deep learning can learn more and more abstract features of data layer by layer. Generally speaking, the purpose of studying deep learning is to enable computers to simulate the mechanism of the human brain to master the ability to learn data and then complete various tasks.

[0003] At present, deep learning has been gradually applied to the field of seismic data processing and interpretation, such as reservoir parameter prediction, fault detection, first arrival wave picking, seismic phase identification, etc. Reservoir prediction based on deep learning technology has achieved good results. Duan Youxiang et al. borrowed the advantages of convolutional neural networks in image processing to reorganize reservoir parameter data into pixel format, and designed single-layer and double-layer convolutional neural network models for porosity prediction. The results show that the double-layer convolutional neural network can extract more features of the data, and thus has better prediction results. Song Hui et al. combined the gated recurrent unit network with the convolutional neural network and proposed a new porosity prediction method. This method can well extract the spatiotemporal characteristics of reservoir parameters. Experimental results show that compared with a single gated recurrent unit network and convolutional neural network, the network model combined with the two has better porosity prediction results. Wu has so far had many methods for suppressing random noise in seismic exploration, distributed in different fields, and has achieved very obvious results. The earliest method can be traced back to 1981, when Larner restored incomplete seismic traces, designed detailed methods in the process of field seismic data acquisition, and made corresponding discussions and research.

[0004] Filter method. The prediction error filter method is often used in the reconstruction of seismic data after acquisition. The characteristic of this filter method is to reconstruct the acquired signal and perform convolution processing with a certain interpolation filter and the seismic data to be restored. However, this method is prone to large errors because it processes non-uniformly sampled data as regular data and then interpolates through a Gaussian window. The uncertainty of the interpolation result is large.

[0005] Operator recovery and reconstruction method. This algorithm uses forward and inversion operators to iteratively solve an inverse problem, which includes the DMO (common offset) operator, the inverse DMO stacking method and the AMO method (wave field operator method). The wave field operator method uses the Kirchhoff operator and suppresses the false frequency part through quadratic migration to complete the data reconstruction. This method can deal with the non-uniformity of the sampled data, but the premise is to have some understanding of the information of the formation structure. Its calculation amount is very large and the sampling rate requirement is also very high. Therefore, this algorithm is not very ideal for data reconstruction.

[0006] There is still a lot of room for development in the application of deep learning technology in the field of random noise attenuation in seismic data. Summary of the invention

[0007] In view of the above problems, the present invention is proposed to provide a seismic denoising method based on an improved generative adversarial network that overcomes the above problems or at least partially solves the above problems.

[0008] According to one aspect of the present invention, a seismic denoising method based on an improved generative adversarial network is provided, the denoising method comprising:

[0009] Improve the core functions in generative adversarial networks;

[0010] Training the adversarial network to obtain a trained adversarial network;

[0011] The trained adversarial network is subjected to network denoising processing.

[0012] Optionally, the improved generative adversarial network method specifically includes: using a U-net structure that has a good effect in the field of image segmentation as a discriminant model of network constraints to extract noise feature information.

[0013] Optionally, the improved generative adversarial network adopts a classic residual convolutional network structure as a generation model.

[0014] Optionally, the mean square error loss function of the improved generative adversarial network is expressed as:

[0015]

[0016] Among them, N represents the total number of samples, x represents the sample data distribution, G(z) represents the data distribution of the generator, and ||*|| represents the Frobenius norm.

[0017] Optionally, the final loss function of the improved generative adversarial network is:

[0018]

[0019] Optionally, the structural similarity function of the improved generative adversarial network is an indicator for measuring the similarity of two seismic signals, which is expressed as:

[0020]

[0021] Among them, λ1 is the adjustment coefficient, μ x is the mean value of x, μ G(z) is the average value of G(z); σ x is the standard deviation of x, σ G(z) is the standard deviation of G(z), σ xG(z) is the covariance of x and G(z); k1 and k2 are set to 0.01 and 0.03 respectively; L is the dynamic range of the pixel; k1L and k2L are two quantities that maintain stability.

[0022] Optionally, the peak signal-to-noise ratio function of the improved generative adversarial network is a standard for measuring the strength of the seismic signal and the background noise level, and is expressed as:

[0023]

[0024] Among them, λ2 is the adjustment coefficient, N represents the total number of samples, x represents the sample data distribution, G(z) represents the data distribution of the generator, and ||*|| represents the Frobenius norm.

[0025] Optionally, the training the adversarial network to obtain the trained adversarial network specifically includes:

[0026] Train the discriminant model D;

[0027] The noisy data is sent to the generative model G for denoising to obtain the processed data;

[0028] The output result is passed to the discriminant model D for judgment.

[0029] Optionally, the training discriminant model D specifically includes:

[0030] The discriminant model D is trained with noisy labeled data so that the discriminant model has the ability to distinguish whether it contains noise.

[0031] Optionally, after passing the output result to the discriminant model D for judgment, the method further includes: passing adjustment parameters to the generation model.

[0032] The present invention provides a seismic denoising method based on an improved generative adversarial network, the denoising method comprising: improving the core function in the generative adversarial network; training the adversarial network to obtain a trained adversarial network; and performing network denoising on the trained adversarial network. The signal-to-noise ratio of the seismic signal is improved and is not affected by artificially selected feature functions or threshold parameters.

[0033] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0035] Figure 1 A schematic diagram of the improved generative adversarial network structure provided in an embodiment of the present invention.

[0036] Figure 2 Schematic diagram of the training structure of the generative adversarial network;

[0037] Figure 3 It is the comparison of the effects before and after random noise suppression;

[0038] (a) Actual seismic data containing random noise; (b) FX filtering result; (c) DnCNN processing result; (d) improved generative adversarial network processing result. DETAILED DESCRIPTION

[0039] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0040] The terms "comprises" and "having" and any variations thereof in the description embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.

[0041] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0042] The present invention proposes a denoising method based on an improved generative adversarial network. In view of the problem that the traditional denoising algorithm is affected by the artificially selected feature function or threshold parameter, which makes it difficult to remove the noise in the signal, a seismic signal denoising model based on a CNN network is proposed. The CNN network can effectively extract the characteristics of seismic signals through residual learning and batch normalization. By training and testing the model with a constructed data set, it is verified that the CNN model can efficiently remove the noise in the seismic signal; by using the wavelet threshold denoising algorithm, the empirical mode decomposition denoising algorithm and the trained CNN model to denoise the same section of seismic signals at different noise levels, the experimental results show that the denoising effect of the CNN model is better than these traditional denoising algorithms, and the signal-to-noise ratio of the seismic signal is improved, and it is not affected by the artificially selected feature function or threshold parameter.

[0043] The effect of supervised deep learning networks is highly dependent on the quality of sample data. Producing appropriate sample data can improve the training effect of generative adversarial networks. By improving the loss function and the network structure, generative adversarial networks can suppress random noise and surface wave noise in seismic data, which is better than ordinary neural network noise suppression. There is still room for improvement. Producing training samples is time-consuming and laborious. Using semi-supervised or unsupervised training methods may be the trend of subsequent development. Normalizing and preprocessing training data and test data can effectively solve the difference between model data and actual data, making the processing effect of actual data better.

[0044] like Figure 1 As shown in the figure, the denoising method based on the improved generative adversarial network is implemented as follows:

[0045] First, the core functions in the generative adversarial network are improved.

[0046] The improved generative adversarial network method uses the U-net structure, which has a good effect in the field of image segmentation, as a discriminant model of network constraints to extract noise feature information. The classic residual convolutional network structure is used as the generative model. The internal residual block can improve the overall network training speed and achieve better noise suppression effect. Figure 1 Schematic diagram of the structure of the improved generative adversarial model.

[0047] The mean square error loss function of the improved generative adversarial network is expressed as:

[0048]

[0049] Among them, N represents the total number of samples, x represents the sample data distribution, G(z) represents the data distribution of the generator, and ||*|| represents the Frobenius norm.

[0050] The improved final loss function is expressed as:

[0051]

[0052] The structural similarity function is an indicator used to measure the similarity of two seismic signals, which is expressed as:

[0053]

[0054] Among them, λ1 is the adjustment coefficient, μ x is the mean value of x, μ G(z) is the average value of G(z); σ x is the standard deviation of x, σ G(z) is the standard deviation of G(z), σ xG(z) is the covariance of x and G(z); k1 and k2 are set to 0.01 and 0.03 respectively; L is the dynamic range of the pixel; k1L and k2L are two quantities that maintain stability.

[0055] The peak signal-to-noise ratio function is a standard used to measure the strength of seismic signals and background noise levels, expressed as:

[0056]

[0057] Among them, λ2 is the adjustment coefficient, N represents the total number of samples, x represents the sample data distribution, G(z) represents the data distribution of the generator, and ||*|| represents the Frobenius norm.

[0058] Secondly, after improving the core functions in the generative adversarial network, in order to realize the improved generative adversarial network method, two major steps are adopted: one is network training; the other is network denoising.

[0059] The network training process is as follows Figure 2 As shown:

[0060] The discriminant model D is trained with noisy labeled data so that the discriminant model has the ability to distinguish whether it contains noise.

[0061] The noisy data is sent to the generative model G for denoising to obtain the processed data.

[0062] The output result is passed to the discriminant model D for judgment, and the adjustment parameters are passed to the generation model.

[0063] In order to verify the application effect of this method, this study is based on the three-dimensional seismic data of a certain exploration area, and this method is used to process the data to verify the effect of this method.

[0064] Prepare real seismic data containing random noise;

[0065] De-noising results of seismic data after FX filtering;

[0066] Denoising results based on commonly used convolutional neural networks;

[0067] Denoising results based on the improved generative adversarial network.

[0068] from Figure 3 It can be seen that the improved generative adversarial network has the best processing effect, which shows that this type of denoising method plays an important role in denoising actual data.

[0069] Beneficial effect: The seismic denoising method using the improved generative adversarial network proposed in the present invention can effectively improve the effect and efficiency of seismic data denoising, and can better serve the subsequent reservoir description and development.

[0070] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A seismic denoising method based on an improved generative adversarial network, characterized in that: The denoising method comprises: Improve the core functions in generative adversarial networks; Training the adversarial network to obtain a trained adversarial network; The trained adversarial network is subjected to network denoising processing.

2. A seismic denoising method based on an improved generative adversarial network according to claim 1, characterized in that: The improved generative adversarial network method specifically includes: using the U-net structure, which has a good effect in the field of image segmentation, as a discriminant model of network constraints to extract noise feature information.

3. The seismic denoising method based on improved generative adversarial network according to claim 1, characterized in that: The improved generative adversarial network adopts the classic residual convolutional network structure as the generation model.

4. The seismic denoising method based on improved generative adversarial network according to claim 1, characterized in that: The mean square error loss function of the improved generative adversarial network is expressed as: Among them, N represents the total number of samples, x represents the sample data distribution, G(z) represents the data distribution of the generator, and ||*|| represents the Frobenius norm.

5. The seismic denoising method based on improved generative adversarial network according to claim 1, characterized in that: The final loss function of the improved generative adversarial network is:

6. The seismic denoising method based on improved generative adversarial network according to claim 1, characterized in that: The structural similarity function of the improved generative adversarial network is an indicator used to measure the similarity of two seismic signals, which is expressed as: Among them, λ1 is the adjustment coefficient, μ x is the mean value of x, μ G(z) is the average value of G(z); σ x is the standard deviation of x, σ G(z) is the standard deviation of G(z), σ xG(z) is the covariance of x and G(z); k1 and k2 are set to 0.01 and 0.03 respectively; L is the dynamic range of the pixel; k1L and k2L are two quantities that maintain stability.

7. The seismic denoising method based on improved generative adversarial network according to claim 1, characterized in that: The peak signal-to-noise ratio function of the improved generative adversarial network is a standard for measuring the strength of the seismic signal and the background noise level, which is expressed as: Among them, λ2 is the adjustment coefficient, N represents the total number of samples, x represents the sample data distribution, G(z) represents the data distribution of the generator, and ||*|| represents the Frobenius norm.

8. The seismic denoising method based on improved generative adversarial network according to claim 1, characterized in that: The training of the adversarial network to obtain the trained adversarial network specifically includes: Train the discriminant model D; The noisy data is sent to the generative model G for denoising to obtain the processed data; The output result is passed to the discriminant model D for judgment.

9. The seismic denoising method based on improved generative adversarial network according to claim 8, characterized in that: The training discriminant model D specifically includes: The discriminant model D is trained with noisy labeled data so that the discriminant model has the ability to distinguish whether it contains noise.

10. The seismic denoising method based on improved generative adversarial network according to claim 8, characterized in that: After the output result is passed to the discriminant model D for judgment, the method further includes: passing adjustment parameters to the generation model.

Images