Adaptive multiple suppression method based on deep convolutional neural network

Abstract

The present invention provides a method for adaptive suppression of multiple waves based on a deep convolutional neural network, comprising: step 1: setting initial values ​​of variables according to seismic data; step 2: constructing a deep convolutional neural network; step 3: dividing the original seismic data and predicted multiple waves into multiple data windows, and training the deep convolutional neural network; step 4: using the trained deep convolutional neural network to calculate the estimated multiple waves of a single window; step 5: directly subtracting the estimated multiple waves from the original data to obtain the multiple wave suppression result; step 6: weighted fusion of the multiple wave suppression results of all data windows to obtain the final multiple wave suppression result. The method uses a deep convolutional neural network to reduce the adaptive subtraction of multiple waves to a nonlinear regression problem, which can better eliminate the complex differences between predicted multiple waves and real multiple waves, and effectively reduce residual multiple waves while protecting the primary wave.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploration seismic data processing, and in particular to a multiple wave adaptive suppression method based on a deep convolutional neural network. Background Art

[0002] In oil and gas seismic exploration, the original seismic data contains primary waves and multiple waves. Compared with primary waves, multiple waves reflect primary waves or multiple waves at the free surface or the interface below it. Usually, primary waves are used to image the underground medium. Therefore, before using primary waves for imaging, multiple waves need to be removed. Prediction and subtraction are two important steps in multiple wave removal. Compared with real multiple waves, predicted multiple waves contain complex differences in time, space, and sub-waves. Therefore, adaptive subtraction of predicted multiple waves instead of direct subtraction is the key to effectively suppress multiple waves.

[0003] At present, there are many adaptive subtraction methods, which can be performed in the time-space domain or the transform domain, including the Radon domain, the Curvelet domain, etc. Usually, adaptive subtraction is reduced to a linear regression problem. The traditional linear regression method uses a 2D matched filter for adaptive subtraction. According to different optimization problems, adaptive subtraction is divided into different methods. The adaptive subtraction method can impose L1 mode or L2 mode minimization constraints on the primary wave. Where the primary wave and the multiple waves overlap, the L1 mode can obtain better separation results of the primary wave and the multiple waves than the L2 mode. The fast iterative shrinkage threshold algorithm and the iterative reweighted least squares algorithm can be used to solve the L1 mode minimization optimization problem. Compared with the iterative reweighted least squares algorithm, the fast iterative shrinkage threshold algorithm can further improve the computational efficiency while maintaining the computational accuracy.

[0004] How to effectively balance the protection of primary waves and the suppression of multiple waves is the core issue of adaptive subtraction. The traditional linear regression method is performed in overlapping 2D data windows. The 2D data windows overlap in time and space, and a 2D matched filter is estimated in each 2D data window. Therefore, the traditional method cannot use the prediction information in multiple data windows for adaptive subtraction at the same time, and the 2D matched filter cannot effectively express the complex differences between predicted multiple waves and real multiple waves, which will cause primary wave damage or residual multiple waves. How to effectively eliminate the complex differences between predicted multiple waves and real multiple waves and improve the effects of primary wave protection and multiple wave suppression has important research significance and practical value.

[0005] In the Chinese patent application with application number: CN202010013146.5, a method for adaptive subtraction of multiple waves based on predicted feature extraction is involved. The method for adaptive subtraction of multiple waves based on predicted feature extraction includes: step 1, setting the initial value of the variable; step 2, constructing a convolutional neural network; step 3, training the convolutional neural network; step 4, extracting predicted multiple wave features using the trained convolutional neural network; step 5, constructing a mathematical model and optimization problem for adaptive subtraction of multiple waves using original data and predicted multiple wave features; step 6, solving the matched filter for adaptive subtraction of multiple waves; step 7, estimating the primary wave

[0006] In the Chinese patent application with application number: CN201911235220.1, a method and system for identifying the arrival time and phase of microseismic signals based on a deep neural network are involved. The method includes: dividing the training data set, the verification data set and the test data set; building a convolutional neural network model with a depth of 10 layers; calculating the errors of the input and output labels of the network to complete the update of the network parameters; performing model training; setting the number of iterative training to N, and stopping the network training after N iterative training; that is, the deep convolutional neural network model for identifying the arrival time and phase of microseismic signals is constructed; inputting the unprocessed data set into the trained network model to identify the probabilities of P waves, S waves and other signals.

[0007] In the Chinese patent application with application number: CN201910829017.0, a method for suppressing surface waves of seismic data based on a generative adversarial network is involved. The method mainly includes: creating a set of training data obtained through data processing; providing training data to the neural network, training the generator and the discriminator, and through network training, the data generated by the generator can deceive the discriminator, and the discriminator can identify the data generated by the generator and the training data; in the process of network training, by adopting the mutual confrontation training method between the generator and the discriminator, the generator can obtain good data and the discriminator can make more accurate identification; after completing the network training, the original seismic data containing surface waves is provided to the generator, and the generator generates seismic data without surface waves.

[0008] The above existing technologies are quite different from the present invention and fail to solve the technical problem we want to solve. Therefore, we have invented a new multiple wave adaptive suppression method based on deep convolutional neural network. Summary of the invention

[0009] The purpose of the present invention is to provide a multiple wave adaptive suppression method based on deep convolutional neural network, which can better eliminate the complex differences between predicted multiple waves and real multiple waves.

[0010] The purpose of the present invention can be achieved by the following technical measures: a multiple wave adaptive suppression method based on a deep convolutional neural network, the multiple wave adaptive suppression method based on a deep convolutional neural network includes:

[0011] Step 1: Set the initial value of the variable according to the earthquake data;

[0012] Step 2: Build a deep convolutional neural network;

[0013] Step 3: Divide the original seismic data and predicted multiple waves into multiple data windows and train the deep convolutional neural network;

[0014] Step 4: Use the trained deep convolutional neural network to calculate the estimated multiple waves of a single window;

[0015] Step 5: Subtract the estimated multiple waves of a single window directly from the original data to obtain the multiple wave suppression result;

[0016] Step 6: Weighted fusion of the multiple wave suppression results of all data windows to obtain the final multiple wave suppression result.

[0017] The purpose of the present invention can also be achieved by the following technical measures:

[0018] In step 1, set the initial values ​​of the variables. The initial value variables that need to be set include the 2D data window size n×n, the 2D convolution kernel size m×m, the number of layers N for downsampling of the deep convolutional neural network, the number of convolution kernels P in the first downsampling, the iteration round M, the batch size S of the training samples, and the initial learning rate l.

[0019] In step 2, a deep convolutional neural network is constructed, which is divided into two steps: downsampling and upsampling calculations.

[0020] In step 2, the basic calculations included in the downsampling calculation are convolution calculation, batch normalization calculation, nonlinear activation function ReLu (max (0, x)) calculation, and maximum pooling calculation, where x is the input element.

[0021] In step 2, after the downsampling calculation step and before the upsampling calculation step, there are two consecutive calculation sequences, each of which includes convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation.

[0022] In step 2, the basic calculations included in the upsampling calculation are up-convolution calculation, convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation.

[0023] In step 2, the deep convolutional neural network connects the output data after the up-convolution calculation step with the output data after the symmetric down-sampling calculation step, and the connected data is used as the input of the down-sampling convolution calculation; in addition, the last layer of the deep convolutional neural network includes convolution calculation, batch normalization calculation and nonlinear calculation. x represents the independent variable in the nonlinear calculation, and the data is output after the nonlinear calculation Tanh.

[0024] In step 3, the seismic data is divided into multiple data windows, and the predicted multiple waves of each data window are used as the input of the deep convolutional neural network. The original data of the corresponding data window is used as a label to train the deep convolutional neural network. After all data windows are iterated, a trained deep convolutional neural network is formed.

[0025] In step 4, the predicted multiple waves of a single data window are input into the trained deep convolutional neural network to obtain the multiple wave estimation results in the single data window.

[0026] In step 5, the multiple wave estimation result in a single data window is directly subtracted from the original data in the window to obtain the multiple wave suppression result in the data window.

[0027] In step 6, steps 4 and 5 are repeated to obtain the multiple wave estimation and suppression results of all data windows, and the multiple wave suppression results of all data windows are weightedly fused to obtain the final multiple wave adaptive suppression results.

[0028] The method for adaptive suppression of multiple waves based on deep convolutional neural network in the present invention uses the original seismic data of multiple windows and predicted multiple waves to train the deep convolutional neural network for adaptive suppression of multiple waves, and then uses the deep convolutional neural network to achieve multiple wave suppression. Compared with the traditional linear regression method, the present invention uses the deep convolutional neural network to reduce the adaptive subtraction of multiple waves to a nonlinear regression problem, which can better eliminate the complex differences between predicted multiple waves and real multiple waves, and effectively reduce the residual multiple waves while protecting the primary wave. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flowchart of a specific embodiment of the method for adaptive suppression of multiple waves based on deep convolutional neural network of the present invention;

[0030] Figure 2 It is a schematic diagram of raw data, predicted multiple waves and true primary waves in a specific embodiment of the present invention;

[0031] Figure 3 is a schematic diagram of the structure of a deep convolutional neural network in a specific embodiment of the present invention;

[0032] Figure 4 It is a schematic diagram of the multiple wave estimation result of the deep convolutional neural network, the result after suppressing the multiple waves, and the difference profile between the multiple wave suppression result and the true primary wave in a specific embodiment of the present invention;

[0033] Figure 5 It is a schematic diagram of the multiple wave estimation result, the result after suppressing the multiple waves, and the difference profile between the multiple wave suppression result and the true primary wave using the traditional linear regression method in a specific embodiment of the present invention when a large data window is used;

[0034] Figure 6 It is a schematic diagram of the multiple wave estimation result, the result after suppressing the multiple waves, and the difference profile between the multiple wave suppression result and the true primary wave using the traditional linear regression method in a specific embodiment of the present invention when a small data window is used;

[0035] Figure 7 It is a schematic diagram of raw data, predicted multiple waves, multiple wave suppression results of a deep convolutional neural network, and multiple waves estimated by a deep convolutional neural network in a specific embodiment of the present invention;

[0036] Figure 8 It is a schematic diagram of a multiple wave suppression result when a traditional linear regression method uses a large data window length, a multiple wave estimated when a traditional linear regression method uses a large data window length, a multiple wave suppression result when a traditional linear regression method uses a small data window length, and a multiple wave estimated when a traditional linear regression method uses a small data window length in a specific embodiment of the present invention;

[0037] Fig. 9 It is a schematic diagram of the amplification result of the original data, the amplification result of the predicted multiple waves, the amplification result after the deep convolutional neural network suppresses the multiple waves, the amplification result after the traditional linear regression method suppresses the multiple waves when the large data window length is used, and the amplification result after the traditional linear regression method suppresses the multiple waves when the small data window length is used in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0039] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations and / or combinations thereof.

[0040] like Figure 1 As shown, Figure 1 The flowchart of the method for adaptively suppressing multiple waves based on a deep convolutional neural network of the present invention is as follows:

[0041] Step 101, setting the initial values ​​of variables, the initial value variables that need to be set include the 2D data window size n×n, the 2D convolution kernel size m×m, the number of layers N of the deep convolutional neural network downsampling, the number of convolution kernels P in the first downsampling, the iteration round M, the batch size S of the training sample, and the initial learning rate l;

[0042] Step 102, constructing a deep convolutional neural network, is divided into two steps: downsampling and upsampling calculations. The first is the downsampling calculation, which includes the basic calculations of convolution calculation, batch normalization calculation, nonlinear activation function ReLu (max (0, x)) calculation, and maximum pooling calculation, where x is the input element. After the downsampling calculation step and before the upsampling calculation step, there are two consecutive calculation sequences, each of which includes convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation. Then comes the upsampling calculation, which includes the basic calculations of up-convolution calculation, convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation. The deep convolutional neural network connects the output data after the up-convolution calculation step with the output data after the symmetric downsampling calculation step, and the connected data is used as the input of the downsampling convolution calculation. In addition, the last layer of the deep convolutional neural network includes convolution calculation, batch normalization calculation, and nonlinear calculation. Output data after nonlinear calculation of Tanh.

[0043] Step 103, divide the data windows, iteratively train to form a deep convolutional neural network. Divide the seismic data into multiple data windows, use the predicted multiple waves of each data window as the input of the deep convolutional neural network, and use the original data of the corresponding data window as a label to train the deep convolutional neural network. After all data windows are iterated, a trained deep convolutional neural network is formed;

[0044] Step 104, using the trained deep convolutional neural network to estimate multiple waves in a single window; inputting the predicted multiple waves of the single data window into the trained deep convolutional neural network to obtain the multiple wave estimation result in the single data window;

[0045] Step 105, directly subtract the estimated multiple waves from the original data to obtain the multiple wave suppression result of the single window. Directly subtract the multiple wave estimation result in a single data window from the original data in the window to obtain the multiple wave suppression result in the data window;

[0046] Step 106, repeating steps 104 and 105, weighted fusion of the multiple wave suppression results of all data windows, and obtaining the final multiple wave automatic suppression result. Repeating steps 104 and 105, obtaining the multiple wave estimation and suppression results of all data windows, weighted fusion of the multiple wave suppression results of all data windows, and obtaining the final multiple wave automatic suppression result.

[0047] In a specific embodiment 1 of the present invention, the multiple wave adaptive suppression method based on deep convolutional neural network includes the following steps:

[0048] (1) Set the initial values ​​of variables; Set the initial values ​​of variables. The initial value variables that need to be set include the 2D data window size n×n, the 2D convolution kernel size m×m, the number of layers N of the deep convolutional neural network downsampling, the number of convolution kernels P in the first downsampling, the iteration round M, the batch size S of the training samples, and the initial learning rate l;

[0049] (2) Constructing a deep convolutional neural network; Constructing a deep convolutional neural network is divided into two steps: downsampling and upsampling calculations. The first is the downsampling calculation, which includes the basic calculations of convolution calculation, batch normalization calculation, nonlinear activation function ReLu (max(0,x)) calculation, and maximum pooling calculation, where x is the input element. After the downsampling calculation step and before the upsampling calculation step, there are two consecutive calculation sequences, each of which includes convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation. Then comes the upsampling calculation, which includes the basic calculations of up-convolution calculation, convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation. The deep convolutional neural network connects the output data after the up-convolution calculation step with the output data after the symmetric downsampling calculation step, and the connected data is used as the input of the downsampling convolution calculation. In addition, the last layer of the deep convolutional neural network includes convolution calculation, batch normalization calculation, and nonlinear calculation. x represents the independent variable in the nonlinear calculation, and the data is output after the nonlinear calculation Tanh.

[0050] (3) Dividing the data windows and iteratively training to form a deep convolutional neural network: The seismic data is divided into multiple data windows, and the predicted multiple waves of each data window are used as the input of the deep convolutional neural network. The original data of the data window is used as a label to train the deep convolutional neural network. After all data windows are iterated, a trained deep convolutional neural network is formed;

[0051] (4) Using the trained deep convolutional neural network to estimate multiple waves in a single window;

[0052] (5) The estimated multiple waves are directly subtracted from the original data to obtain the multiple wave suppression result of a single window.

[0053] (6) Repeat steps 4 and 5 to obtain the multiple wave estimation and suppression results of all data windows, and perform weighted fusion on the multiple wave suppression results of all data windows to obtain the final multiple wave suppression results.

[0054] Figure 2 The original seismic data and predicted multiple waves and true primary waves are shown, among which, Figure 2 The original seismic data in (a) contains primary waves and multiple waves. Figure 2 (b) To predict multiple waves, Figure 2 (c) is the real primary wave. By using the operation steps of the present invention, by setting the initial value of the variable, construct Figure 3 The deep convolutional neural network shown in the figure, where Conv means convolution, BN means batch normalization calculation, ReLu means nonlinear activation function, MP means maximum pooling calculation, Concatenate means connecting two data together, Up-Conv means up-convolution, and Tanh means Tanh nonlinear calculation. Figure 3 The deep convolutional neural network shown is trained and used to estimate multiple waves. Figure 4 (a) is the multiple waves estimated by the deep convolutional neural network in the present invention. The estimated multiple waves are directly subtracted from the original data to obtain the single-window primary wave estimation result. The primary wave estimation results of multiple data windows are weighted and fused to obtain the final multiple wave suppression result, which is shown in Figure 4 (b), Figure 4 (c) is the difference profile obtained by subtracting the multiple wave suppression result of the present invention from the true primary wave. It can be seen that there is basically no shadow of the primary wave in the difference profile, which shows that the method of the present invention can better protect the energy of the primary wave, and the residual multiple wave energy is weak, which proves the accuracy and effectiveness of the adaptive suppression method of multiple waves. In order to further prove the advantages of the present method, the suppression effect of the conventional method and the difference profile are compared, which is shown in Figure 5 and Figure 6 middle. Figure 5 (a) is the multiple waves estimated by the traditional method using a large data window, Figure 5 (b) is the multiple wave suppression result obtained by the traditional method using a large data window. Figure 5 (c) is the difference profile between the multiple wave suppression result using the traditional method with a large data window and the true primary wave. Figure 6 (a) is the multiple waves estimated by the traditional method using a small data window, Figure 6 (b) is the multiple wave suppression result obtained by the traditional method using a small data window. Figure 6 (c) is the difference profile between the multiple wave suppression result using a small data window and the true primary wave using the traditional method. Figure 4 , Figure 5 and Figure 6 At the location indicated by the white arrow, it can be seen that the traditional method will cause residual multiple waves. Figure 6 The black arrow in (c) shows that the traditional method using a small data window will cause primary wave damage. Compared with the traditional method, the method of the present invention can better remove residual multiple waves while protecting the primary wave, which reflects the technical advantages of the method of the present invention.

[0055] In the specific embodiment 2 of the present invention, Figure 7 (a) is the original data, Figure 7 (b) To predict multiple waves, Figure 7 (c) is the result after suppressing multiple waves using the deep convolutional neural network method. Figure 7 (d) is the multiple waves estimated by the deep convolutional neural network method. Figure 8 (a) is the result of the traditional linear regression method after suppressing multiple waves when using a large data window length. Figure 8 (b) Multiple waves estimated by the traditional linear regression method when using a large data window length. Figure 7 (c) and Figure 8 As can be seen from the white arrows in (a), the deep convolutional neural network method can suppress the residual multiple waves better than the traditional linear regression method (when using a large data window length). Figure 8 (c) is the result of the traditional linear regression method after suppressing multiple waves when a small data window length is used. Figure 8 (d) Multiple waves estimated by the traditional linear regression method when a small data window length is used. Figure 8 The white arrows in (c) show that the traditional linear regression method can effectively suppress the residual multiple waves when a small data window length is used. Fig. 9 (a) is the magnified result of the original data. Fig. 9 (b) is the predicted amplification result of multiple waves. Fig. 9 (c) is the amplification result after the deep convolutional neural network suppresses multiple waves. Fig. 9 (d) is the amplified result after suppressing multiple waves when the traditional linear regression method uses a large data window length. Fig. 9(e) is the amplified result after suppressing multiple waves when the traditional linear regression method uses a small data window length. Fig. 9 (c) Fig. 9 (d) and Fig. 9 The black arrow in (e) shows that the traditional linear regression method will damage the primary wave when a small data window length is used. The deep convolutional neural network method can better protect the primary wave than the traditional linear regression method (when a small data window length is used).

[0056] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0057] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. A multiple wave adaptive suppression method based on a deep convolutional neural network, characterized in that: The multiple wave adaptive suppression method based on deep convolutional neural network includes: Step 1: Set the initial value of the variable according to the earthquake data; Step 2: Build a deep convolutional neural network; Step 3: Divide the original seismic data and predicted multiple waves into multiple data windows and train the deep convolutional neural network; Step 4: Use the trained deep convolutional neural network to calculate the estimated multiple waves of a single window; Step 5: Subtract the estimated multiple waves directly from the original data to obtain the multiple wave suppression result; Step 6: Weighted fusion of the multiple wave suppression results of all data windows to obtain the final multiple wave suppression result; In step 1, set the initial values ​​of the variables. The initial value variables that need to be set include the 2D data window size n×n, the 2D convolution kernel size m×m, the number of layers N for downsampling of the deep convolutional neural network, the number of convolution kernels P in the first downsampling, the iteration round M, the batch size S of the training sample, and the initial learning rate l; In step 2, a deep convolutional neural network is constructed, which is divided into two steps: downsampling and upsampling calculations; The basic calculations included in the downsampling calculation are convolution calculation, batch normalization calculation, nonlinear activation function ReLu (max (0, x)) calculation, and maximum pooling calculation, where x is the input element; After the downsampling calculation step and before the upsampling calculation step, there are two consecutive calculation sequences, each of which contains convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation; The basic calculations included in the upsampling calculation are up-convolution calculation, convolution calculation, batch normalization calculation, and nonlinear activation function ReLu calculation; In step 2, the deep convolutional neural network connects the output data after the up-convolution calculation step with the output data after the symmetric down-sampling calculation step, and the connected data is used as the input of the down-sampling convolution calculation; in addition, the last layer of the deep convolutional neural network includes convolution calculation, batch normalization calculation and nonlinear calculation. x represents the independent variable in the nonlinear calculation, and the data is output after the nonlinear calculation Tanh.

2. The method for adaptive suppression of multiple waves based on deep convolutional neural network according to claim 1, characterized in that: In step 4, the predicted multiple waves of a single data window are input into the trained deep convolutional neural network to obtain the multiple wave estimation results in the single data window.

3. The method for adaptive suppression of multiple waves based on deep convolutional neural network according to claim 1, characterized in that: In step 5, the multiple wave estimation result in a single data window is directly subtracted from the original data in the window to obtain the multiple wave suppression result in the data window.

4. The method for adaptive suppression of multiple waves based on deep convolutional neural network according to claim 1, characterized in that: In step 6, steps 4 and 5 are repeated to obtain the multiple wave estimation and suppression results of all data windows, and the multiple wave suppression results of all data windows are weightedly fused to obtain the final multiple wave adaptive suppression results.

Images