Seismic data reconstruction method based on convolutional neural network

Abstract

The invention provides a seismic data reconstruction method based on a convolutional neural network, and the method comprises the steps: 1, randomly initializing a network parameter theta, and inputting noise Z; step 2, the reconstruction network f takes Z as an input and takes preliminary reconstruction data f theta (Z) with the same size as an output; 3, calculating local optimal parameters of the network; 4, reconstructing test data based on the trained network model; and step 5, evaluating the performance of the method in irregular and regular seismic data reconstruction. The seismic data reconstruction method based on the convolutional neural network is used for interpolation of irregular and rule missing channels in pre-stack and post-stack seismic data, and experimental results show that the method is superior to a processing type Cadzow method commonly used in traditional singular spectrum analysis and reconstruction of the data.

Description

Technical Field

[0001] The present invention relates to the technical field of seismology, and in particular to a method for reconstructing seismic data based on a convolutional neural network. Background Art

[0002] Seismic data acquisition provides actionable geophysical information on the characteristics of geological structures and the distribution of stratigraphic units. The quality of the data obtained from the receiver directly affects subsequent seismic processing and interpretation operations, such as migration and inversion. Due to physical or financial constraints, seismic data are often spatially undersampled, effectively resulting in missing traces in the acquired data, which increases the difficulty of subsequent seismic data processing. Therefore, it is necessary to reconstruct the seismic data so that subsequent data processing can better describe complex geological structures.

[0003] With the development of machine learning, deep learning (DL) methods have attracted great attention in seismic data reconstruction. Generative adversarial networks have been explored for seismic noise attenuation and reconstruction of irregular seismic traces. The training labels for the neural network (i.e., the original complete data in the training dataset) are synthetic shots with real data geometry generated using a finite difference modeling engine. Wang et al. (2018) proposed a DL-based method to reconstruct regularly missing traces using a trained residual network. However, the performance of this method degrades when the features in the test dataset differ from those in the training dataset, and the reconstruction error increases as this difference increases. To reconstruct seismic data, regardless of the type of sampling method used to collect the data, Siahkoohi et al. (2018) used a generative adversarial network that can handle severely undersampled seismic data; however, this network requires a certain proportion of fully sampled shots as a training dataset. To reconstruct irregular missing traces, Oliveira et al. (2018) used a conditional generative adversarial network, which can result in lower resolution and some non-geological artifacts when the data structure is complex. All of these DL-based methods are under the framework of "train-validate-test", and they are all trained using large datasets. However, it is challenging to obtain sufficient earthquake datasets.

[0004] Most of the state-of-the-art DL methods are trained using large seismic datasets, and their outstanding performance may be attributed to their ability to learn the projection between corrupted and intact data from large datasets. However, their success is limited due to the number of available training datasets. In contrast, Ulyanov et al. (2018) demonstrated that convolutional neural networks (CNNs) have an intrinsic ability to handle image restoration without pre-training. Due to the lax nature of the image reconstruction problem, regularization is traditionally applied by constraining the solution to be consistent with prior knowledge (i.e., historical information; henceforth referred to as “priors”), such as sparsity and non-local self-similarity. Instead of using these predefined image priors, Ulyanov et al. (2018) used a CNN structure that acts as a regularizer to guide network learning, and the learned network weights are the parameters that represent the image. The U-net network is a classic CNN with a U-shaped structure. It does not have any fully connected layers (i.e., the nodes of each previous layer are connected to the nodes of the previous layer), and it only includes convolutional layers, which has shown excellent performance in computer vision tasks such as image inpainting, image denoising, and super-resolution reconstruction.

[0005] In the Chinese patent application with application number: CN202111440070.5, a method for reconstructing seismic data based on a non-local attention convolutional neural network is involved. The present invention establishes long-distance feature relationships through a non-local attention mechanism, and establishes dependencies between local features by calculating the correlation weights between different local features, so that local features can perceive global information, and enhance the learning ability of local features by introducing more global information. Therefore, non-local attention can establish long-distance feature relationships without significantly increasing the depth of the network, and effectively integrate local features and global features. The invention adopts a deep residual network module, which contains multiple short jump connections and one long jump connection. While effectively preventing overfitting, it can also allow a large number of shallow features to be transmitted to the deep layer, facilitating the flow of information and allowing the main network to focus on reconstructing missing seismic data.

[0006] In the Chinese patent application with application number: CN201811540278.2, a method for training a model for reconstructing well logging data is involved, including: 1) reducing the dimension of the collected well logging attribute data; 2) using the false data for the well logging attribute generated by the generator network G based on its current parameters as an input of the discriminant network D, and using the result of the dimensionality reduction process as another input of the discriminant network D, and adjusting the parameters of the generator network G so that the discriminant network D judges the false data as the real collected data; wherein the generator network G is a convolutional neural network, and the false data corresponds to a plane coordinate. The problem of difficulty in obtaining seismic data of well logging with an earlier mining period to train the neural network is avoided, and a model that can produce prediction results closest to the real data is obtained through adversarial training.

[0007] In the Chinese patent application with application number: CN202011071604.7, a method and system for seismic data reconstruction based on deep learning is involved, which belongs to the field of geophysical exploration data preprocessing. The method includes: (1) collecting seismic data; (2) generating a neural network; (3) using seismic data to train the neural network to obtain a reconstruction model; (4) inputting the seismic data to be reconstructed into the reconstruction model, and the reconstruction model outputs the reconstructed seismic data. This invention uses the U-Net network in deep learning to extract the deep features of the seismic data through convolution and pooling operations of varying degrees for a large amount of data with various irregularities collected during seismic exploration, reduce the dimension of the data, and then upsample, that is, restore the original image size through interpolation and feature image fusion to achieve data reconstruction.

[0008] In the Chinese patent application with application number: CN202110645817.4, a seismic data reconstruction method, device, equipment and medium are involved, which are applied to the field of seismic exploration technology to solve the problems of high acquisition cost and low exploration accuracy of high-density seismic data existing in the prior art. Specifically, the sparse seismic data and high-density seismic data of each geological model are used to train a deep neural network model that characterizes the nonlinear mapping relationship between sparse seismic data and high-density seismic data, so that the seismic data reconstruction model finally trained can obtain high-density seismic data based on the sparse seismic data to be reconstructed, thereby realizing the reconstruction of seismic data, thereby reducing the acquisition cost of high-density seismic data and improving the exploration accuracy of high-density seismic data.

[0009] The above existing technologies are significantly different from the present invention and fail to solve the technical problem we want to solve. For this reason, we have invented a new seismic data reconstruction method based on convolutional neural network. Summary of the invention

[0010] The purpose of the present invention is to provide a convolutional neural network seismic data reconstruction method which utilizes U-net to capture the intrinsic properties hidden in seismic data and has excellent interpolation performance.

[0011] The purpose of the present invention can be achieved by the following technical measures: a convolutional neural network-based seismic data reconstruction method, the convolutional neural network-based seismic data reconstruction method comprising:

[0012] Step 1: Randomly initialize network parameters θ and input noise Z;

[0013] Step 2: The reconstruction network f takes Z as input and converts the preliminary reconstruction data f with the same size θ (Z) as output;

[0014] Step 3: Calculate the local optimal parameters of the network;

[0015] Step 4: Reconstruct the test data based on the trained network model;

[0016] Step 5: Evaluate the performance of the proposed method in the reconstruction of irregular and regular seismic data.

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

[0018] In step 1, the seismic data are normalized to the range of 0-1; the input noise Z is a random number sequence ranging from 0 to 0.1, with uniform distribution and similar size, i.e., dimension, to X.

[0019] In step 2, the CNN network output is used to reconstruct the data f θ (Z); The CNN network has an encoder-decoder structure with skip connections, and adds a normalization algorithm and a nonlinear activation function; these can speed up training.

[0020] In step 2, the CNN network has the following features: (1) Downsampling is performed by convolution with a kernel size of 3×3, and a normalization algorithm and a nonlinear activation function are added; these speed up the training; the number of filters is 16, 32, 64, 128, and 128; (2) Upsampling is performed by bilinear interpolation with a kernel size of 3×3; the number of filters is 128, 128, 64, 32, and 16; (3) The skip connection is formed by copying the features of the left encoder path and adding them to the right decoder path, thus changing a small number of hyperparameters; the height and width of the input can be randomly selected as needed; noise-based regularization is used when fitting the network; that is, at each iteration, additive normal noise is used to perturb the input Z; this regularization can help better fit the network and achieve the desired goal.

[0021] In step 3, in an iterative process, the reconstructed data f is calculated by θ The mean square error MSE between (Z) and the observed seismic data M is used to obtain the local optimal parameters of the network.

[0022] In step 3, in the seismic data reconstruction problem, M∈R m×n represents the observed undersampled seismic data, i.e., the damaged missing seismic traces, X∈R m×n Represents the original complete seismic data; for the reconstruction task, the goal is to reconstruct X from M, only the missing seismic traces must be reconstructed by the network, while the non-missing seismic traces remain unchanged.

[0023] In step 3, using the CNN network, the loss function is determined as:

[0024]

[0025] Among them, f θ represents the designed CNN, with parameters θ, including network layers, the number of cores in each network layer, and the weight and bias of each core; Z represents the fixed input, i.e., random noise, which obeys some common distributions, including Gaussian or uniform distribution; P Ω : R m×n →R m×n represents the sampling operator, defined as

[0026]

[0027] where Ω represents the subset of indices corresponding to the observed entries; P Ω represents the projection onto the observed data while keeping the entries in Ω and setting zeros at the locations of missing traces; represents the Frobenius norm, defined as θ * represents the local optimal parameters of the network obtained by minimizing the loss function given in Equation 1; the sampling operator P Ω Copy to Output , making the entries in Ω identical to those in the observed seismic data M.

[0028] In step 3, during network training, f is generated from the input Z at each iteration. θ (Z), and P Ω Projection to f θ (Z); Calculate the loss based on the MSE criterion Then update the network parameters θ by gradient descent method; finally get the optimal parameters

[0029] In step 4, after saving the trained network parameters, the test data is reconstructed based on the trained network model, that is, by setting P Ω (X) = P Ω (M) Get

[0030] In step 5, the quality of the reconstruction is represented by the signal-to-noise ratio S / N, defined as:

[0031]

[0032] where X and X * represent the original seismic data and the reconstructed data, respectively.

[0033] The convolutional neural network-based seismic data reconstruction method of the present invention captures the prior based on the special structure of CNN, but it does not require any training data set. Therefore, the bad inverse problem in seismic interpolation is solved using the CNN structure as a prior, and the learned network weights are the parameters representing the seismic data. Since the weights of the convolution filter are shared to achieve spatial invariance, the CNN structure can be used as a regularizer to guide network learning. The damaged seismic data is reconstructed by minimizing the mean square error between the network output and the original data during the iteration process. The present invention is used to interpolate irregular and regularly missing traces in pre-stack and post-stack seismic data. Experimental results show that this method is superior to traditional singular spectrum analysis and the processed Cadzow method commonly used in reconstructing such data. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flowchart of a specific embodiment of the method for seismic data reconstruction based on convolutional neural network of the present invention;

[0035] Figure 2 A schematic diagram of a seismic data reconstruction experiment of model data in a specific embodiment of the present invention;

[0036] Figure 3 It is a schematic diagram of a seismic data reconstruction experiment of actual data of a certain onshore exploration area in a specific embodiment of the present invention;

[0037] Figure 4 It is a schematic diagram of a seismic data reconstruction experiment of actual data of a certain offshore exploration area 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 reconstructing seismic data based on a convolutional neural network of the present invention is as follows. The method for reconstructing seismic data based on a convolutional neural network comprises:

[0041] Step 1: Randomly initialize the network parameters θ and input noise Z. The seismic data are all normalized to the range of 0-1. The input noise Z is a random number sequence (range, 0-0.1) with uniform distribution and similar size (i.e. dimension) to X.

[0042] Step 2: Reconstruct the network f taking Z as input and outputting data f with the same size θ (Z) as output.

[0043] The CNN network used in the present invention has an encoder-decoder structure with skip connections. It has the following features: (1) Downsampling is performed by convolution, the convolution kernel is 3×3, and a normalization algorithm and a nonlinear activation function are added; these speed up the training speed. The number of filters is 16, 32, 64, 128 and 128. (2) Upsampling is performed by bilinear interpolation, and the convolution kernel is 3×3. The number of filters is 128, 128, 64, 32 and 16. (3) The skip connection is formed by copying the features of the left encoder path and adding them to the right decoder path, thereby changing a small number of hyperparameters. Note that the height and width of the input can be randomly selected as needed. Noise-based regularization is used when fitting the network; that is, additive normal noise is used to perturb the input Z at each iteration. This regularization can help better fit the network and achieve the desired goals.

[0044] Step 3: In the iterative process, by minimizing the output f θ The mean square error (MSE) between (Z) and the observed seismic data M is used to obtain the (local) optimal parameters of the network. In the seismic data reconstruction problem, M∈R m×n represents the observed undersampled seismic data (i.e., the destroyed missing seismic traces), X∈R m×n Represents the original complete seismic data. For the reconstruction task, the goal is to reconstruct X from M, and only the missing seismic traces must be reconstructed by the network, while the non-missing seismic traces remain unchanged. Therefore, using the CNN network, the loss function is determined as

[0045]

[0046] Among them, f θ represents the designed CNN, with parameters θ, including network layers, the number of cores in each network layer, and the weight and bias of each core; Z represents fixed input (random noise), which obeys some common distributions such as Gaussian or uniform distribution; P Ω : R m×n →R m ×n represents the sampling operator, defined as

[0047]

[0048] where Ω represents the subset of indices corresponding to the observed entries; P Ω represents the projection onto the observed data while keeping the entries in Ω and setting zeros at the locations of missing traces; represents the Frobenius norm, defined as θ * represents the (local) optimal parameters of the network obtained by minimizing the loss function given in Equation 1. Ω Copy to Output , making the entries in Ω identical to those in the observed seismic data M.

[0049] During network training, f is generated from the input Z at each iteration. θ (Z), and P Ω Projection to f θ (Z). Loss is calculated based on the MSE criterion Then the network parameters θ are updated by gradient descent method. Finally, the optimal parameters are obtained

[0050] Step 4: After saving the trained network parameters, reconstruct the test data based on the trained network model, that is, by setting P Ω (X) = P Ω (M) Get

[0051] Step 5: Evaluate the performance of the proposed method in the reconstruction of irregular and regular seismic data. The quality of the reconstruction is represented by the signal-to-noise ratio (S / N), defined as

[0052]

[0053] where X and X * Represent the original seismic data and the reconstructed data respectively

[0054] According to the reconstruction process described above, the parameters of the network θ are first randomly initialized. During the network training process, f is generated from the input Z at each iteration. θ (Z), and P Ω Projection to f θ (Z). Loss is calculated based on the MSE criterion Then the network parameters θ are updated by gradient descent method. Finally, the optimal parameters are obtained After saving the trained network parameters, the test data is reconstructed based on the trained network model, that is, by setting P Ω (X) = P Ω (M) Get

[0055] The following are several specific embodiments of the present invention.

[0056] Example 1

[0057] In a specific embodiment 1 of the present invention, Figure 2 As shown in a, the model data is used to conduct a convolutional neural network seismic data reconstruction experiment. The simulated data is obtained based on the layered model forward modeling. Multiple waves are eliminated by smoothing the velocity and the direct wave is cut off. The single shot record is noise-free and has no boundary reflection. The seismic data after random missing 30% is shown in Figure 2 As shown in b, the event axis is no longer continuous. Figure 2 c shows the reconstructed data. The phase axis becomes continuous and conforms to the time-distance relationship of hyperbolic reflection. No computational noise is introduced into the overall profile, which is consistent with the original complete data. This shows that the reconstruction of seismic data based on convolutional neural network has achieved ideal results.

[0058] Example 2

[0059] In a specific embodiment 2 of the present invention, in order to verify the practical application effect of the patent method, the actual data of a land exploration area is reconstructed based on the convolutional neural network seismic data. Figure 3 As shown in , its time sampling and seismic trace spacing are 0.004 seconds and 20 meters respectively. 40% of the traces are randomly deleted from the original data, and the first five traces are also deleted to simulate the traces lost at near offsets, as shown in Figure 3 As shown in a, the signal-to-noise ratio = 3.63dB. Figure 3 b shows the reconstructed data, whose signal-to-noise ratio restored using this method is 10.01 dB. Figure 3 c shows the original complete field data. Figure 3 b and Figure 3c, we can see that the convolutional neural network-based seismic data reconstruction has well restored the missing seismic data, the wave field morphology and dynamic characteristics have also been restored, and the position, amplitude and waveform of the event axis have good continuity. This proves the effectiveness and practicality of this method in land data reconstruction.

[0060] Example 3

[0061] In a specific embodiment 3 of the present invention, in order to further verify the practical application effect of the patent method, the actual data of a certain offshore exploration area is reconstructed based on the convolutional neural network seismic data. The original non-missing seismic data is as follows Figure 4 As shown in a, the time sampling and seismic trace interval are 0.002 seconds and 12.5 meters respectively. 35% of the seismic traces are randomly missing from the original data, as shown in Figure 4 b. In the missing seismic data, both the reflection wave field and the diffraction wave field are seriously affected, which will reduce the quality of subsequent stacking and migration imaging. Figure 4 b. Recover the missing data while keeping the non-missing data unchanged. The reconstructed data is as follows: Figure 4 c. It can be seen that the underground seismic reflection wave field and diffraction wave field are reconstructed with high quality, and the hyperbolic reflection wave and diffraction wave can be clearly seen in the shot record. In addition to the effective signal being restored, the reconstructed shot record has a certain improvement in signal-to-noise ratio because the noise cannot be reconstructed. This proves that this method is also applicable to the reconstruction of offshore data.

[0062] As can be seen from the figure, the convolutional neural network seismic data reconstruction method proposed in this patent can effectively restore seismic data, thereby better serving subsequent reservoir description and development.

[0063] 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.

[0064] 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 method for seismic data reconstruction based on convolutional neural network, characterized in that: The convolutional neural network-based seismic data reconstruction method includes: Step 1: Randomly initialize network parameters θ and input noise Z; Step 2: The reconstruction network f takes Z as input and the preliminary reconstruction data f with the same size θ (Z) as output; Step 3: Calculate the local optimal parameters of the network; Step 4: Reconstruct the test data based on the trained network model; Step 5: Evaluate the performance of the proposed method in the reconstruction of irregular and regular seismic data.

2. The method for seismic data reconstruction based on convolutional neural network according to claim 1, characterized in that: In step 1, the seismic data are normalized to the range of 0-1; the input noise Z is a random number sequence ranging from 0 to 0.1, with uniform distribution and similar size, i.e., dimension, to X.

3. The method for seismic data reconstruction based on convolutional neural network according to claim 1, characterized in that: In step 2, the CNN network output is used to reconstruct the data f θ (Z); The CNN network has an encoder-decoder structure with skip connections, and adds a normalization algorithm and a nonlinear activation function; these can speed up training.

4. The method for seismic data reconstruction based on convolutional neural network according to claim 3, characterized in that: In step 2, the CNN network has the following features: (1) Downsampling is performed by convolution with a kernel size of 3×3, and a normalization algorithm and a nonlinear activation function are added; these speed up the training; the number of filters is 16, 32, 64, 128, and 128; (2) Upsampling is performed by bilinear interpolation with a kernel size of 3×3; the number of filters is 128, 128, 64, 32, and 16; (3) The skip connection is formed by copying the features of the left encoder path and adding them to the right decoder path, thus changing a small number of hyperparameters; the height and width of the input can be randomly selected as needed; noise-based regularization is used when fitting the network; that is, at each iteration, additive normal noise is used to perturb the input Z; this regularization can help better fit the network and achieve the desired goal.

5. The method for seismic data reconstruction based on convolutional neural network according to claim 1, characterized in that: In step 3, in an iterative process, the reconstructed data f is calculated by θ The mean square error MSE between (Z) and the observed seismic data M is used to obtain the local optimal parameters of the network.

6. The method for seismic data reconstruction based on convolutional neural network according to claim 5, characterized in that: In step 3, in the seismic data reconstruction problem, M∈R m×n represents the observed undersampled seismic data, i.e., the damaged missing seismic traces, X∈R m×n Represents the original complete seismic data; for the reconstruction task, the goal is to reconstruct X from M, only the missing seismic traces must be reconstructed by the network, while the non-missing seismic traces remain unchanged.

7. The method for seismic data reconstruction based on convolutional neural network according to claim 6, characterized in that: In step 3, using the CNN network, the loss function is determined as: Among them, f θ represents the designed CNN, with parameters θ, including network layers, the number of cores in each network layer, and the weight and bias of each core; Z represents the fixed input, i.e., random noise, which obeys some common distributions, including Gaussian or uniform distribution; P Ω : R m×n →R m×n represents the sampling operator, defined as where Ω represents the subset of indices corresponding to the observed entries; P Ω represents the projection onto the observed data while keeping the entries in Ω and setting zeros at the locations of missing traces; represents the Frobenius norm, defined as θ * represents the local optimal parameters of the network obtained by minimizing the loss function given in Equation 1; the sampling operator P Ω Copy to Output , making the entries in Ω identical to those in the observed seismic data M.

8. The method for seismic data reconstruction based on convolutional neural network according to claim 7, characterized in that: In step 3, during network training, f is generated from the input Z at each iteration. θ (Z), and P Ω Projection to f θ (Z); Calculate the loss based on the MSE criterion Then the network parameters θ are updated by the gradient descent method; finally the optimal parameters are obtained 9. The method for seismic data reconstruction based on convolutional neural network according to claim 8, characterized in that: In step 4, after saving the trained network parameters, the test data is reconstructed based on the trained network model, that is, by setting P Ω (X) = P Ω (M) Get 10. The method for seismic data reconstruction based on convolutional neural network according to claim 1, characterized in that: In step 5, the quality of the reconstruction is represented by the signal-to-noise ratio S / N, defined as: where X and X * represent the original seismic data and the reconstructed data, respectively.

Images