An unsupervised method for generating seismic data based on the Nash equilibrium principle

Abstract

The present invention discloses an unsupervised method for generating seismic data based on the Nash equilibrium principle. The generation network generates seismic data with classification features, that is, the generated seismic data contains reservoir features and gas-bearing features, which can greatly expand the training sample set and can directly use seismic data as training samples; the discrimination network can judge the consistency of the propagation law of seismic data and can directly check the processing quality of seismic data; since the generated data comes from real seismic data, it implies the physical propagation law of the target layer, which can indirectly constrain the learning process of deep learning so that it does not deviate from the physical propagation model.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing and fluid prediction, and in particular to a method for generating seismic data in an unsupervised manner based on the Nash equilibrium principle. Background Art

[0002] In the field of oil and gas exploration, certain progress has been made in reservoir prediction and gas content prediction based on deep learning methods. Artificial intelligence methods are used to carry out related research on reservoir prediction and gas content prediction. Due to the application characteristics of exploration and development, although the amount of seismic data in the work area is large, usually only the seismic data of the target layer will be selected for research and application. However, since the number of sample wells in the target layer is usually small, in the actual application of deep learning methods, the amount of seismic data samples matching the target layer is often insufficient. Therefore, it is difficult to directly use seismic data for training and learning.

[0003] At present, the most common practice is to first perform various linear transformations on the seismic data to obtain various seismic attributes, such as root mean square amplitude, minimum amplitude, longitudinal wave impedance, etc., and then use the various attributes obtained as learning samples, and use the seismic multiple attribute bodies to learn, and then find the reservoir characteristics and gas-bearing characteristics. There are also some problems with this method: (1) Seismic data is the only real data obtained during the acquisition process. Since seismic waves pass through the target layer, seismic data contains multiple characteristics. However, after multiple linear attribute solution operations, some characteristics, especially very weak gas-bearing characteristic information, may be weakened or eliminated. In addition, the indirect linear transformation method also introduces other uncertain factors. (2) Seismic data itself implicitly has strong physical laws, and the acquisition of attribute bodies is a mathematical linear operation. Using attribute bodies as samples may cause the entire learning process to deviate from physical laws. This implicit risk is difficult to control manually. When using deep learning methods to solve problems such as reservoir prediction or gas content prediction, it is more appropriate to directly use seismic data as training samples. However, due to the small number of known wells in the work area, the seismic data samples with clear classification characteristics corresponding to the known wells are usually insufficient. Therefore, it is very necessary to design a method to generate a large amount of seismic data with classification characteristics. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for generating seismic data in an unsupervised manner based on the Nash equilibrium principle in order to address the problem of insufficient seismic data samples with clear classification characteristics in the prior art.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] An unsupervised method for generating seismic data based on the Nash equilibrium principle, which interconnects the generation network and the discrimination network, then trains the generation network and the discrimination network, finally reaches the state of Nash equilibrium, and uses the trained generation network to generate seismic data with similar oil and gas laws in the target layer of real seismic data; among them, the training of the generation network and the discrimination network includes:

[0007] S1 Use the np function library to construct multiple randomly generated one-dimensional Gaussian distribution noises;

[0008] S2 Regularize the random noise data, divide it into multiple groups of data, and input it into the generation network in the form of two-dimensional data;

[0009] S3 The generation network linearly processes the input noise data and outputs a high-dimensional generated data through the two-dimensional deconvolution method;

[0010] S4 Use the generated data as the input of the discrimination network, and the discrimination network outputs the probability value of the discrimination flag. Based on the cross-entropy of the real seismic data and the generated data, it is used as the objective function H:

[0011] H((x1, y1), D) = -y1 log D(x1) - (1 - y1) log(1 - D(x1))

[0012] x1 is the generated data, y1 is the real seismic data, and D is the discrimination network.

[0013] An unsupervised method for generating seismic data based on the Nash equilibrium principle. The generation network generates seismic data with classification features, that is, the generated seismic data contains reservoir features and gas-bearing features, which can greatly expand the training sample set and can directly use seismic data as training samples; the discrimination network can judge the consistency of the seismic data propagation law and can directly check the quality of seismic data processing; since the generated data comes from real seismic data, it implies the physical propagation law of the target layer, which can indirectly constrain the learning process of deep learning and prevent it from deviating from the physical propagation model.

[0014] Preferably, the method for obtaining the real seismic data: input all seismic data within the specified block range, interpret the seismic horizons, and perform preprocessing and extraction on the seismic data of each layer according to the interpreted horizons.

[0015] Preferably, in step S4, the discrimination network is used to guide the iteration and correction of the seismic generation network model according to the distribution difference between the real seismic data and the generated data.

[0016] Preferably, when the output value of the discriminator in step S4 is between 0.4 and 0.5, the training ends.

[0017] Preferably, the generation network and the identification network structures adopt similar symmetrical network structures, the generation network adopts a 7-layer two-dimensional deconvolutional CNN neural network with different scales, and the identification network adopts a 7-layer two-dimensional convolutional CNN neural network with different scales.

[0018] Preferably, the step S4 further comprises: training and optimizing the identification network, wherein the loss function d_loss of the identification network is:

[0019] d_loss=d_loss_real+d_loss_fake

[0020] d_loss_real is the cross entropy obtained by inputting real earthquake data into the identification network, and d_loss_fake represents the cross entropy value of the generated data through the identification network.

[0021] Preferably, the step S4 further comprises: training the generated network to generate a loss function g_loss of the generated network:

[0022] g_loss=g_loss_fake

[0023] g_loss_fake is the cross entropy of the generated data passing through the generated network.

[0024] Preferably, during the training process, the Adam method is used to optimize the identification network and the generation network respectively.

[0025] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0026] A method for unsupervised generation of seismic data based on the Nash equilibrium principle. The generative network generates seismic data with classification characteristics, that is, the generated seismic data contains reservoir characteristics and gas-bearing characteristics, which can greatly expand the training sample set and enable the direct use of seismic data as training samples. The discriminative network can judge the consistency of the propagation laws of seismic data and directly check the quality of seismic data processing. Since the generated data comes from real seismic data, it implies the physical propagation laws of the target layer, which can indirectly constrain the learning process of deep learning so that it does not deviate from the physical propagation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the unsupervised seismic data generation method based on Nash equilibrium.

[0028] Figure 2 Flowchart of the unsupervised seismic data generation method based on Nash equilibrium.

[0029] Figure 3 is an error convergence curve of the iteration and correction process according to this embodiment. Detailed implementation mode

[0030] The present invention will be described in detail below with reference to the accompanying drawings.

[0031] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0032] Based on the principle of Nash equilibrium, an unsupervised method directly uses existing seismic data for unsupervised learning, such as Figure 2 shown, including:

[0033] Step 101: Input all seismic data within the specified block range, interpret the horizons of the seismic data according to the interpretation process, preprocess and extract the seismic data for each layer according to the interpreted horizons, and use it as the training sample of the seismic data model discrimination network;

[0034] Step 102: Use the real seismic data to perform data chunking according to horizons, use it as the input of the discrimination network, train the discrimination network to obtain the initial discrimination feature model of the seismic data, and continuously improve the seismic data discrimination model;

[0035] Step 103: Use white noise as the input of the seismic data generation network, and obtain the generated seismic data through the seismic data generation network, which is used as the initial feature model of the seismic data generation model;

[0036] Step 104: Then use the generated seismic data as the input of the discrimination model, and use the discrimination network to guide the seismic generation network model to iterate and correct according to the distribution difference between the real seismic data and the generated seismic data. The error convergence curve of this embodiment is as Figure 3 shown;

[0037] Using the principle of mutual game between the seismic data generation network and the discrimination network, finally reach the state of Nash equilibrium, that is, the discrimination network cannot distinguish the difference between the generated data and the real data;

[0038] Step 105: Use the finally obtained seismic data generation model to generate the seismic data of the target layer. The finally generated data has a similar law to the real seismic data and can be used as a sample for deep learning applications.

[0039] Among them, the entire network is as Figure 1 shown, and it is composed of a seismic data generation network and a seismic data discrimination network. The seismic data generation network is used to generate seismic data, and the discrimination network is used to discriminate the generated seismic data. The main content includes:

[0040] First, regularize the seismic data, group it according to a two-dimensional arrangement, and output a one-dimensional data structure through a 7-layer CNN convolutional neural network with different scales;

[0041] The seismic data generation network mainly uses a 7-layer two-dimensional transposed convolution CNN neural network with different scales, and the output is a vector set representing the generated seismic data;

[0042] The seismic data discrimination network mainly uses a 7-layer two-dimensional convolutional CNN neural network with different scales, and outputs a;

[0043] By interconnecting the two networks, the generation network and the discrimination network structures adopt similar symmetric network structures. The specific training calculations are as follows:

[0044] Use the np library to construct multiple random one-dimensional Gaussian distribution noises;

[0045] Regularize the random noise data, divide it into multiple groups of data, and input it into the generation network in the form of two-dimensional data;

[0046] The generation network linearly processes the input noise data and outputs a high-dimensional data through the two-dimensional transposed convolution method;

[0047] Use the high-dimensional data output by the generation network as the input of the discrimination network. The discrimination network passes through a 7-layer convolutional network. At this time, the problem can be regarded as a binary classification problem, and the cross-entropy between the real seismic data and the generated data is used as the objective function;

[0048] When the distribution of the generated seismic data is exactly the same as the distribution of the real seismic data, the probability value of the discrimination becomes 1.0. During the training process, the cross-entropy is mainly used as the loss function, as shown in Equation 1

[0049] H((x1, y1), D) = -y1 log D(x1) - (1 - y1) log(1 - D(x1)) (Equation 1)

[0050] x1 and y1 represent the generated data and the real data respectively, D represents the discrimination network, and H represents the cross-entropy. First, train and optimize the discrimination network. The loss function used at this time is as shown in Equation 2,

[0051] d_loss = d_loss_real + d_loss_fake (Equation 2)

[0052] At this time, d_loss is the sum of the losses of the generated data and the real data output by the discriminative network. d_loss_real is the cross-entropy obtained by inputting the real seismic data into the discriminative network, and d_loss_fake represents the value of the cross-entropy of the generated data passing through the discriminative network. Then, the generative network is trained, and its loss function is shown in Equation 3.

[0053] g_loss = g_loss_fake (Equation 3)

[0054] g_loss is the cross-entropy of the generated data passing through the generative network. During the optimization process, the generative network and the discriminative network are independently optimized using the adam method respectively.

[0055] The seismic data generation network and the seismic data discriminative deep learning network are constructed based on the convolutional neural network framework TensorFlow and use the Python programming language for network control.

[0056] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An unsupervised method for generating seismic data based on the Nash equilibrium principle, characterized in that, By interconnecting the generation network and the discrimination network, then training the generation network and the discrimination network, finally reaching the state of Nash equilibrium, and using the trained generation network to generate seismic data with similar oil and gas laws in the target layer of real seismic data; among them, the training of the generation network and the discrimination network includes: S1 Use the np function library to construct multiple random one-dimensional Gaussian distribution noises; S2 Regularize the random noise data, divide it into multiple groups of data, and input it into the generation network in the form of two-dimensional data; S3 The generation network linearly processes the input noise data and outputs a high-dimensional generated data through the two-dimensional transposed convolution method; S4 Use the generated data as the input of the discrimination network. The discrimination network outputs the probability value of the label bit resolution. Based on the cross-entropy of the real seismic data and the generated data as the objective function H: H((x1, y1), D) = -y1 log D(x1) - (1 - y1) log(1 - D(x1)) x1 is the generated data, y1 is the real seismic data, and D is the discrimination network; The acquisition method of the real seismic data: Input all seismic data within the specified block range, interpret the horizons of the seismic data, and perform preprocessing and extraction on the seismic data of each layer according to the interpreted horizons.

2. The method according to claim 1, characterized in that, In step S4, the discrimination network is used to guide the seismic generation network model to iterate and correct according to the distribution difference between the real seismic data and the generated data.

3. The method according to claim 2, wherein When the output value of the discriminator in step S4 is between 0.4 and 0.5, the training ends.

4. The method according to claim 3, characterized in that The generation network and the discrimination network structures adopt similar symmetric network structures. The generation network adopts a 7-layer two-dimensional transposed convolution CNN neural network with different scales, and the discrimination network adopts a 7-layer two-dimensional convolution CNN neural network with different scales.

5. The method according to claim 1, characterized in that, Optimize the discrimination network, and the loss function of the discrimination network is d_loss: d_loss = d_loss_real + d_loss_fake d_loss_real is the cross-entropy obtained by inputting the real seismic data into the discrimination network, and d_loss_fake represents the value of the cross-entropy of the generated data passing through the discrimination network.

6. The method according to claim 5, characterized in that Step S4 also includes: training the generation network, and the loss function of the generation network is g_loss: g_loss = g_loss_fake g_loss_fake is the cross-entropy of the generated data passing through the generation network.

7. The method according to claim 6, characterized in that Including: During the training process, the adam method is used to optimize the discrimination network and the generation network respectively.

Images