An intelligent seismic inversion prediction method based on deep learning
By constructing a multi-scale ultra-asymmetric Cycle-JNET network model and using deep learning to process well-seismic data, the problem of difficulty in identifying thin sand bodies in seismic inversion was solved, and high-resolution seismic inversion and reservoir prediction were achieved.
Patent Information
- Application Number
- CN202311533290.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-17
- Publication Date
- 2026-05-15
- Estimated Expiration
- 2043-11-17
AI Technical Summary
Existing seismic inversion techniques cannot effectively identify thin sand bodies, and in complex geological conditions, the inversion accuracy is low, the resolution is poor, and it is difficult to accurately identify reservoirs.
A multi-scale ultra-asymmetric Cycle-JNET network model is constructed. Well-seismic data is processed using deep learning methods to extract features at different scales and generate high-resolution seismic inversion volumes, thereby realizing the learning of the correspondence between well-seismic data.
It improves the resolution and accuracy of seismic inversion, enabling accurate identification and prediction of thin sand layers, and guiding reservoir identification and well placement.
Smart Images
Figure CN117434607B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an intelligent earthquake inversion and prediction method based on deep learning, belonging to the field of earthquake inversion technology. Background Technology
[0002] Seismic inversion, a core technology for reservoir prediction, is essentially a method of transforming geophysical data into subsurface lithological data through data processing. The inversion results include rich stratigraphic information, porosity, and acoustic impedance, used for reservoir description. Acoustic impedance inversion theory posits that different lithofacies correspond to different acoustic impedances, a prerequisite for inversion. After years of development and application, seismic inversion has evolved into various methods for reservoir description. Existing seismic inversion theories are mainly based on wave equations and convolution models. The former, due to its complexity in practical applications, has not been widely adopted. Therefore, the most commonly used method is convolution model-based seismic inversion. This theory posits that seismic data is the result of the convolution of reflection coefficients and seismic wavelets. Reflection coefficients can be calculated from acoustic impedance. The process of obtaining seismic data through the convolution of seismic wavelets and reflection coefficients is defined as forward modeling, and the reverse is called inversion. In reality, seismic data often contains significant noise, which affects seismic resolution and increases the ambiguity of the inversion results.
[0003] Therefore, there is currently no seismic inversion technology that fully considers uncertainty and complexity, nor can it have high resolution and identify thin sand bodies.
[0004] Traditional seismic inversion typically involves two steps: first, establishing a target equation based on various assumptions about the inversion equations; and then adding constraints to solve this equation. However, actual geological conditions and data are often highly complex and difficult to describe in a single, definitive way. Significant challenges exist in both establishing and solving inversion equations. Actual inversion accuracy is low, resolution is poor, and the prediction and identification of sand bodies largely depend on well curve data. When well spacing is large or 3D data such as seismic data is complex and of poor quality, accurate reservoir prediction is difficult. Summary of the Invention
[0005] The main objective of this invention is to construct a multi-scale ultra-asymmetric network (Cycle-JNET). The problem to be solved is how to identify complex seismic data and accurately identify and predict thin sand layers, fully considering the correspondence between well-seismic data, and generating various realistic seismic inversion models based on the seismic data. Ultimately, a high-resolution seismic inversion model consistent with the seismic data is established, achieving accurate identification and prediction of sand bodies and interlayers, thereby improving the inversion resolution and accuracy, and making it more suitable for practical application.
[0006] To achieve the above objectives, this invention provides a deep learning-based intelligent seismic inversion prediction method, the method comprising the following steps:
[0007] S1. Construct a deep convolutional multi-attribute multi-scale ultra-asymmetric Cycle-JNET network model;
[0008] S2. Obtain the dataset, i.e., various types of well seismic data;
[0009] S3. Process the well-seismic data volume. According to the training requirements, divide the well curve data and seismic data in the well-seismic data volume into the required 1*1*128 size respectively. Each well curve data has corresponding seismic data.
[0010] Step S3 specifically includes the following sub-steps:
[0011] S301. Based on the diversity of well curve data and the uniqueness of seismic data, multiple well curve data on a seismic dataset can be trained simultaneously. Therefore, during data processing, the attributes of the multiple well curves to be learned are bound to the corresponding seismic data, and invalid and abnormal data in the well curves are removed to maintain the quality of the dataset.
[0012] S302. Create an empty data volume, divide the empty data volume according to the number of attributes to be trained, and write all the data required in the data volume into the data volume. The size of each well curve data and seismic data is 1*1*128.
[0013] S4. Perform wavelet transform on the model data in the training dataset to obtain well seismic data at different scales. Each well curve data has corresponding seismic data.
[0014] Step S4 specifically includes the following sub-steps:
[0015] S401. Based on the one-dimensionality of well curve data and seismic data, wavelet transform can obtain data with two resolutions. When performing wavelet transform on the data volume, each type of well curve data is transformed separately to obtain two resolutions: low resolution aaa and high resolution daa.
[0016] S402. Perform wavelet transform on the corresponding seismic data to obtain two resolutions: low-resolution aaa and high-resolution daa.
[0017] S5. Input the well seismic data at different scales into the multi-attribute, multi-scale, ultra-asymmetric Cycle-JNET network model for training to obtain the trained JNET model.
[0018] Step S5 specifically includes the following sub-steps:
[0019] The S501 and Cycle-JNET networks are designed as two parallel aaa and daa networks, respectively learning features at different resolutions. Taking the aaa network as an example, the real low-frequency part Segy(R-aaa) after wavelet transform of real seismic data is input into the JNET(S1-aaa) network for convolution, and the network generates the corresponding spurious well curve Well(F-aaa). At the same time, the daa network generates the corresponding spurious well curve Well(F-daa).
[0020] S502, Well(F-aaa), Well(F-daa) are compared with the actual well curves Well(R-aaa) and Well(R-daa) after wavelet decomposition. The residuals obtained from the two aaa networks and the daa network are linearly added to obtain the residual Loss1(S1) of the JNET(S1) network.
[0021] S503. Taking the aaa network as an example, the output well curve Well(F-aaa) is input into JNET(S2) to convolve the well curve Well(F-aaa) and generate seismic data Segy(F-aaa). At the same time, the ddd network generates the corresponding seismic data Segy(F-aaa).
[0022] S504. Perform inverse wavelet transform on Segy(F-aaa) and Segy(F-daa) to obtain the seismic data Segy(fake) after this round of machine learning. Compare it with the real seismic data Segy(real) to obtain the residual Loss2(S2) of the JNET(S2) network for this round.
[0023] S505. Linearly add Loss1(S1) and Loss2(S2) to obtain the total residual Loss(S). Use Loss(S) to optimize JNET(S1) and JNET(S2), and then let the machine enter the next round of training.
[0024] The multi-attribute, multi-scale, super-asymmetric Cycle-JNET network was constructed, generating a network with 12 deconvolutional layers. The kernel size of the first layer is 4, stride is 2, and width is 1; the kernel size of the second layer is 4, stride is 2, and width is 1; the kernel size of the third layer is 4, stride is 2, and width is 1; the kernel size of the fourth layer is 4, stride is 2, and width is 1; the kernel size of the fifth layer is 4, stride is 2, and width is 1; the kernel size of the sixth layer is 4, stride is 2, and width is 1; the kernel size of the seventh layer is 4, stride is 2, and width is 1; the kernel size of the eighth layer is 4, stride is 2, and width is 1; the ninth layer is... The kernel size is 4, stride is 2, and width is 1. The tenth convolutional layer has a kernel size of 4, stride is 2, and width is 1. The eleventh convolutional layer has a kernel size of 4, stride is 2, and width is 1. The twelfth convolutional layer has a kernel size of 4, stride is 2, and width is 1. The initial learning rate of the generator network is 0.0001. The optimization function is Adam, the loss function is BCEloss, and the activation functions are ReLU, ReLU, ReLU, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid.
[0025] The multi-attribute, multi-scale, super-asymmetric Cycle-JNET network was constructed. The discriminator network has a total of 7 convolutional layers. The first convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The second, third, fourth, fifth, sixth, and seventh convolutional layers all have kernel sizes of 4, strides of 2, and widths of 1. The initial learning rate of the generator network is 0.0001. The optimization function is Adam, the loss function is BCEloss, and the activation functions are LeakyReLU, LeakyReLU, LeakyReLU, LeakyReLU, Sigmoid, Sigmoid, Sigmoid.
[0026] The DCGAN network was built with a batch size of 8 and 250 epochs.
[0027] S6. Based on the obtained JNET model, construct the seismic inversion body by writing it channel by channel.
[0028] Compared with the prior art, the technical effects of the present invention are as follows:
[0029] 1. The seismic inversion prediction method proposed in this invention is a deep learning-based approach that achieves high-accuracy formation identification and prediction. During seismic inversion, the density of well data affects the accuracy of the inversion. By employing wavelet transform and deep learning, well-seismic data can be processed rapidly and in large quantities, extracting features at different scales, learning the correspondence between well-seismic data, and generating a highly consistent seismic inversion model based on seismic data, which can be used for reservoir identification and prediction.
[0030] 2. The seismic inversion prediction method proposed in this invention greatly improves the resolution of inversion prediction by improving the network structure, and realizes the identification and prediction of thin sand layers. In practical work, it can lay the foundation for the development of the work area and well placement. Attached Figure Description
[0031] Figure 1 This is a flowchart of an intelligent earthquake inversion and prediction method based on deep learning according to the present invention.
[0032] Figure 2 This is a schematic diagram illustrating how the present invention processes data in deep learning;
[0033] Figure 3 This is a single-well curve analysis diagram during the first training of this invention;
[0034] Figure 4 This is the single-well curve analysis diagram of the present invention during the 250th training iteration. Detailed Implementation
[0035] The implementation steps of the present invention will be further described in detail below with reference to the accompanying drawings.
[0036] This invention discloses a deep learning-based intelligent seismic inversion prediction method, which can improve the accuracy and resolution of seismic inversion, enhance the accuracy of reservoir identification and prediction, and guide the development of work areas and well placement.
[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0038] This invention provides a deep learning-based intelligent earthquake inversion and prediction method. Figure 1 This is a flowchart of an intelligent earthquake inversion and prediction method based on deep learning, according to the present invention. The method includes the following steps:
[0039] S1. Construct a deep convolutional multi-attribute multi-scale ultra-asymmetric Cycle-JNET network model;
[0040] S2. Obtain the dataset, i.e., various types of well seismic data;
[0041] S3. Process the well-seismic data volume. According to the training requirements, divide the well curve data and seismic data in the well-seismic data volume into the required 1*1*128 size respectively. Each well curve data has corresponding seismic data.
[0042] Furthermore, in step S3, such as Figure 2 The following are the specific steps for processing well seismic data:
[0043] S301. Based on the diversity of well curve data and the uniqueness of seismic data, multiple well curve data from a single seismic dataset can be trained simultaneously. Therefore, during data processing, the attributes of the various well curves to be learned are bound to the corresponding seismic data, and invalid and outlier data in the well curves are removed to maintain the quality of the dataset; such as... Figure 3 and Figure 4 As shown;
[0044] S302. Create an empty data volume, divide the empty data volume according to the number of attributes to be trained, and write all the data required in the data volume into the data volume. The size of each well curve data and seismic data is 1*1*128.
[0045] S4. Perform wavelet transform on the model data in the training dataset to obtain well seismic data at different scales. Each well curve data has corresponding seismic data.
[0046] Furthermore, in step S4, the specific steps for performing wavelet transform on the model data in the training dataset are as follows:
[0047] S401. Based on the one-dimensionality of well curve data and seismic data, wavelet transform can obtain data with two resolutions. When performing wavelet transform on the data volume, each type of well curve data is transformed separately to obtain two resolutions: low resolution aaa and high resolution daa.
[0048] S402. Perform wavelet transform on the corresponding seismic data to obtain two resolutions: low-resolution aaa and high-resolution daa.
[0049] S5. Input the well seismic data at different scales into the multi-attribute, multi-scale, ultra-asymmetric Cycle-JNET network model for training to obtain the trained JNET model.
[0050] Furthermore, in step S5, the specific steps for performing wavelet transform on the model data in the training dataset are as follows:
[0051] The S501 and Cycle-JNET networks are designed as two parallel aaa and daa networks, which learn features at different resolutions respectively. Taking the aaa network as an example, the real low-frequency part Segy(R-aaa) of the real seismic data Segy(real) after wavelet transform is input into the JNET(S1-aaa) network for convolution. The network generates the corresponding fake well curve Well(F-aaa), and at the same time, the daa network generates the corresponding fake well curve Well(F-daa).
[0052] S502, Well(F-aaa), Well(F-daa) are compared with the actual well curves Well(R-aaa) and Well(R-daa) after wavelet decomposition. The residuals obtained from the two aaa networks and the daa network are linearly added to obtain the residual Loss1(S1) of the JNET(S1) network.
[0053] S503. Taking the aaa network as an example, the output well curve Well(F-aaa) is input into JNET(S2) to convolve the well curve Well(F-aaa) and generate seismic data Segy(F-aaa). At the same time, the ddd network generates the corresponding seismic data Segy(F-aaa).
[0054] S504. Perform inverse wavelet transform on Segy(F-aaa) and Segy(F-daa) to obtain the false earthquake data Segy(fake) after this round of machine learning. Compare it with the real earthquake data Segy(real) to obtain the residual Loss2(S2) of the JNET(S2) network in this round.
[0055] S505. Linearly add Loss1(S1) and Loss2(S2) to obtain the total residual Loss(S). Use Loss(S) to optimize JNET(S1) and JNET(S2), and then let the machine enter the next round of training and learning.
[0056] In step S5, the multi-attribute, multi-scale, super-asymmetric Cycle-JNET network is constructed. The generated network has a total of 12 deconvolutional layers. The first layer has a kernel size of 4, a stride of 2, and a width of 1. The second layer has a kernel size of 4, a stride of 2, and a width of 1. The third layer has a kernel size of 4, a stride of 2, and a width of 1. The fourth layer has a kernel size of 4, a stride of 2, and a width of 1. The fifth layer has a kernel size of 4, a stride of 2, and a width of 1. The sixth layer has a kernel size of 4, a stride of 2, and a width of 1. The seventh layer has a kernel size of 4, a stride of 2, and a width of 1. The eighth layer has a kernel size of 4, a stride of 2, and a width of 1. The ninth layer has a kernel size of 4, a stride of 2, and a width of 1. The kernel size of the first convolutional layer is 4, the stride is 2, and the width is 1. The kernel size of the tenth convolutional layer is 4, the stride is 2, and the width is 1. The kernel size of the eleventh convolutional layer is 4, the stride is 2, and the width is 1. The kernel size of the twelfth convolutional layer is 4, the stride is 2, and the width is 1. The initial learning rate of the generator network is 0.0001. The optimization function is Adam, the loss function is BCEloss, and the activation functions are ReLU, ReLU, ReLU, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid.
[0057] In step S5, a multi-attribute, multi-scale, super-asymmetric Cycle-JNET network is constructed. The discriminator network has a total of 7 convolutional layers. The first convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The second convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The third convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The fourth convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The fifth convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The sixth convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The seventh convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The initial learning rate of the generator network is 0.0001. The optimization function is Adam, the loss function is BCEloss, and the activation functions are LeakyReLU, LeakyReLU, LeakyReLU, LeakyReLU, Sigmoid, Sigmoid, Sigmoid.
[0058] The DCGAN network was built with a batch size of 8 and 250 epochs.
[0059] S6. Based on the obtained JNET model, construct the seismic inversion body by writing it channel by channel.
Claims
1. A deep learning-based intelligent seismic inversion prediction method, characterized in that, Includes the following steps: S1. Construct a deep convolutional multi-attribute multi-scale ultra-asymmetric Cycle-JNET network model; S2. Obtain the dataset, i.e., various types of well seismic data; S3. Process the well-seismic data volume. According to the training requirements, divide the well curve data and seismic data in the well-seismic data volume into the required 1*1*128 size respectively. Each well curve data has corresponding seismic data. S4. Perform wavelet transform on the model data in the training dataset to obtain well seismic data at different scales. Each well curve data has corresponding seismic data. S5. Input the well seismic data at different scales into the multi-attribute, multi-scale, ultra-asymmetric Cycle-JNET network model for training to obtain the trained JNET model. The S501 and Cycle-JNET networks are designed as two parallel aaa and daa networks, respectively learning features at different resolutions. In the aaa network, the real low-frequency part Segy (R-aaa) of the real seismic data after wavelet transform is input into the JNET (S1-aaa) network for convolution, and the network generates the corresponding spurious well curve Well (F-aaa). At the same time, the daa network generates the corresponding spurious well curve Well (F-daa). S502, Well(F-aaa), Well(F-daa) are compared with the actual well curves Well(R-aaa) and Well(R-daa) after wavelet decomposition. The residuals obtained from the two aaa networks and the daa network are linearly added to obtain the residual Loss1(S1) of the JNET(S1) network. S503. In the aaa network, the output well curve Well (F-aaa) is input into JNET (S2) to convolve the well curve Well (F-aaa) to generate seismic data Segy (F-aaa). At the same time, the daa network generates the corresponding seismic data Segy (F-daa). S504. Perform inverse wavelet transform (IWT) on Segy(F-aaa) and Segy(F-daa) to obtain the seismic data Segy(fake) after this round of machine learning. Compare it with the real seismic data Segy(real) to obtain the residual Loss2(S2) of the JNET(S2) network. S505. Linearly add Loss1 (S1) and Loss2 (S2) to obtain the total residual Loss (S). Use Loss (S) to optimize JNET (S1) and JNET (S2), and then let the machine enter the next round of training and learning. S6. Based on the obtained JNET model, construct the seismic inversion body by writing it channel by channel.
2. The intelligent seismic inversion prediction method based on deep learning according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S301. Based on the diversity of well curve data and the uniqueness of seismic data, multiple well curve data on a seismic data segment are trained. Therefore, during data processing, the attributes of the multiple well curves to be learned are bound with the corresponding seismic data, and invalid and abnormal data in the well curves are removed to maintain the quality of the dataset. S302. Create an empty data volume, divide the empty data volume according to the number of attributes to be trained, and write all the data required in the data volume into the data volume. The size of each well curve data and seismic data is 1*1*128.
3. The intelligent seismic inversion prediction method based on deep learning according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S401. Based on the one-dimensionality of well curve data and seismic data, wavelet transform is used to obtain data with two resolutions. When performing wavelet transform on the data volume, each type of well curve data is transformed separately to obtain low-resolution aaa and high-resolution daa data. S402. Perform wavelet transform on the corresponding seismic data to obtain two resolutions: low-resolution aaa and high-resolution daa.
4. The intelligent seismic inversion prediction method based on deep learning according to claim 1, characterized in that, The multi-attribute, multi-scale, super-asymmetric Cycle-JNET network was constructed, generating a network with a total of 12 deconvolutional layers: the first layer has a kernel size of 4, a stride of 2, and a width of 1; the second layer has a kernel size of 4, a stride of 2, and a width of 1; the third layer has a kernel size of 4, a stride of 2, and a width of 1; the fourth layer has a kernel size of 4, a stride of 2, and a width of 1; the fifth layer has a kernel size of 4, a stride of 2, and a width of 1; the sixth layer has a kernel size of 4, a stride of 2, and a width of 1; the seventh layer has a kernel size of 4, a stride of 2, and a width of 1; the eighth layer has a kernel size of 4, a stride of 2, and a width of 1; the ninth ... The kernel size is 4, stride is 2, and width is 1. The tenth convolutional layer has a kernel size of 4, stride is 2, and width is 1. The eleventh convolutional layer has a kernel size of 4, stride is 2, and width is 1. The twelfth convolutional layer has a kernel size of 4, stride is 2, and width is 1. The initial learning rate of the generator network is 0.0001. The optimization function is Adam, the loss function is BCEloss, and the activation functions are ReLU, ReLU, ReLU, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid, Sigmoid.
5. The intelligent seismic inversion prediction method based on deep learning according to claim 4, characterized in that, In step S5, a multi-attribute, multi-scale, super-asymmetric Cycle-JNET network is constructed. The discriminator network has a total of 7 convolutional layers. The first convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The second convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The third convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The fourth convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The fifth convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The sixth convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The seventh convolutional layer has a kernel size of 4, a stride of 2, and a width of 1. The initial learning rate of the generator network is 0.0001. The optimization function is Adam, the loss function is BCEloss, and the activation functions are LeakyReLU, LeakyReLU, LeakyReLU, LeakyReLU, Sigmoid, Sigmoid, Sigmoid.