Seismic wave impedance inversion method under deep learning framework, prediction model, equipment and medium
Through a deep learning framework combining LSTM and CNN to extract the wideband wave impedance characteristics in seismic data, and using deconvolution and fully connected neural networks for scale matching and nonlinear mapping, the problem of inaccurate seismic wave impedance inversion results in the prior art is solved, and efficient inversion of wideband seismic wave impedance is achieved.
Patent Information
- Application Number
- CN202311673268.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-07
- Publication Date
- 2025-06-10
AI Technical Summary
The existing seismic wave impedance inversion methods are greatly affected by the initial model, making it difficult to break out of the local optimal solution, and the convergence speed is slow, making it difficult to obtain accurate seismic wave impedance inversion results.
A deep learning framework is adopted, combining long and short-term memory recurrent network (LSTM) and convolutional neural network (CNN) to extract low-frequency and high-frequency wave impedance characteristics in seismic data, and scale matching and nonlinear mapping are carried out through deconvolution networks and fully connected neural networks to build a wide-band seismic wave impedance inversion model.
The wave impedance characteristics of the wide band are effectively extracted, which solves the scale differences between earthquake and logging data, improves the accuracy and efficiency of the inversion results, and realizes the precise inversion of the seismic wave impedance of the wide band.
Smart Images
Figure CN120122151A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of oil and gas geophysical exploration, and particularly to a broadband seismic wave impedance inversion method, a prediction model, an electronic device, and a storage medium under a deep learning framework. Background Art
[0002] Seismic wave impedance is one of the important bridges connecting well logging and seismic data. By accurately obtaining the seismic wave impedance, the interface-type seismic reflection data can be converted into formation-type wave impedance data, which is convenient for comparing and analyzing seismic data with well logging data, and further clarifying the lithology, physical properties and other characteristics of the reservoir. Therefore, accurately obtaining seismic wave impedance information is of great significance for the accurate prediction of the reservoir.
[0003] The existing seismic wave impedance inversion methods mainly include the following several types:
[0004] (1) Direct wave impedance inversion based on the convolution model: mainly includes trace integration inversion, recursive inversion, etc. This type of method is developed based on deconvolution. Limited by the band-limited characteristics of seismic data, the inverted wave impedance is a relative wave impedance, and the absolute wave impedance can be obtained only by adding low-frequency components.
[0005] (2) Seismic wave impedance inversion based on the model: includes generalized linear inversion, well logging constrained inversion, etc. This type of method establishes a geological model through well logging and geological data, and continuously corrects the model through iteration to reduce the multi-solution of the model. Although this type of method can obtain a relatively accurate seismic wave impedance body, it is greatly affected by the initial model and is prone to falling into local optimal solutions.
[0006] (3) Wave impedance inversion based on the optimization algorithm: includes simulated annealing algorithm, genetic algorithm, particle swarm algorithm, etc. This type of method overcomes the influence of the initial model and can effectively jump out of the local optimal solution by using the global optimization ability of the algorithm to achieve global optimization. However, this type of algorithm performs random search in the model space, resulting in a slow convergence speed. Summary of the Invention
[0007] Seismic wave impedance is one of the important bridges connecting well logging and seismic data. By accurately obtaining the seismic wave impedance, the interface-type seismic reflection data can be converted into formation-type wave impedance data, which is convenient for comparing and analyzing seismic data with well logging data, and further clarifying the lithology, physical properties and other characteristics of the reservoir. Therefore, accurately obtaining seismic wave impedance information is of great significance for the accurate prediction of the reservoir.
[0008] To achieve the above object, the present invention provides a seismic wave impedance prediction model under a deep learning framework, including:
[0009] A low-frequency feature extraction module and a high-frequency feature extraction module are respectively used to extract the low-frequency features and high-frequency features of the wave impedance from the input data;
[0010] A scale matching module is used to fuse the extracted high-frequency features and low-frequency features to obtain wave impedance features at the same scale;
[0011] A non-linear mapping module is used to obtain the non-linear mapping relationship between the extracted wave impedance features and the well-log wave impedance.
[0012] Specifically, the low-frequency feature extraction module is based on a long short-term memory recurrent network and includes three LSTM layers.
[0013] Specifically, the high-frequency feature extraction module is based on a convolutional neural network and includes three convolutional layers with different convolutional kernel sizes.
[0014] Specifically, the scale matching module is based on a deconvolutional neural network and includes two deconvolutional layers.
[0015] Specifically, the non-linear mapping module is based on a fully connected neural network and consists of a fully connected layer with 2 hidden layers.
[0016] According to another aspect of the present invention, the present invention provides a seismic wave impedance inversion method under a deep learning framework, including:
[0017] Construct a training data set, and the training data set includes the relationship between well-log data and seismic data;
[0018] Construct a seismic wave impedance prediction model under a deep learning framework;
[0019] Use the training data set to train the seismic wave impedance prediction model and update the model parameters;
[0020] Based on the trained seismic wave impedance prediction model, realize wide-band seismic wave impedance inversion.
[0021] Furthermore, constructing the training data set includes:
[0022] Fine well-seismic calibration, extract seismic data near the well and well-log data, and construct a sample set;
[0023] Sample set division, divide the training data set and the validation data set according to a ratio close to 7:3.
[0024] Furthermore, updating the model parameters includes using the root mean square error of the training data set as an evaluation criterion, continuously updating the model parameters, and obtaining an optimal wave impedance inversion model.
[0025] According to another aspect of the present invention, the present invention provides an electronic device, and the electronic device includes:
[0026] A memory for storing executable instructions;
[0027] A processor that runs the executable instructions in the memory to implement the seismic wave impedance inversion method under the deep learning framework described above.
[0028] According to another aspect of the present invention, the present invention provides a non-transitory computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the seismic wave impedance inversion method under the deep learning framework described above.
[0029] The present invention has the following innovative points compared with the existing technologies:
[0030] (1) The present invention extracts high-frequency features of wave impedance based on the LSTM network and extracts low-frequency features of wave impedance by using the CNN network. By effectively combining the LSTM network and the CNN network, wide-band wave impedance features can be extracted.
[0031] (2) The present invention uses the deconvolution network to effectively solve the scale difference between seismic and logging data.
[0032] (3) The present invention constructs a technical process for wave impedance inversion under the deep learning framework, which can simultaneously extract low-frequency and high-frequency features of wave impedance and effectively realize wide-band seismic wave impedance inversion. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] By describing the exemplary embodiments of the present invention in more detail in conjunction with the accompanying drawings, the above and other objects, features and advantages of the present invention will become more obvious. Among them, in the exemplary embodiments of the present invention, the same reference numerals generally represent the same components.
[0034] Figure 1 Schematic diagram of a seismic wave impedance inversion model according to an embodiment of the present invention.
[0035] Figure 2 Well-seismic calibration diagram according to an embodiment of the present invention.
[0036] Figure 3 Schematic diagram of the structure of a low-frequency feature extraction module according to an embodiment of the present invention.
[0037] Figure 4 Schematic diagram of the structure of a high-frequency feature extraction module according to an embodiment of the present invention.
[0038] Figure 5 Schematic diagram of a scale matching module according to an embodiment of the present invention.
[0039] Figure 6Schematic diagram of a non - linear mapping module according to an embodiment of the present invention.
[0040] Figure 7 Cross - well profile of wave impedance inversion according to an embodiment of the present invention.
[0041] Figure 8 Flowchart of a seismic wave impedance inversion method under a deep learning framework according to an embodiment of the present invention.
[0042] Figure 9 Flowchart of the method implementation according to an embodiment of the present invention. Detailed implementation manners
[0043] The preferred embodiments of the present invention will be described in more detail below. Although the preferred embodiments of the present invention are described below, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.
[0044] Traditional seismic wave impedance inversion methods are greatly affected by the initial model, and the seismic wave impedance inversion methods based on optimization algorithms have a slow convergence speed, and it is difficult to obtain a satisfactory seismic wave impedance inversion result. In recent years, machine learning and deep learning algorithms have been used to solve inverse problems and have achieved certain results. The present invention uses deep learning algorithms as technical means and invents a wide - band seismic wave impedance inversion method under a deep learning framework. This method extracts wave impedance information in different frequency bands through different deep learning models, so as to achieve a wide - band seismic wave impedance inversion result.
[0045] The present invention belongs to the field of oil and gas geophysical exploration, and discloses a wide - band seismic wave impedance inversion method under a deep learning framework. Conventional seismic wave impedance inversion is restricted by factors such as hypothesis conditions, initial models, and calculation efficiency. The frequency band of the inverted seismic wave impedance is limited, and it is difficult to obtain an accurate seismic wave impedance inversion result. In view of these problems, the present invention proposes a wide - band seismic wave impedance inversion method under a deep learning framework, which uses a method combining an LSTM network and a CNN network to extract wide - band seismic wave impedance information contained in seismic data, and at the same time matches the scale difference between seismic data and well - logging data based on a de - convolution network, so as to achieve a wide - band seismic wave impedance inversion.
[0046] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but it is not a limitation of the present invention. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.
[0047] Embodiment 1
[0048] As Figure 1 shown, this embodiment provides a seismic wave impedance prediction model under a deep learning framework, including:
[0049] A low-frequency feature extraction module and a high-frequency feature extraction module, which are respectively used to extract the low-frequency features and high-frequency features of wave impedance from the input data;
[0050] A scale matching module, which is used to fuse the extracted high-frequency features and low-frequency features to obtain wave impedance features at the same scale;
[0051] A non-linear mapping module, which is used to obtain the non-linear mapping relationship between the extracted wave impedance features and well logging wave impedance.
[0052] Specifically, the seismic wave impedance prediction model under the deep learning framework mainly includes a low-frequency feature extraction module based on a long short-term memory recurrent neural network (LSTM), a high-frequency feature extraction module based on a convolutional neural network (CNN), a scale matching module based on a deconvolutional neural network, and a non-linear mapping module based on a fully connected neural network (DNN).
[0053] Specifically, the low-frequency feature extraction module based on a long short-term memory recurrent neural network (LSTM). Seismic and well logging data are typical sequence signals. The LSTM network can explore the internal relationship between the front and back samples of the sequence signal, and at the same time overcome the problem that the traditional recurrent neural network (RNN) cannot effectively solve the long-term dependence relationship, and can effectively mine the internal relationship and mutual connection between seismic and well logging data. At the same time, using the characteristic that the deep network can capture more complex non-linear relationships and produce smoother outputs, the present invention designs a three-layer LSTM layer in this module. Therefore, it can be considered that this module can obtain the low-frequency features of wave impedance.
[0054] The low-frequency feature extraction module based on a long short-term memory recurrent neural network (LSTM) in this embodiment is composed of three LSTM layers, and each LSTM layer is as Figure 3 shown. The low-frequency feature extraction module uses the LSTM network to explore the internal relationship between the front and back samples of the sequence signal, so as to mine the low-frequency wave impedance features contained in the seismic data.
[0055] Specifically, the high-frequency feature extraction module based on a convolutional neural network (CNN). Different from the LSTM network, the CNN network has no state variable and has poor effect in processing low-frequency information, but it can capture high-frequency components through a small convolutional kernel. Therefore, the present invention uses the CNN network to extract the high-frequency features of wave impedance.
[0056] The high-frequency feature extraction module based on the convolutional neural network (CNN) in this embodiment is different from the LSTM network. It has no state variables and performs poorly in processing low-frequency information. However, it can capture high-frequency components through small convolutional kernels. Therefore, the present invention uses the CNN network to extract the high-frequency features of wave impedance. To effectively extract the wave impedance information in seismic data, three convolutional layers with different convolutional kernel sizes are designed in this module to extract the high-frequency features of wave impedance. The schematic diagram of the network structure is as shown in Figure 4 shown.
[0057] Specifically, the scale matching module based on the deconvolutional neural network. The sampling interval of seismic data is much higher than that of well logging data, resulting in the frequency band width of seismic data being much smaller than that of well logging data. Therefore, how to match the frequency differences between seismic and well logging data is crucial for the accurate inversion of wave impedance in the follow-up. The deconvolutional network is an upsampling feature processing network, equivalent to a learnable interpolation algorithm. Through deconvolution operations, the previously extracted high-frequency and low-frequency features can be effectively fused to obtain the wave impedance features at the same scale. In this embodiment, to complete the scale matching work between seismic data and well logging data and make the seismic data scale gradually approach the well logging data scale, this module is designed to consist of two deconvolutional layers. The schematic diagram of the network structure is as shown in Figure 5 shown.
[0058] Specifically, the non-linear mapping module based on the fully connected neural network (DNN). The DNN network can realize the features of any non-linear relationship and obtain the non-linear mapping relationship between the previously extracted wave impedance features and the well logging wave impedance. In this embodiment, the non-linear mapping module based on the fully connected neural network (DNN) consists of a fully connected layer with 2 hidden layers. The schematic diagram of the network structure is as shown in Figure 6 shown.
[0059] Embodiment 2
[0060] As Figure 8 shown, this embodiment provides a seismic wave impedance inversion method under a deep learning framework, including:
[0061] Construct a training data set, where the training data set includes the relationship between well logging data and seismic data;
[0062] Construct a seismic wave impedance prediction model under the deep learning framework;
[0063] Use the training data set to train the seismic wave impedance prediction model and update the model parameters;
[0064] Based on the trained seismic wave impedance prediction model, achieve wide-band seismic wave impedance inversion.
[0065] Specifically, constructing the training data set includes:
[0066] Perform fine well-seismic calibration to establish the connection between logging data and seismic data, providing high-quality data samples for subsequent wave impedance inversion;
[0067] Construction and division of the sample set. Through fine well-seismic calibration, extract the seismic data adjacent to the well and logging data to form a sample set, and divide the training data set and validation data set according to a ratio close to 7:3.
[0068] Next, construct a seismic wave impedance prediction model under the deep learning framework. This model is mainly composed of four modules, which will be introduced separately below.
[0069] Module 1: Low-frequency feature extraction module based on the long short-term memory recurrent neural network (LSTM). Seismic and logging data are typical sequence signals. The LSTM network can explore the internal relationship between the front and back samples of the sequence signal, and at the same time overcome the problem that the traditional recurrent neural network (RNN) cannot effectively solve the long-term dependence relationship, and can effectively mine the internal relationship and connection between seismic and logging data. At the same time, using the characteristic that the deep network can capture more complex non-linear relationships and produce smoother outputs, the present invention designs a three-layer LSTM layer in this module. Therefore, it can be considered that this module can obtain the low-frequency features of wave impedance.
[0070] Module 2: High-frequency feature extraction module based on the convolutional neural network (CNN). The CNN network is different from the LSTM network and has no state variables, and its effect is poor when processing low-frequency information. However, it can capture high-frequency components through a small convolutional kernel. Therefore, the present invention uses the CNN network to extract the high-frequency features of wave impedance.
[0071] Module 3: Scale matching module based on the deconvolution neural network. The sampling interval of seismic data is much higher than that of logging data, resulting in the frequency band width of seismic data being much smaller than that of logging data. Therefore, how to match the frequency differences between seismic and logging data is crucial for the accurate inversion of subsequent wave impedance. The deconvolution network is an upsampling feature processing network, equivalent to a learnable interpolation algorithm. Through deconvolution operations, the previously extracted high-frequency features and low-frequency features can be effectively fused to obtain wave impedance features at the same scale.
[0072] Module 4: Non-linear mapping module based on the fully connected neural network (DNN). The DNN network can realize the features of any non-linear relationship and obtain the non-linear mapping relationship between the previously extracted wave impedance features and the logging wave impedance.
[0073] Next, perform model training and parameter optimization. Using the high-quality data samples in the first step above, taking the root mean square error of the training samples as the evaluation criterion, continuously update the model parameters to obtain the optimal wave impedance inversion model.
[0074] Finally, based on the trained model, broadband seismic wave impedance inversion is achieved.
[0075] Embodiment 3
[0076] like Figure 9 As shown, this embodiment uses an example to illustrate the implementation process and application effect of the present invention. Taking the data of a certain actual work area in China as an example, a seismic wave impedance inversion study is carried out based on the method described in the present invention. There are 5 well data in the study area, all of which have wave impedance curves.
[0077] Step 1: Fine well seismic calibration. Fine calibration of 5 wells in the study area was completed. Figure 2 The calibration of Well S1 is shown. Figure 2 It can be seen that the wave group characteristics of the synthetic record and the seismic data near the target layer are consistent and have good correlation, providing high-quality data samples for subsequent wave impedance inversion.
[0078] Step 2: Construction and division of sample sets. Through fine well-seismic calibration, seismic wellside data and logging data are extracted to form a sample set. The training set and validation set are divided in a ratio of nearly 7:3. Four wells are randomly selected as sample wells, and the remaining well (S4 well) is used as a validation well.
[0079] Step 3: Construct a seismic wave impedance prediction model under the deep learning framework. The model consists of four modules: a low-frequency feature extraction module based on a long short-term memory recurrent network (LSTM), a high-frequency feature extraction module based on a convolutional neural network (CNN), a scale matching module based on a deconvolutional neural network, and a nonlinear mapping module based on a fully connected neural network (DNN).
[0080] Module 1: Low-frequency feature extraction module based on long short-term memory recurrent network (LSTM). This module consists of three LSTM layers, each of which is as follows: Figure 3 The low-frequency feature extraction module uses the LSTM network to explore the intrinsic connection between the samples before and after the sequence signal, thereby mining the low-frequency wave impedance characteristics contained in the seismic data.
[0081] Module 2: High-frequency feature extraction module based on convolutional neural network (CNN). Unlike LSTM network, CNN network has no state variables and is less effective in processing low-frequency information. However, it can capture high-frequency components through smaller convolution kernels. Therefore, the present invention uses CNN network to extract high-frequency features of wave impedance. In order to effectively extract wave impedance information from seismic data, this module designs three convolution layers with different convolution kernel sizes to extract high-frequency features of wave impedance. The network structure diagram is shown in the figure. Figure 4 shown.
[0082] Module 3: Scale matching module based on deconvolution neural network. To complete the scale matching between seismic data and logging data and make the scale of seismic data gradually approach that of logging data, this module is designed to consist of two deconvolution layers. The schematic diagram of the network structure is as shown in Figure 5 shown.
[0083] Module 4: Nonlinear mapping module based on fully connected neural network (DNN). This module consists of a fully connected layer with 2 hidden layers. The schematic diagram of the network structure is as shown in Figure 6 shown.
[0084] Step 4: Training and parameter optimization of the model. Using the training samples in the aforementioned Step 2, with the root mean square error between the forward propagation of the model and the expected output as the evaluation criterion, continuously update the model parameters to obtain the optimal wave impedance inversion model.
[0085] Step 5: Based on the trained model, achieve wide-band seismic wave impedance inversion. Figure 7 The wave impedance inversion profile passing through the verification well (Well S4) is shown. It can be seen from the figure that the inverted seismic wave impedance of the well-side trace is in good agreement with the logging wave impedance, and at the same time, the lateral continuity of the inverted seismic wave impedance profile is good, indicating that the wave impedance inversion result of this method is relatively ideal.
[0086] Example 4
[0087] This example provides an electronic device, which includes:
[0088] A memory storing executable instructions;
[0089] A processor that runs the executable instructions in the memory to implement the seismic wave impedance inversion method under the above-mentioned deep learning framework. The method includes:
[0090] Construct a training data set, which includes the relationship between logging data and seismic data;
[0091] Construct a seismic wave impedance prediction model under the deep learning framework;
[0092] Use the training data set to train the seismic wave impedance prediction model and update the model parameters;
[0093] Based on the trained seismic wave impedance prediction model, achieve wide-band seismic wave impedance inversion.
[0094] Example 5
[0095] This example provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the seismic wave impedance inversion method under the above-mentioned deep learning framework. The method includes:
[0096] Construct a training data set, where the training data set includes the relationship between well logging data and seismic data;
[0097] Construct a seismic wave impedance prediction model under a deep learning framework;
[0098] Use the training data set to train the seismic wave impedance prediction model and update the model parameters;
[0099] Based on the trained seismic wave impedance prediction model, achieve broadband seismic wave impedance inversion.
[0100] The above computer-readable storage media include, but are not limited to: optical storage media (e.g., CD-ROM and DVD), magneto-optical storage media (e.g., MO), magnetic storage media (e.g., magnetic tape or external hard drive), media with built-in rewritable non-volatile memory (e.g., memory card), and media with built-in ROM (e.g., ROM cartridge).
[0101] In summary, the present invention extracts the high-frequency features of wave impedance based on the LSTM network and extracts the low-frequency features of wave impedance using the CNN network. By effectively combining the LSTM network and the CNN network, broadband wave impedance features can be extracted. The present invention uses a deconvolution network to effectively solve the scale difference between seismic and well logging data. The wave impedance inversion model based on the deep learning framework of the present invention can effectively achieve broadband seismic wave impedance inversion.
[0102] The various embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments.
Claims
1. A seismic wave impedance prediction model under a deep learning framework, characterized in that, it includes: A low-frequency feature extraction module and a high-frequency feature extraction module, which are respectively used to extract the low-frequency features of wave impedance and the high-frequency features of wave impedance from the input data; A scale matching module, which is used to fuse the extracted high-frequency features and low-frequency features to obtain wave impedance features at the same scale; A non-linear mapping module, which is used to obtain the non-linear mapping relationship between the extracted wave impedance features and the well log wave impedance.
2. The seismic wave impedance prediction model under the deep learning framework according to claim 1, characterized in that, The low-frequency feature extraction module is based on a long short-term memory recurrent network and includes three LSTM layers.
3. The seismic wave impedance prediction model under the deep learning framework according to claim 1, characterized in that, The high-frequency feature extraction module is based on a convolutional neural network and includes three convolutional layers with different kernel sizes.
4. The seismic wave impedance prediction model under the deep learning framework according to claim 1, characterized in that, The scale matching module is based on a deconvolutional neural network and includes two deconvolutional layers.
5. The seismic wave impedance prediction model under the deep learning framework according to claim 1, characterized in that, The non-linear mapping module is based on a fully connected neural network and consists of a fully connected layer containing 2 hidden layers.
6. A seismic wave impedance inversion method under a deep learning framework, characterized in that, it includes: Construct a training data set, and the training data set includes the relationship between well log data and seismic data; Construct a seismic wave impedance prediction model under the deep learning framework according to any one of claims 1-5; Use the training data set to train the seismic wave impedance prediction model and update the model parameters; Based on the trained seismic wave impedance prediction model, achieve broadband seismic wave impedance inversion.
7. The seismic wave impedance inversion method under the deep learning framework according to claim 6, characterized in that, Constructing the training data set includes: Fine well-seismic calibration, extracting seismic data adjacent to the well and well log data, and constructing a sample set; Sample set division, dividing the training data set and the validation data set according to a ratio close to 7:
3.
8. The seismic wave impedance inversion method under the deep learning framework according to claim 6, characterized in that, Updating the model parameters includes using the root mean square error of the training data set as an evaluation criterion, continuously updating the model parameters, and obtaining an optimal wave impedance inversion model.
9. An electronic device, characterized in that, the electronic device includes: A memory, which stores executable instructions; A processor, and the processor runs the executable instructions in the memory to implement the seismic wave impedance inversion method under the deep learning framework according to any one of claims 6-8.
10. A non-transitory computer-readable storage medium, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, it implements the seismic wave impedance inversion method under the deep learning framework according to any one of claims 6-8.