An earthquake inversion method and device based on a multi-task reversible network

Abstract

The present invention provides a seismic inversion method and device based on a multi-task reversible network, which relates to the technical field of seismic exploration. The method first obtains well-seismic data in the study area and processes it to construct a data set, then establishes a multi-task reversible network model and uses the constructed data set for training to obtain a complex relationship model between target data and seismic data, and finally performs inversion on actual data to obtain a target data volume. The present invention can make up for the problem that conventional inversion methods are difficult to fully utilize the mutual dependence relationship between input and output data and the mutual coupling relationship between different output data, and improve the accuracy of inversion results.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic exploration, and more specifically, to a seismic inversion method and device based on a multi-task reversible network. Background Art

[0002] Seismic inversion is one of the important contents of geophysical inversion and plays a crucial role in oil and gas seismic exploration. Scholars at home and abroad have developed some effective seismic inversion methods through continuous exploration and achieved good application effects under certain conditions, but there are also certain limitations. For example, the current commonly used AVO inversion mostly uses an approximation of the Zoepppritz equation to establish a model, and the approximation holds under certain assumptions, and it is sometimes difficult to obtain satisfactory results in the exploration of complex oil and gas reservoirs. In addition, conventional inversion methods often ignore the mutual dependence between target data and seismic data and the mutual coupling relationship between different target data. Summary of the Invention

[0003] A seismic inversion method and device based on a multi-task reversible network proposed by the present invention aims to solve or partially solve the problem that conventional inversion methods are difficult to make full use of the mutual dependence relationship in the forward and inverse inversion processes and the mutual coupling relationship between different target data, and improve the accuracy and efficiency of the inversion results.

[0004] To achieve the above object and reach the above technical effect, the present invention provides the following technical solutions:

[0005] According to one aspect of the present invention, a seismic inversion method based on a multi-task reversible network is provided. The method includes three main steps: data acquisition, model establishment, and data inversion:

[0006] Data acquisition: acquiring seismic and logging data of the study area and processing them to construct a data set; including: acquiring seismic and logging data of the study area and processing them; converting the processed seismic data into angle gather data to obtain a common angle gather stacked data volume; based on the processed logging data and different angle gather stacked data volumes, comprehensively calibrating horizons with fine geological information; obtaining target data based on the logging data and performing scale transformation according to the calibrated horizons; constructing a data set based on the obtained target data and angle gather stacked data;

[0007] Model establishment, establishing a multi-task reversible network model and using the constructed data set for training to obtain a complex geophysical relationship model between target data and seismic data; including: determining the model input and output dimensions according to the multi-task objectives and establishing a reversible module of the reversible network; building a multi-task reversible network model based on the established reversible module; setting network model parameters; calculating the optimization target value during network training; updating the network model parameters according to the calculated optimization target value; saving the network model parameters; obtaining a multi-task reversible network inversion model;

[0008] Data inversion, using the trained multi-task reversible network model to invert the field seismic data to obtain the target data volume of the study area; including: inputting the implicit variables and extracted angular seismic data output in the network training into the inversion process of the trained multi-task reversible network model, inverting to obtain the target data volume of the study area, where the dimension of the forward output needs to be determined according to the target task.

[0009] As a further preferred embodiment of the present invention, the processing of seismic and logging data includes trace editing, amplitude compensation, denoising and prestack migration of seismic data, and outlier removal, correction and lateral standardization of logging data.

[0010] As a further preferred embodiment of the present invention, in addition to calibrating the horizons by synthetic seismic records and through-well seismic data, the comprehensive geological information fine calibration also needs to be checked and adjusted one by one in combination with geological information and other well logging data.

[0011] As a further preferred embodiment of the present invention, the target data may include longitudinal and transverse wave velocities and density, etc.

[0012] As a further preferred embodiment of the present invention, the scale transformation is to convert the obtained target data into the time domain by using time-depth conversion. After the depth-time conversion, it is first smoothed and then downsampled to have the same size as the angular seismic data.

[0013] As a further preference of the present invention, the constructed data set includes a training set and a test set, wherein the training set is used to train the multi-task reversible network model, and the test set is used to test the performance of the model; the training set is composed of target data and angular track data, wherein the dimension of the forward output needs to be determined according to the target task.

[0014] As a further preferred embodiment of the present invention, multi-task is to train the network model by multi-task joint learning in the modeling process, and to mine the mutual coupling relationship between different target data while learning to characterize the complex geophysical relationship between seismic amplitude characteristics and target data.

[0015] As a further preference of the present invention, the optimized target value of the calculation includes three loss functions, one supervised loss function and two unsupervised loss functions. Among them, the forward process uses the supervised loss function to measure the error between the observed and predicted seismic data; the first unsupervised loss function aims to capture the loss information around the input data in the forward process. Since some information is lost in the forward process, additional latent output variables are introduced, which are trained to capture the information related to the forward input but not included in the forward output. In addition, the second unsupervised loss function further trains the model by comparing the distributions of the reverse prediction and the forward input.

[0016] According to another aspect of the present invention, there is provided a seismic inversion device based on a multi-task reversible network, comprising:

[0017] A data acquisition module, configured to acquire seismic and logging data of a study area and process them to construct a data set; including: acquiring seismic and logging data of the study area and processing them; converting the processed seismic data into common angle gather data to obtain a common angle gather stacked data volume; based on the processed logging data and different angle gather stacked data volumes, comprehensively calibrating horizons with fine geological information; obtaining target data based on the logging data and performing scale transformation according to the calibrated horizons; constructing a data set based on the obtained target data and angle gather stacked data;

[0018] A model establishment module, configured to establish a multi-task reversible network model and use the constructed data set for training to obtain a complex geophysical relationship model between target data and seismic data; including: determining the input and output dimensions of the model according to the multi-task objective and establishing a reversible module of the reversible network; constructing a multi-task reversible network model based on the established reversible module; setting network model parameters; calculating the optimized target value during network training; updating the network model parameters according to the calculated optimized target value; saving the network model parameters; obtaining a multi-task reversible network inversion model;

[0019] A data inversion module, configured to use the trained multi-task reversible network model to invert on-site seismic data to obtain a target data volume of the study area; including: inputting the implicit variable output during network training and the extracted angle seismic data into the inversion process of the trained multi-task reversible network model to invert and obtain the target data volume of the study area, where the dimension of the forward output needs to be determined according to the target task.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: By combining the advantages of multi-task learning and reversible networks, the present invention can make up for the problems that conventional inversion methods cannot fully utilize the interdependent relationship between seismic data and target data and the mutual coupling relationship between different target data, and improve the accuracy and efficiency of the inversion results. Description of the Drawings

[0021] Figure 1 Flow chart of a seismic inversion method based on a multi - task reversible network provided by an embodiment of the present invention;

[0022] Figure 2 Comparison diagram of single - task and multi - task learning processes provided by an embodiment of the present invention;

[0023] Figure 3 Schematic diagram of the reversible network provided by an embodiment of the present invention;

[0024] Figure 4 Comparison diagram of different inversion processes provided by an embodiment of the present invention.

[0025] Figure 5 Structural diagram of a seismic inversion device based on a multi - task reversible network provided by an embodiment of the present invention. Detailed implementation manners

[0026] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of the specification, rather than all the embodiments.

[0027] First, a brief introduction to multi - task learning and reversible networks involved in this application is as follows:

[0028] Multi - task learning aims to utilize the similarity between different tasks to jointly learn the implicit information in multiple related tasks, solve multiple different tasks simultaneously, and improve the performance of individual tasks.

[0029] Figure 2 For the comparison between single - task and multi - task processes, in single - task, different target data are regarded as isolated tasks and modeled and inverted separately. Each network model has only one output target for the same input, and there is no connection between different models. The internal structure of the network models trained separately has no impact on other network models. While multi - task focuses on the mutual connection between different tasks, can better mine the rich correlation information contained in different target data, and overcome the problem that single - task cannot well consider the mutual coupling relationship between different target data.

[0030] The reversible network optimizes the modeling process in both forward and reverse directions and uses an additional implicit output variable to capture the information that may be lost in the forward process. That is, after the input data is propagated forward to obtain the output, the original input data can be obtained through reverse propagation according to the output in the opposite process, and no information of the input data is lost in this process.

[0031] The basic unit of the reversible network is a reversible module with two complementary affine coupling layers. The forward process of the affine coupling layer is as shown in Figure 3 Figure a, where x 1 and x 2 are two parts divided from the input x. These two parts are transformed by the learning functions t i and s i (i = 1, 2) and coupled in an alternating manner. x 1 and x 2 can be transformed into y 1 and y 2 through the following formula:

[0032] y 1 = x 1 ⊙ exp(s 2 (x 2 )) + t 2 (x 2 )

[0033] y 2 = x 2 ⊙ exp(s 1 (x 1 )) + t 1 (x 1 )

[0034] where ⊙ represents element-wise multiplication. t 1 , t 2 , s 1 and s 2 can be arbitrarily complex functions, such as fully connected networks or convolutional networks, and the functions themselves do not need to be reversible.

[0035] The reverse process of the affine coupling layer is as shown in Figure 3 Figure b. The reversible process of the above formula can be obtained through the following formula:

[0036] x 1 = (y 1 - t 2 (x 2 )) ⊙ exp(-s 2 (x 2 ))

[0037] x 2 = (y 2 - t 1 (x 1 )) ⊙ exp(-s 1 (x 1 ))

[0038] An earthquake inversion method based on a multi-task reversible network provided by an embodiment of the present invention is described with reference to Figure 1, the specific implementation steps of the present invention are as follows:

[0039] Step 101: Data acquisition. Acquire seismic and logging data in the study area and process them to construct a data set, including: acquiring seismic and logging data in the study area, performing trace editing, amplitude compensation, denoising, and prestack migration on the seismic data, and performing outlier rejection, correction, and lateral normalization on the logging data; converting the processed seismic data into common angle gather data to obtain a common angle gather stacked data volume. Since this embodiment takes the inversion of P-wave and S-wave velocities and density as an example for illustration, it is necessary to extract the stacked data volumes corresponding to different incident angle ranges; based on the processed logging data and the stacked data volumes of different angle gathers, finely calibrate the horizons by integrating geological information; obtain target data based on the logging data and perform scale transformation using the calibrated horizons, that is, convert the obtained target data to the time domain using time-depth conversion. After deep-time conversion, first perform smoothing processing on it, and then perform downsampling to make it have the same size as the angle seismic data; construct a training set and a test set based on the obtained target data and the stacked data of the angle gather.

[0040] Step 102: Model establishment. Establish a multi-task reversible network model and use the constructed data set for training to obtain a complex geophysical relationship model between target data and seismic data, including: determining the input and output dimensions of the model according to multi-task objectives and establishing a reversible module of the reversible network, where the reversible module consists of affine coupling layers with bidirectional mapping; constructing a multi-task reversible network model based on the established reversible module, where the forward input is the target data to be inverted and the forward output is the stacked data volumes corresponding to different incident angle ranges; setting the depth and learning function of the multi-task reversible network model, where the model depth is the number of layers corresponding to the reversible module, and the learning function can be any complex function, such as a fully connected network or a convolutional network; calculating the optimization objective value during network training, and the calculated optimization objective value includes a supervised loss function and two unsupervised loss functions; updating the model parameters according to the calculated optimization objective value; saving the network model parameters; obtaining a multi-task reversible network inversion model.

[0041] Step 103: Data inversion. Use the trained multi-task reversible network model to invert the on-site seismic data to obtain the target data volume of the study area, including: inputting the implicit variable output during network training and the extracted real angle seismic data into the inversion process of the trained multi-task reversible network model to invert and obtain the target data volume of the study area, where the dimension of the forward output needs to be determined according to the target task.

[0042] Figure 4For the comparison of the inversion processes of different networks, ordinary networks generally model the forward and inverse processes separately, that is, establish the forward relationship between the target data and the observed data or the inverse relationship between the observed data and the target data. The former is the basis of the latter. However, in the inversion based on ordinary networks, we generally only focus on the latter, that is, how to accurately construct the complex geophysical relationship between the observed seismic data and the target data, and realize obtaining various geophysical parameters characterizing the strata from the observed seismic data. This is the problem that needs to be solved in practical applications. However, in actual situations, there is often a one-to-many relationship between the observed seismic data and various geophysical parameters, that is, a set of observed seismic data may correspond to multiple different target data combinations, resulting in an ill-posed inversion process. Using a reversible network for modeling and inversion can, to a certain extent, overcome the problems existing in the separate modeling of the forward and inverse processes of ordinary network models and improve the accuracy of the inversion results.

[0043] Corresponding to the above-mentioned seismic inversion method based on a multi-task reversible network, an embodiment of the present invention provides a seismic inversion device based on a multi-task reversible network. Refer to Figure 5 As shown, in some embodiments, the inversion device mainly includes three modules: a data acquisition module 201, a model establishment module 202, and a data inversion module 203.

[0044] The data acquisition module 201 is used to acquire seismic and logging data of the study area and process them to construct a data set; including: acquiring seismic and logging data of the study area and processing them; converting the processed seismic data into angle gather data to obtain a common angle gather stacked data volume; based on the processed logging data and different angle gather stacked data volumes, finely calibrating horizons by integrating geological information; obtaining target data based on the logging data and performing scale transformation according to the calibrated horizons; constructing a data set based on the obtained target data and angle gather stacked data;

[0045] The model establishment module 202 is used to establish a multi-task reversible network model and use the constructed data set for training to obtain a complex geophysical relationship model between the target data and the seismic data; including: determining the input and output dimensions of the model according to the multi-task objectives and establishing a reversible module of the reversible network; constructing a multi-task reversible network model based on the established reversible module; setting network model parameters; calculating the optimization objective value during network training; updating the network model parameters according to the calculated optimization objective value; saving the network model parameters; obtaining a multi-task reversible network inversion model;

[0046] The data inversion module 203 is configured to perform inversion on the on-site seismic data by using the trained multi-task invertible network model to obtain the target data volume of the study area, including: inputting the implicit variable output during network training and the extracted angular seismic data into the inversion process of the trained multi-task invertible network model to inversely obtain the target data volume of the study area, where the dimension of the forward output needs to be determined according to the target task.

[0047] The seismic inversion device based on the multi-task invertible network provided by the embodiment of the present invention can execute the technical solution of the method embodiment described above Figure 1 The specific implementation process is detailed in the method embodiment and will not be elaborated here.

[0048] Those skilled in the art should understand that the above embodiments are only used to exemplarily illustrate the beneficial effects of the present invention and are not exhaustive. Any modifications, equivalent replacements, improvements, etc. made without departing from the scope and spirit of the described embodiments should not be excluded from the protection scope of the present invention.

Claims

1. A seismic inversion method based on multi-task reversible network, It is characterized in that The method includes three main steps: data acquisition, model building and data inversion: Data acquisition, acquiring seismic and logging data in the study area and processing them to construct a data set; including: acquiring seismic and logging data in the study area and processing them; converting the processed seismic data into angle gather data to obtain a common angle gather stacked data volume; based on the processed logging data and the stacked data volume of different angle gathers, finely calibrating the horizons with comprehensive geological information; obtaining target data based on the logging data and performing scale transformation according to the calibrated horizons; constructing a data set based on the obtained target data and the angle gather stacked data; Model establishment, establishing a multi-task reversible network model and using the constructed data set for training to obtain a complex geophysical relationship model between target data and seismic data; including: determining the model input and output dimensions according to the multi-task objectives and establishing a reversible module of the reversible network; building a multi-task reversible network model based on the established reversible module; setting network model parameters; calculating the optimization target value during network training; updating the network model parameters according to the calculated optimization target value; saving the network model parameters; obtaining a multi-task reversible network inversion model; Data inversion, using the trained multi-task reversible network model to invert the field seismic data to obtain the target data volume of the study area; including: inputting the implicit variables and extracted angular seismic data output in the network training into the inversion process of the trained multi-task reversible network model, inverting to obtain the target data volume of the study area, where the dimension of the forward output needs to be determined according to the target task.

2. A seismic inversion method based on a multi-task reversible network according to claim 1, Features: The method introduces a reversible network of bidirectional mapping to construct an inversion model structure, which can make up for the problem that conventional inversion methods cannot make good use of the mutual dependence between target data and seismic data.

3. A seismic inversion method based on a multi-task reversible network according to claim 1, Features: Multi-task learning is to train the network model by multi-task joint learning in the modeling process, and to explore the mutual coupling relationship between different target data while learning to characterize the complex geophysical relationship between seismic amplitude characteristics and target data.

4. A seismic inversion device based on a multi-task reversible network, It is characterized in that The device comprises: The data acquisition module is used to acquire the seismic and logging data of the study area and process and construct a data set; including: acquiring the seismic and logging data of the study area and processing; converting the processed seismic data into angle gather data to obtain a common angle gather stacked data volume; based on the processed logging data and the different angle gather stacked data volume, finely calibrating the horizon by integrating geological information; acquiring the target data based on the logging data and performing scale transformation according to the calibrated horizon; constructing a data set based on the acquired target data and the angle gather stacked data; The model establishment module is used to establish a multi-task reversible network model and train it using the constructed dataset to obtain a complex geophysical relationship model between the target data and the seismic data. It includes: determining the input and output dimensions of the model according to the multi-task objectives and establishing a reversible module of the reversible network; constructing a multi-task reversible network model based on the established reversible module; setting the network model parameters; calculating the optimization objective value during the network training; updating the network model parameters according to the calculated optimization objective value; saving the network model parameters; and obtaining a multi-task reversible network inversion model. The data inversion module is used to invert the on-site seismic data using the trained multi-task reversible network model to obtain the target data volume of the study area. It includes: inputting the implicit variables output during the network training and the extracted angular seismic data into the inversion process of the trained multi-task reversible network model to invert and obtain the target data volume of the study area, where the dimension of the forward output needs to be determined according to the target task.

Images