Sparse dictionary inversion method and device based on transfer learning

Abstract

The invention relates to the field of machine learning and geophysical exploration processing, inversion and interpretation of petroleum and natural gas energy, and particularly discloses a sparse dictionary inversion method and device based on transfer learning, and the method comprises the steps: obtaining logging information feature libraries of different work areas through dictionary learning; constructing an advanced feature library based on the logging information feature libraries of the different work areas; extracting a seismic signal from the logging information and then constructing a positive operator; and performing inversion prediction by using the forward operator and the advanced feature library as constraints. According to the sparse dictionary inversion method based on transfer learning, logging information feature libraries of different work areas are obtained through dictionary learning, the problem that the number of atoms in the dictionary library is insufficient due to insufficient logging data is relieved, and the precision and resolution of an inversion result are improved.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and geophysical exploration processing, inversion and interpretation of oil and natural gas energy, and in particular to a sparse dictionary inversion method and device based on transfer learning. Background Art

[0002] In recent years, with the widespread application of machine learning in signal processing, image processing and other fields, more and more scholars have begun to focus on the use of machine learning technology to solve various complex problems in the field of seismic exploration. As a powerful tool for signal representation and data compression, sparse representation technology has attracted widespread attention from domestic and foreign researchers. The theoretical basis of sparse representation was perfected in the compressed sensing theory proposed by scientists such as Donoho, Candès, Romberg and Terence Tao in 2004. The outbreak of this technology began in 2006 with the classic K-SVD dictionary learning algorithm proposed by Aharon and Elad of the Israel Institute of Technology. Different from the traditional method of using orthogonal basis functions (such as Fourier transform and discrete cosine transform) to represent signals, sparse representation technology is based on an overcomplete basis function set (i.e., an overcomplete dictionary) to represent the signal as a linear combination of as few basis functions (also known as atoms) as possible in the set, so as to capture the characteristic information contained in the signal and facilitate the subsequent encoding, compression, reconstruction, classification and other processing of the signal. Seismic inversion based on sparse dictionaries can use high-frequency logging data to learn a sparse dictionary, and use the sparse dictionary as a priori information constraint for inversion. This method can improve the vertical resolution of the inversion result. There are two main methods in the process of sparse characterization of logging geological structure characteristics. One is to use a fixed dictionary, such as Fourier transform, wavelet transform, seislet transform, etc. The other method is to learn an adaptive dictionary suitable for sample characteristics from training samples based on a learning method, mainly including principal component analysis (PCA), optimal direction method (MOD), and singular value class algorithm (K-SVD). Compared with fixed dictionaries, this type of method has the advantage that different dictionaries can be generated for different fixed signals, and the dictionary atoms are more suitable for signal characteristics, so better sparse representation effects can be obtained; the disadvantage is that the amount of calculation for dictionary learning is large, so the size of the dictionary is limited. In the actual work area, the financial resources required to drill a well are huge, so the number of wells is relatively small, so the obtained structural feature atoms are relatively small, which is not enough to characterize the elastic parameters to be inverted. Therefore, how to obtain a rich structural feature library with a small number of wells is a direction worth studying.

[0003] Based on this technical background, the present invention studies a sparse dictionary inversion method and device based on transfer learning. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a sparse dictionary inversion method and device based on transfer learning. The method obtains a logging information feature library of different work areas through dictionary learning, alleviates the problem of insufficient number of dictionary library atoms caused by insufficient logging data, and improves the accuracy and resolution of the inversion results.

[0005] In order to achieve the above object, a first aspect of the present invention provides a sparse dictionary inversion method based on transfer learning, comprising:

[0006] Obtain the logging information feature library of different work areas through dictionary learning;

[0007] Building an advanced feature library based on the logging information feature library of the different work areas;

[0008] The forward operator is constructed after extracting the seismic signal from the well logging information;

[0009] The forward operator and the advanced feature library are used as constraints to perform inversion prediction.

[0010] A second aspect of the present invention provides a sparse dictionary inversion device based on transfer learning, comprising:

[0011] Dictionary learning module, used to obtain the logging information feature library of different work areas through dictionary learning;

[0012] A feature construction module, used for constructing an advanced feature library based on the logging information feature library of the different work areas;

[0013] An operator construction module is used to construct a forward modeling operator after extracting seismic signals from well logging information;

[0014] The inversion prediction module is used to perform inversion prediction using the forward operator and the advanced feature library as constraints.

[0015] A third aspect of the present invention provides an electronic device, the electronic device comprising:

[0016] A memory storing executable instructions;

[0017] A processor, wherein the processor runs the executable instructions in the memory to implement the sparse dictionary inversion method based on transfer learning described in the first aspect.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the sparse dictionary inversion method based on transfer learning described in the first aspect.

[0019] The beneficial effects of the present invention include:

[0020] (1) The sparse dictionary inversion method based on transfer learning proposed in the present invention obtains the logging information feature library of different work areas through dictionary learning, which alleviates the problem of insufficient number of dictionary library atoms caused by insufficient logging data and improves the accuracy and resolution of the inversion results.

[0021] (2) The sparse dictionary inversion method based on transfer learning proposed in the present invention improves the stability of the inversion process and the accuracy of the inversion results by migrating the dictionary obtained from the logging data in one work area to another work area.

[0022] (3) The sparse dictionary inversion method based on transfer learning proposed in the present invention obtains a feature library of logging information through dictionary learning, migrates the feature library to another work area, and adds it to the feature library of another work area. After forming an advanced feature library, it is added to the inversion process as a constraint, and combined with the inherent physical mechanism of forward and inversion of seismic signals, an inversion method driven by both data and model is realized.

[0023] (4) The sparse dictionary inversion method based on transfer learning proposed in the present invention does not require human intervention through the constraints of logging information in multiple work areas. Its inversion result only depends on the inherent physical mechanism of seismic signal propagation and the data itself including seismic observation signals and logging data, and can be applied to inversion tasks with different geological conditions.

[0024] Other features and advantages of the present invention will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings.

[0026] Figure 1 This is a flow chart of the sparse dictionary inversion method based on transfer learning proposed in the present invention.

[0027] Figure 2 This is a schematic diagram of the comparison between the longitudinal wave impedance feature library of the overthrust model and the longitudinal wave impedance feature library of the Marmousi model in a specific implementation of the sparse dictionary inversion method based on transfer learning proposed in the present invention.

[0028] Figure 3 This is a schematic diagram comparing the impedance model of the overthrow body and the Marmousi impedance model in a specific implementation of the sparse dictionary inversion method based on transfer learning proposed in the present invention.

[0029] Figure 4This is a schematic diagram comparing the inversion results of the thrust impedance model and the inversion results of the Marmousi model in a specific implementation of the sparse dictionary inversion method based on transfer learning proposed in the present invention.

[0030] Figure 5 A comparative schematic diagram of the spectrum analysis of the inversion results of the sparse dictionary Marmousi impedance model based on transfer learning proposed in the present invention.

[0031] Figure 6 A schematic diagram of the comparison between the spectrum of real seismic records and synthetic data of the Marmousi model and a schematic diagram of the comparison between the spectrum of real seismic records and synthetic data of the thrust body impedance model in a specific implementation of the sparse dictionary inversion method based on transfer learning proposed in the present invention.

[0032] Figure 7 This is a schematic diagram of the comparison of the spectrum of the well logging data with the numerical model in a specific implementation of the sparse dictionary inversion method based on transfer learning proposed in the present invention. DETAILED DESCRIPTION

[0033] 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 to the embodiments set forth herein.

[0034] The present invention provides a sparse dictionary inversion method based on transfer learning, such as Figure 1 As shown, including:

[0035] Obtain the logging information feature library of different work areas through dictionary learning;

[0036] The logging information feature library based on different work areas is constructed as an advanced feature library;

[0037] The forward operator is constructed after extracting the seismic signal from the well logging information;

[0038] Inversion prediction is performed using forward operators and advanced feature libraries as constraints.

[0039] In the present invention, the logging information feature library of different work areas is obtained through dictionary learning, which alleviates the problem of insufficient number of dictionary library atoms due to insufficient logging data and improves the accuracy and resolution of the inversion result.

[0040] According to the present invention, the feature library of well logging information of different work areas obtained by dictionary learning includes:

[0041] Extract multiple overlapping blocks from the logging information of each work area, expand the multiple overlapping blocks into column vectors and arrange them in sequence to obtain a training set;

[0042] The K-SVD algorithm is used to perform feature learning on the training set to obtain the logging information feature library of the work area.

[0043] In the present invention, the dictionary obtained from the logging data in one work area is transferred to another work area, thereby improving the stability of the inversion process and the accuracy of the inversion result.

[0044] In the present invention, a feature library of logging information is obtained through dictionary learning, and the feature library is migrated to another work area and added to the feature library of another work area. After forming an advanced feature library, it is added to the inversion process as a constraint, and combined with the inherent physical mechanism of forward and inversion of seismic signals, an inversion method driven by both data and model is realized.

[0045] According to the present invention, the mathematical expression of the K-SVD algorithm is:

[0046]

[0047]

[0048] Among them, Y is the training set, X is the sparse representation of the signal, A is the logging information feature library, T 0 is the upper limit of non-zero elements in the sparse representation coefficients.

[0049] According to the present invention, the well logging information feature library based on different work areas is constructed as an advanced feature library including:

[0050] Except for the migrated work area, the logging information feature library of each other work area is used as a sparse sub-feature library and migrated to the logging information feature library of the migrated work area to form an advanced feature library.

[0051] Preferably, the formula used to construct the forward operator is:

[0052]

[0053] Among them, S is the seismic signal, G is the forward operator, M is the P-wave impedance model parameter, λ is the regularization parameter, m is the elastic parameter to be inverted, φ is the prior information constraint, and the prior information constraint is L 1 Norm constraint or L 2 Norm constraints are used to introduce specific or expected geological structural features, including blocky structures and smooth structures. i is the sparse coefficient in the inversion process, L is the objective function;

[0054] The regularization parameter is used to balance the weight of the regularization constraint and the data matching term.

[0055] Preferably, the formula of the objective function used for inversion is:

[0056]

[0057]

[0058] Among them, R i is a constant matrix used to extract the local small blocks of the parameters to be inverted and fit them with the reconstructed small blocks. D is the advanced feature library.

[0059] According to the present invention, the forward operator serves as a model-driven constraint;

[0060] Advanced feature libraries as data-driven constraints.

[0061] In the present invention, through the constraints of logging information in multiple work areas, the present invention does not require human intervention, and its inversion result only depends on the inherent physical mechanism of seismic signal propagation and the data itself including seismic observation signals and logging data, and can be applied to inversion tasks with different geological conditions.

[0062] The present invention will be described in more detail below by way of examples.

[0063] Embodiment 1:

[0064] In this embodiment, two numerical models are selected for mutual transfer learning test: the Marmousi impedance model with a more complex geological structure and the nappe impedance model with a simpler geological structure; considering the complexity of the geological structure and the size of the numerical model, the longitudinal wave impedance feature library obtained by different numerical models, the feature extraction sliding window size of the characteristic atoms in the nappe model feature library is 120, the number of atoms is 7500, the characteristic atom length in the Marmousi model feature library is 90, and the number of atoms is 10000, and the longitudinal wave impedance feature library of one numerical model is substituted into another numerical model for diversified geological structure feature constraint inversion; the specific process of the sparse dictionary inversion method based on transfer learning in this embodiment is as follows:

[0065] Discretization of the plane wave propagation process in the underground: Assume that the underground target area is composed of N l Small flat layers are constructed, and two adjacent small layers will form a dielectric interface. Let i = 1, 2, ···, N l -1 represents the interface between layer i and layer i+1, is the longitudinal wave impedance reflection coefficient of the i-th interface. Since the elastic difference between two adjacent layers is small, Buland and Omre proposed the following approximation:

[0066]

[0067] Among them, V p is the average value of the corresponding longitudinal wave velocity on both sides, ΔV pThe longitudinal wave velocity on both sides of the corresponding interface; based on the approximate simplification of the Zoeppritz equation, an approximate equation that is easier to calculate numerically is obtained to construct the inversion target functional. The following formula can be obtained through formula (1):

[0068]

[0069] in, and is the longitudinal wave impedance of the i-th layer. For multiple sampling points, equation (2) is written as the following matrix form:

[0070]

[0071] The above formula can be simplified as:

[0072]

[0073] Among them, R p represents the longitudinal wave impedance reflection coefficient vector, L p represents the natural logarithm of the longitudinal wave impedance vector, D represents the first matrix on the right side of the equal sign in the first-order difference matrix corresponding to equation (3); according to the convolution model, the forward model from the reflection signal to the seismic signal can be expressed as follows:

[0074] S = GM (5);

[0075] Among them, S represents the synthetic seismic signal, the matrix G represents the forward operator, which is constructed by the wavelet, the P-wave background velocity and the constant linear fitting coefficient, and finally M represents the P-wave impedance model parameters. Therefore, solving the impedance model parameters can essentially be achieved by solving the following least squares problem:

[0076]

[0077] Since S is usually seriously disturbed by noise and the system matrix G has a very large condition number, the solution to equation (6) is unstable in most cases. This is the famous "ill-conditioned" inverse problem. Therefore, given S and G, the estimation of M also depends on regularization technology. Adding appropriate prior constraints and regularizing the objective function can help enhance the stability of the inversion and reduce the multi-solution of the inverse problem. At the same time, it can also effectively supplement the missing low-frequency and high-frequency component information in the seismic data and improve the resolution of the inversion results.

[0078] By L 2 The norm is used as a sparse constraint term to establish a new objective function, which is expressed as follows:

[0079]

[0080] Among them, λ is the regularization parameter, which is used to balance the weight of the regularization constraint and the data matching term, m is the elastic parameter to be inverted, φ is the prior information constraint, which is the L1 norm constraint or the L2 norm constraint, which is used to introduce specific or expected geological structure characteristics, including blocky structure and smooth structure, a i is the sparse coefficient in the inversion process, and L is the objective function; as mentioned above, the traditional regularization technology is a model-driven process that gives some formulated mathematical models, which is usually poorly adaptable to complex geological conditions and only uses the low-frequency components of high-resolution logging data, which is a potential waste of resources. Therefore, the solution method for the objective function uses the underground geological structure as a priori information to constrain the inversion of seismic data, so that the high-frequency information of logging and its fine structure are integrated with the seismic data to obtain a higher-resolution inversion result;

[0081] In this embodiment, a learning-based method is used to generate a feature library expressing diverse geological structures. During the training process, multiple blocks with certain overlap are extracted from the well logging information, and all the blocks are expanded into column vectors and arranged in sequence to form a training set Y; the original feature library is initialized, and the K-SVD algorithm is used to perform feature learning on the training set Y. Each iteration updates each atom in the feature library. Its mathematical model has the following form:

[0082]

[0083] Among them, Y represents the training set, X is the sparse representation of the signal, and A represents the feature library containing the diverse underground geological structures. 0 Represents the upper limit of non-zero elements in the sparse representation coefficients;

[0084] After obtaining the advanced feature library D from the logging elastic parameter curve, these features generally contain diverse geological structures, which are used as data-driven regularization to constrain the model inversion, that is, to indirectly and organically integrate the logging information into the inversion process, to select limited logging phases from the advanced features, and to combine them in the simplest way to form a model-driven solution, which can greatly reduce the solution space of conventional reservoir parameter inversion. The expression of the inversion objective function is:

[0085]

[0086] Among them, S represents the observed seismic record, G represents the forward operator, m represents the elastic parameter to be inverted, λ represents the regularization parameter, D represents the feature library generated after multi-scale transformation, α represents the i represents the sparse representation of the signal, K represents the maximum number of non-zero elements in the sparse representation, and Ri is a constant matrix used to extract the local small blocks of the parameters to be inverted and fit them with the reconstructed small blocks.

[0087] from Figure 2 road Figure 7 The comparison results show that the migration of the feature library from the complex model (Marmousi model) to the simple model (nappe model) can obtain better results. It can be seen that compared with the conventional regularization constraints, the addition of geological structure feature constraints has expanded both the low frequency and the high frequency. In the low frequency region within the range of 0 to 15 Hz, the frequency spectrum of the inversion result obtained by the inversion method of logging data and geological structure feature constraints is more consistent with the real impedance data.

[0088] However, in the feature library migration from a simple model (nappe model) to a complex model (Marmousi model), it can be seen that the feature library migration did not work. The main reason is that the two feature libraries come from logging data of different models, and because the size of the sliding window needs to be set when generating the feature library, the feature library contains different frequencies and feature information, which is mainly due to the poor effect of migration learning from the spectrum analysis of seismic data and logging data.

[0089] Embodiment 2:

[0090] This embodiment provides a sparse dictionary inversion method based on transfer learning, such as Figure 1 As shown, including:

[0091] Obtain the logging information feature library of different work areas through dictionary learning;

[0092] The logging information feature library based on different work areas is constructed as an advanced feature library;

[0093] The forward operator is constructed after extracting the seismic signal from the well logging information;

[0094] Use forward operators and advanced feature libraries as constraints for inversion prediction;

[0095] The feature library of logging information obtained from different work areas through dictionary learning includes:

[0096] Extract multiple overlapping blocks from the logging information of each work area, expand the multiple overlapping blocks into column vectors and arrange them in sequence to obtain a training set;

[0097] The K-SVD algorithm is used to learn the features of the training set to obtain the logging information feature library of the work area;

[0098] The mathematical expression of the K-SVD algorithm is:

[0099]

[0100]

[0101] Among them, Y is the training set, X is the sparse representation of the signal, A is the logging information feature library, T 0is the upper limit of non-zero elements in the sparse representation coefficients;

[0102] The advanced feature library based on the logging information feature library of different work areas includes:

[0103] Except for the migrated work area, the logging information feature library of each other work area is used as a sparse sub-feature library and migrated to the logging information feature library of the migrated work area to form an advanced feature library;

[0104] The formula used to construct the forward operator is:

[0105]

[0106] Among them, S is the seismic signal, G is the forward operator, M is the P-wave impedance model parameter, λ is the regularization parameter, m is the elastic parameter to be inverted, φ is the prior information constraint, and the prior information constraint is L 1 Norm constraint or L 2 Norm constraints are used to introduce specific or expected geological structural features, including blocky structures and smooth structures. i is the sparse coefficient in the inversion process, L is the objective function;

[0107] The regularization parameter is used to balance the weight of the regularization constraint and the data matching term;

[0108] The formula of the objective function used in the inversion is:

[0109]

[0110]

[0111] Among them, R i is a constant matrix, used to extract the local small blocks of the parameters to be inverted and fit them with the reconstructed small blocks, and D is the high-level feature library;

[0112] Forward operators as model-driven constraints;

[0113] Advanced feature libraries as data-driven constraints.

[0114] Embodiment three:

[0115] This embodiment provides a sparse dictionary inversion device based on transfer learning, including:

[0116] Dictionary learning module, used to obtain the logging information feature library of different work areas through dictionary learning;

[0117] Feature construction module, used to construct an advanced feature library based on the logging information feature library of different work areas;

[0118] An operator construction module is used to construct a forward modeling operator after extracting seismic signals from well logging information;

[0119] Inversion prediction module, used to perform inversion prediction using forward operators and advanced feature libraries as constraints;

[0120] The feature library of logging information obtained from different work areas through dictionary learning includes:

[0121] Extract multiple overlapping blocks from the logging information of each work area, expand the multiple overlapping blocks into column vectors and arrange them in sequence to obtain a training set;

[0122] The K-SVD algorithm is used to learn the features of the training set to obtain the logging information feature library of the work area;

[0123] The mathematical expression of the K-SVD algorithm is:

[0124]

[0125]

[0126] Among them, Y is the training set, X is the sparse representation of the signal, A is the logging information feature library, T 0 is the upper limit of non-zero elements in the sparse representation coefficients;

[0127] The advanced feature library based on the logging information feature library of different work areas includes:

[0128] Except for the migrated work area, the logging information feature library of each other work area is used as a sparse sub-feature library and migrated to the logging information feature library of the migrated work area to form an advanced feature library;

[0129] The formula used to construct the forward operator is:

[0130]

[0131] Among them, S is the seismic signal, G is the forward operator, M is the P-wave impedance model parameter, λ is the regularization parameter, m is the elastic parameter to be inverted, φ is the prior information constraint, and the prior information constraint is L 1 Norm constraint or L 2 Norm constraints are used to introduce specific or expected geological structural features, including blocky structures and smooth structures. i is the sparse coefficient in the inversion process, and L is the objective function;

[0132] The regularization parameter is used to balance the weight of the regularization constraint and the data matching term;

[0133] The formula of the objective function used in the inversion is:

[0134]

[0135]

[0136] Among them, R i is a constant matrix, used to extract the local small blocks of the parameters to be inverted and fit them with the reconstructed small blocks, and D is the high-level feature library;

[0137] Forward operators as model-driven constraints;

[0138] Advanced feature libraries as data-driven constraints.

[0139] Embodiment 4:

[0140] An embodiment of the present invention provides an electronic device including a memory and a processor.

[0141] A memory storing executable instructions;

[0142] The processor runs the executable instructions in the memory to implement the sparse dictionary inversion method based on transfer learning.

[0143] The memory is used to store non-temporary computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0144] The processor may be a central processing unit (CPU) or other forms of processing units with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. In one embodiment of the present invention, the processor is used to run the computer-readable instructions stored in the memory.

[0145] Those skilled in the art should be able to understand that in order to solve the technical problem of how to obtain a good user experience, the present embodiment may also include well-known structures such as a communication bus and an interface, and these well-known structures should also be included in the protection scope of the present invention.

[0146] For detailed description of this embodiment, reference may be made to the corresponding descriptions in the aforementioned embodiments, which will not be repeated here.

[0147] Embodiment five:

[0148] An embodiment of the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, a sparse dictionary inversion method based on transfer learning is implemented.

[0149] The computer-readable storage medium according to the embodiment of the present invention stores non-transitory computer-readable instructions, and when the non-transitory computer-readable instructions are executed by a processor, all or part of the steps of the above-mentioned methods of the embodiments of the present invention are executed.

[0150] The above-mentioned 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 mobile hard disk), media with built-in rewritable non-volatile memory (e.g., memory card) and media with built-in ROM (e.g., ROM box).

[0151] The sparse dictionary inversion method based on transfer learning proposed in the embodiment of the present invention obtains the logging information feature library of different work areas through dictionary learning, alleviates the problem of insufficient number of dictionary library atoms caused by insufficient logging data, and improves the accuracy and resolution of the inversion result.

[0152] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A sparse dictionary inversion method based on transfer learning, characterized in that, it includes: Obtaining a logging information feature library for different work areas through dictionary learning; Constructing a high-level feature library based on the logging information feature libraries of different work areas; Extracting seismic signals from logging information and then constructing a forward operator; Using the forward operator and the high-level feature library as constraints for inversion prediction.

2. The inversion method according to claim 1, characterized in that, Obtaining a logging information feature library for different work areas through dictionary learning includes: Extracting multiple overlapping blocks from the logging information of each work area, unfolding the multiple overlapping blocks into column vectors and arranging them in sequence to obtain a training set; Performing feature learning on the training set using the K-SVD algorithm to obtain the logging information feature library of this work area.

3. The inversion method according to claim 2, characterized in that, The mathematical expression of the K-SVD algorithm is: Among them, Y is the training set, X is the sparse representation of the signal, A is the well logging information feature library, and T 0 is the upper limit of the non-zero elements in the sparse representation coefficient.

4. The inversion method according to claim 1, characterized in that, Constructing a high-level feature library based on the logging information feature libraries of different work areas includes: Except for the work area to be migrated, the logging information feature library of each other work area is used as a sparse sub-feature library and migrated into the logging information feature library of the work area to be migrated to form a high-level feature library.

5. The inversion method according to claim 3, characterized in that, The formula used to construct the forward operator is: Wherein, S is the seismic signal, G is the forward operator, M is the P-wave impedance model parameter, λ is the regularization parameter, m is the elastic parameter to be inverted, φ is the prior information constraint, and the prior information constraint is the L 1 norm constraint or the L 2 norm constraint, which is used to introduce specific or expected geological structure features, and the geological structure features include blocky structures and smooth structures, a i is the sparse coefficient in the inversion process, and L is the objective function; The regularization parameter is used to balance the weights of the regularization constraint and the data matching term.

6. The inversion method according to claim 5, characterized in that, The formula of the objective function used for inversion is: Among them, R i is a constant matrix used to extract local patches of the parameters to be inverted and fit them with the reconstructed patches, and D is a high-level feature library.

7. The inversion method according to claim 1, characterized in that, The forward operator is used as a model-driven constraint; The high-level feature library is used as a data-driven constraint.

8. An inversion device based on data-driven and model-driven, characterized in that, it includes: A dictionary learning module for obtaining a logging information feature library for different work areas through dictionary learning; A feature construction module for constructing a high-level feature library based on the logging information feature libraries of different work areas; An operator construction module for extracting seismic signals from logging information and then constructing a forward operator; An inversion prediction module for using the forward operator and the high-level feature library as constraints for inversion prediction.

9. An electronic device, characterized in that, the electronic device includes: A memory storing executable instructions; A processor that runs the executable instructions in the memory to implement the data-driven and model-driven inversion method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, this computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the data-driven and model-driven inversion method according to any one of claims 1-7.

Images