A sparse time-frequency analysis method, device and readable storage medium based on self-supervised generative adversarial learning and reconstruction of physical constraints

Abstract

The present invention discloses a sparse time-frequency analysis method, device, and readable storage medium based on self-supervised generative adversarial learning and reconstruction of physical constraints. The method includes using a deep learning framework to construct a sparse time-frequency analysis network based on a self-supervised generative adversarial network, wherein the sparse time-frequency analysis network includes a generator, a regulator, and a reconstruction module; calculating the sparse time-frequency spectrum of synthetic seismic data to construct a regulator data set; selecting a portion of actual seismic data and merging it with the synthetic seismic data as the generator data set; designing a generator loss function and a regulator loss function respectively; training a sparse time-frequency analysis network model based on a self-supervised generative adversarial network; and applying the trained sparse time-frequency analysis network model based on a self-supervised generative adversarial network to three-dimensional post-stack seismic data to analyze and interpret geological reservoir structures. This method can improve computational efficiency and reduce the design cost of sparse time-frequency transformation methods.

Description

Technical Field

[0001] The present invention relates to the field of earth exploration technology, and in particular to a sparse time-frequency analysis method, device and readable storage medium based on self-supervised generative adversarial learning and reconstruction of physical constraints. Background Art

[0002] Time-frequency analysis is a key tool for seismic signal processing and interpretation. It converts one-dimensional, nonstationary seismic signals into a two-dimensional time-frequency spectrum, which can then be used to characterize the localized characteristics of seismic signals. Generally speaking, there are three common time-frequency analysis methods: linear, bilinear, and nonlinear. Linear transformations include a series of transforms, such as the short-time Fourier transform (STFT), continuous wavelet transform (CWT), and S-transform (ST). However, these transforms are limited by the Heisenberg uncertainty principle, which limits the time-frequency resolution and hinders accurate time-frequency analysis of seismic signals. While bilinear transformations offer higher time-frequency resolution, they can also introduce cross-interference terms into the time-frequency representation, hindering seismic signal analysis and interpretation.

[0003] To address the challenges of linear and bilinear transformations, nonlinear transformations have been proposed and widely used in seismic signal processing and interpretation. These primarily include synchronous compression transforms, synchronous extraction transforms, and sparse time-frequency representation methods. Among these methods, the sparse time-frequency representation method incorporates compressed sensing theory to model the time-frequency representation as an inverse problem. A sparse time-frequency representation can then be obtained by adding different regularization terms, and the inverse problem is solved using an iterative algorithm. This method can effectively improve the resolution of the time-frequency representation and obtain better localization features of seismic signals. However, this method often requires a large number of iterations, resulting in low computational efficiency. Furthermore, this method requires extensive prior knowledge to design different regularization terms, which is difficult to obtain in advance for real-world data applications. Furthermore, it introduces numerous parameters that require manual adjustment, and the selection of these parameters often significantly impacts the final time-frequency analysis. These drawbacks result in high design costs and low iteration efficiency for sparse time-frequency transform methods, making them unsuitable for large-scale seismic data applications. Summary of the Invention

[0004] Based on this, it is necessary to provide a sparse time-frequency analysis method, device and readable storage medium based on self-supervised generative adversarial learning and reconstruction of physical constraints to solve the technical problems in the existing technology that a large number of iterations are required, resulting in low computational efficiency and high design cost of sparse time-frequency transformation methods.

[0005] The present invention provides a sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstruction of physical constraints, comprising the following steps:

[0006] A sparse time-frequency analysis network based on a self-supervised generative adversarial network is constructed using a deep learning framework. The sparse time-frequency analysis network includes a generator, a regulator, and a reconstruction module. The reconstruction module has physical meaning and is used to constrain the generator to generate a correct sparse time-frequency spectrum. The regulator is used to monitor whether the sparse time-frequency spectrum generated by the generator is close to the actual sparse time-frequency spectrum.

[0007] Calculating the sparse time-frequency spectrum of synthetic seismic data to construct a data set for the supervisor;

[0008] Selecting a portion of the actual seismic data and merging it with the synthetic seismic data to serve as a data set for the generator;

[0009] Design the generator loss function and the regulator loss function separately;

[0010] Train a sparse time-frequency analysis network model based on a self-supervised generative adversarial network;

[0011] The trained sparse time-frequency analysis network model based on self-supervised generative adversarial network is used for three-dimensional post-stack seismic data to analyze and interpret the geological reservoir structure.

[0012] Furthermore, the step of using a deep learning framework to construct a sparse time-frequency analysis network based on a self-supervised generative adversarial network, wherein the sparse time-frequency analysis network includes a generator, a regulator, and a reconstruction module, includes:

[0013] Based on the image segmentation model, a generator is constructed that can map one-dimensional seismic signals into two-dimensional time-frequency spectra.

[0014] According to the inverse transform formula of short-time Fourier transform, a reconstruction module with physical meaning is constructed;

[0015] Building a supervisor based on sparsity measurement methods and convolutional neural networks.

[0016] Furthermore, the step of calculating the sparse time-frequency spectrum of the synthetic seismic data and constructing the data set of the supervisor includes:

[0017] Synthesize seismic data using convolution models;

[0018] Calculate the sparse time-frequency spectrum corresponding to the synthetic seismic data;

[0019] The sparse time-frequency spectrum set of the synthetic seismic data is used as a training regulator to construct a data set of the regulator.

[0020] Furthermore, the step of selecting a portion of actual seismic data and merging the synthesized seismic data as a data set for the generator includes:

[0021] Select some actual seismic data for normalization;

[0022] The normalized real earthquake data and synthetic earthquake data are used to construct the generator's dataset.

[0023] Furthermore, the steps of respectively designing the generator loss function and the supervisor loss function include:

[0024] specifying optimization constraints for the supervisor;

[0025] The reconstruction module uses short-time Fourier transform inverse transformation to construct a reconstructed physical constraint;

[0026] providing a sparsity constraint using the supervisor;

[0027] The generator loss function is constructed using the reconstructed physical constraint and the sparse constraint.

[0028] Furthermore, the step of training the sparse time-frequency analysis network model based on the self-supervised generative adversarial network includes:

[0029] The regulator is trained using the sparse time-frequency spectrum set, so that the regulator in the sparse time-frequency analysis network based on the self-supervised generative adversarial network can distinguish between the real sparse time-frequency spectrum and the sparse time-frequency spectrum generated by the generator;

[0030] A dataset of a mixture of real and synthetic earthquake data is constructed to serve as the generator for the sparse time-frequency analysis network based on the self-supervised generative adversarial network.

[0031] The generator is trained using a suitable learning rate, number of iterations and loss function to obtain a trained sparse time-frequency analysis network model based on a self-supervised generative adversarial network.

[0032] Furthermore, it also includes dividing the data set of the generator into a training set, a validation set and a test set, and the division ratio is 8:1:1.

[0033] Furthermore, the step of applying the trained sparse time-frequency analysis network model based on the self-supervised generative adversarial network to three-dimensional post-stack seismic data to analyze and interpret the geological reservoir structure includes:

[0034] The actual earthquake data described in the trained sparse time-frequency analysis network model based on the self-supervised generative adversarial network are predicted as a whole, and data volumes of different frequencies are calculated;

[0035] The different frequency data volumes are fused to obtain the RGB fusion result;

[0036] The geological reservoir structure is interpreted through the RGB fusion result analysis.

[0037] The present invention also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-mentioned sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstruction of physical constraints.

[0038] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the above-mentioned sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstruction of physical constraints.

[0039] The present invention provides a sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstruction of physical constraints, which is based on the idea of ​​adversarial generative networks and self-supervised learning. The method introduces a reconstruction module with physical meaning, so that the reconstruction module can constrain the generator to generate a correct sparse time-frequency spectrum. At the same time, a regulator is introduced, which is used to supervise whether the sparse time-frequency spectrum generated by the generator is close to the real sparse time-frequency spectrum. The trained sparse time-frequency analysis network model based on the self-supervised generative adversarial network is used for three-dimensional post-stack seismic data to analyze and interpret the geological reservoir structure. This method can effectively reduce the additional parameter selection brought by traditional methods and reduce the dependence on actual data labels, thereby improving the efficiency of calculation and reducing the design cost of the sparse time-frequency transformation method. The method is suitable for large-scale seismic data applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0041] Figure 1 Schematic diagram of the network structure of self-supervised sparse time-frequency analysis based on generative adversarial networks in an embodiment of the present invention;

[0042] Figure 2 Schematic diagram of the network structure of the generator in an embodiment of the present invention.

[0043] Figure 3 FIG. 4 is a schematic diagram of the network structure of the supervisor in an embodiment of the present invention.

[0044] Figure 4 This is a schematic diagram of the division of real data in an embodiment of the present invention. The shaded part is divided into a test set, and the other part is used for the training process, from which the training set and the validation set are extracted.

[0045] Figure 5Figure 1 shows the effects of actual seismic data in an embodiment of the present invention: (a) an actual seismic track randomly extracted from an actual data set, (b) the time-frequency spectrum result of the traditional short-time Fourier transform (STFT), (c) the sparse time-frequency spectrum result of the traditional sparse short-time Fourier transform (sparse STFT), and (d) the sparse time-frequency spectrum result predicted by the STFR-GAN network.

[0046] Figure 6 These are slices of the time-spectrum obtained by different algorithms in the embodiments of the present invention: (a) and (b) are 20 Hz slices obtained by short-time Fourier transform (STFT) and STFR-GAN respectively, (c) and (d) are 40 Hz slices obtained by short-time Fourier transform (STFT) and STFR-GAN respectively, (e) and (f) are 60 Hz slices obtained by short-time Fourier transform (STFT) and STFR-GAN respectively.

[0047] Figure 7 RGB slices obtained by different algorithms in the embodiments of the present invention: (a) The slice combination obtained by short-time Fourier transform calculation is an RGB slice; (b) The RGB slice composed of frequency slices predicted by the STFR-GAN model.

[0048] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0050] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0051] In addition, the descriptions of "first", "second", etc. in the present invention are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, "and / or" in the full text includes three solutions. Taking A and / or B as an example, it includes technical solution A, technical solution B, and technical solution that satisfies both A and B. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0052] In some embodiments, a sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstructed physical constraints includes the following steps:

[0053] S1. A sparse time-frequency analysis network (STFR-GAN) based on a self-supervised generative adversarial network (GAN) is constructed using a deep learning framework. The sparse time-frequency analysis network includes a generator, a regulator, and a reconstruction module. The reconstruction module has physical meaning and is used to constrain the generator to generate the correct sparse time-frequency spectrum. The regulator is used to supervise whether the sparse time-frequency spectrum generated by the generator is close to the actual sparse time-frequency spectrum. Specifically, in S1, based on the Unet model (a model for image segmentation), a generator network is first constructed that can map one-dimensional seismic signals to two-dimensional time-frequency spectrum. Then, based on the inverse transform formula of the short-time Fourier transform, a physically meaningful reconstruction physical constraint module is constructed. Then, based on the method of measuring sparsity and convolutional neural networks (CNN, a neural network model structure), a regulator network is built.

[0054] S2. Calculate the sparse time-frequency spectrum of the synthetic seismic data to construct a supervisor dataset. Specifically, in S2, the convolution model is used to synthesize seismic data, and the corresponding sparse time-frequency spectrum is calculated. The sparse time-frequency spectrum of the synthetic seismic signal is used to train the supervisor, constructing a supervisor training dataset and uniformly dividing it into a training set.

[0055] S3. Preprocess the actual seismic data: A small amount of actual seismic data is selected and merged with the synthetic seismic data to form the generator's dataset. Specifically, in S3, the actual seismic data is preprocessed: a small amount of actual seismic data is normalized. The processed actual seismic data and synthetic seismic data are then combined to form a dataset for training the generator. The dataset is further divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0056] Normalization involves processing the data (via an algorithm) and limiting it to a desired range. Normalization facilitates subsequent data processing and ensures faster convergence during program execution. Normalization specifically aims to unify the statistical distribution of samples. Normalization between 0 and 1 represents a statistical probability distribution, while normalization within a specific interval represents a statistical coordinate distribution.

[0057] S4: Construct the reconstruction physical constraints and design the loss functions of the generator and regulator respectively. Specifically, in S4, you first need to specify the optimization constraints of the regulator (that is, the regulator loss function). Then, use the inverse short-time Fourier transform to construct the reconstruction physical constraints. At the same time, use the regulator to provide sparse constraints. Finally, the reconstruction physical constraints and sparse constraints are used to form the generator loss function.

[0058] S5. Training of sparse time-frequency analysis network model based on self-supervised generative adversarial network. Specifically, in S5, first, it is necessary to use the sparse time-frequency spectrum set to train the regulator, so that the regulator in the sparse time-frequency analysis network based on self-supervised generative adversarial network (STFR-GAN) can distinguish between the real sparse time-frequency spectrum and the sparse time-frequency spectrum generated by the generator. Secondly, a data set of the generator model in the sparse time-frequency analysis network based on self-supervised generative adversarial network (STFR-GAN) is constructed using a mixture of actual seismic data and synthetic seismic data. Then, the generator is trained using an appropriate learning rate, number of iterations, and loss function to obtain a trained sparse time-frequency analysis network based on self-supervised generative adversarial network (STFR-GAN) model.

[0059] S6. Apply the trained sparse time-frequency analysis network model based on a self-supervised generative adversarial network to 3D post-stack seismic data to analyze and interpret geological reservoir structures. Specifically, the trained sparse time-frequency analysis network model based on a self-supervised generative adversarial network (STFR-GAN) is used to perform an overall prediction of the actual seismic data involved in S3, calculate different frequency data volumes, and perform multi-component fusion to obtain RGB (Red-Green-Blue, treating multiple components as three red, green, and blue channels, fusion into a single RGB image, and performing mixed interpretation) fusion results. This result is used to analyze and interpret geological structures such as river channels.

[0060] In order to better explain the technical solution of the present invention, specific implementation methods are given.

[0061] like Figure 1 As shown in the figure, the sparse time-frequency analysis network model based on the self-supervised generative adversarial network includes a generator, a regulator, and a reconstruction module. G represents the generator, D represents the regulator, and H represents the reconstruction module. The subscripts represent the parameter sets of these modules, while represents the real sparse time-frequency spectrum, that is, the sparse time-frequency spectrum of the seismic synthetic signal, and represents the time-frequency spectrum generated by the generator, which can be regarded as a fake time-frequency spectrum.

[0062] The supervisor H is composed of an inversion formula, which is specifically implemented as

[0063]

[0064] The generator G can be regarded as a nonlinear mapping from a one-dimensional signal to a two-dimensional time-frequency spectrum. Figure 2 The schematic diagram of its network structure is shown. It includes two steps: feature extraction step and feature mapping step. When a signal is input, the feature extraction structure first converts the corresponding signal into a higher-dimensional feature through a dimensionality increase operation. This step first uses a one-dimensional convolution with a kernel size of 3 to map the input one-dimensional signal to a two-dimensional matrix, and then uses a two-dimensional convolution with a kernel size of 3x3 to extract features from the two-dimensional matrix. Figure 2 After extracting the effective features, the second step is to further map the features into the real and imaginary parts of the corresponding time-frequency spectrum, that is, Figure 2 The second half of the structure involves dimensionality reduction of the extracted features through deconvolution with a kernel size of 3x3, gradually mapping high-dimensional features into a time-frequency spectrum matrix. After all convolution operations in the encoder, a nonlinear ReLU (Linear Rectification) activation function is used to enhance the encoder's ability to fit nonlinear transformations. Finally, the resulting time-frequency spectrum is passed through a reconstruction module, yielding a reconstructed signal (i.e., a signal identical to the input signal).

[0065] The supervisor D consists of two branches. The first branch is constructed by the traditional method of evaluating the sparsity of the time spectrum, and the second branch is composed of the convolution module in deep learning. Through the convolution operation with a kernel size of 4x4, the time spectrum features can be effectively extracted. Figure 3 As shown in the figure, a schematic diagram of its network structure is shown. The last two branches pass through a fully connected layer to output a value, which indicates the probability of whether the input time-frequency spectrum is a true sparse time-frequency spectrum.

[0066] Furthermore, a reflection seismic model (a model that simulates seismic signals) was used to generate 4608 channels of signal data, and a sparse time-frequency spectrum was generated as the data set for the monitor. The sparse time-frequency spectrum was calculated as follows:

[0067] Sparse time-frequency transform is proposed based on short-time Fourier transform. For the seismic signal f(n), short-time Fourier transform can be defined as follows:

[0068]

[0069] Where m is the discrete time sample and n is the discrete frequency sample. On this basis, the inverse problem of short-time Fourier transform can be expressed as

[0070]

[0071] Written in matrix form as

[0072] f=Ax (3)

[0073] Where f∈ N×1 is a vector of f(n), is the dictionary generated by g(n). Therefore, based on the theory of compressed sensing, the unknown time-frequency coefficients can be described as the following inverse problem:

[0074]

[0075] in, is the two-norm, and respectively represents the regularization term.

[0076] In order to make the time-frequency sparsity sufficiently sparse, the sparse regularization term is added to formula (4), that is,

[0077]

[0078] Among them, ||m||0 represents the l0 norm, but this is a non-convex function and is very difficult to solve. In order to simplify this model, l1 is widely used. Therefore, in order to increase the smoothness of the generated spectrum, we also add the l2 norm, which becomes the following form

[0079]

[0080] In addition, formula (6) can be solved by Bellman iteration method (an iterative solution method), which is defined as follows:

[0081]

[0082] Through the above method, we can obtain the sparse time-frequency spectrum of the synthetic signal and use it as the real sparse time-frequency spectrum to train the supervisor.

[0083] Furthermore, a small amount of actual earthquake data is selected for normalization processing, and then the processed actual earthquake data and synthetic earthquake data are merged to construct a dataset for training the generator, which is further divided into training set, validation set, and test set.

[0084] like Figure 4 Figure 1 shows a schematic diagram of our data partitioning. The shaded area represents the portion used for testing. Finally, the data is divided into training, validation, and test sets in a ratio of 8:1:1. The data is then uniformly normalized, shuffled, and randomly flipped.

[0085] Furthermore, the optimization constraints of the regulator are specified. The regulator is used to distinguish the real sparse time-frequency spectrum obtained from the signal from the sparse time-frequency spectrum generated by the generator, so the loss function of the regulator is as follows:

[0086]

[0087] Where y is a signal, which means that the distance between the real spectrum and the time-frequency spectrum generated by the generator needs to be maximized. Secondly, the inverse short-time Fourier transform is used to construct the reconstruction physical constraints, and the reconstruction formula is formula (2).

[0088] Finally, the generator needs to generate a sparse time-frequency spectrum that is as realistic as possible and be able to reconstruct the correct sparse time-frequency spectrum. Therefore, the loss function of the generator is

[0089]

[0090] y is a signal, which means that the time-frequency spectrum generated by the generator needs to be close to the real sparse time-frequency spectrum while minimizing the reconstruction loss.

[0091] Furthermore, we used the S3 dataset to train the STFR-GAN model and verify its prediction performance. In this training, we used the Adam optimization algorithm as our network optimizer, taking into account its ability to accelerate convergence. First, we used a sparse time-frequency spectrum set to train the supervisor, allowing the supervisor in the STFR-GAN to distinguish between true sparse time-frequency spectra and those generated by the generator, thus ensuring that the supervisor understands the true sparse time-frequency spectrum. Next, we constructed a dataset for the STFR-GAN generator model using a mixture of real and synthetic seismic data. The generator was then trained to ensure that the model generated by the STFR-GAN increasingly resembled the true sparse time-frequency spectrum. Appropriate learning rates, iterations, and loss functions were used to obtain a trained STFR-GAN model. This resulted in new tools that can be used for real-world seismic signal analysis and processing. Furthermore, to prevent overfitting, we used early stopping to prevent the model from overfitting on the training set.

[0092] Furthermore, we validated the model on the test set. After training the STFR-GAN model, we selected another real data set from the actual dataset to verify the effectiveness of STFR-GAN. Figure 5 (a) is the actual earthquake data, Figure 5 (b) is the data of traditional short-time Fourier transform, Figure 5 (c) is the data of traditional sparse short-time Fourier transform, Figure 5 (d) is the time-frequency result obtained by STFR-GAN. It can be seen that the result calculated by our method is more sparse.

[0093] Furthermore, the trained STFR-GAN model is used to interpret reservoir structure using 3D post-stack seismic data.

[0094] We used the trained STFR-GAN model to predict the overall 3D data and extracted results with main frequencies of 20Hz, 40Hz, and 60Hz. Figure 6 Slices of the time-frequency spectrum calculated by different algorithms: (a) and (b) are 20Hz slices calculated by short-time Fourier transform (STFT) and STFR-GAN, respectively; (c) and (d) are 40Hz slices calculated by short-time Fourier transform (STFT) and STFR-GAN, respectively; (e) and (f) are 60Hz slices calculated by short-time Fourier transform (STFT) and STFR-GAN, respectively. It can be seen that the results obtained by our slice calculation can more clearly describe the river channel information. In addition, we use the traditional STFT and STFR-GAN methods to calculate the results of 20Hz, 40Hz, and 60Hz to form the RGB channels. The results are as follows Figure 7As shown in the figure, (a) is an RGB slice composed of slices calculated by short-time Fourier transform, and (b) is an RGB slice composed of frequency slices predicted by the STFR-GAN model. It can be seen that our method can achieve clearer results than the previous two methods and can provide more river channel details.

[0095] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. All equivalent structural transformations made by using the contents of the present invention description and drawings under the inventive concept of the present invention, or direct / indirect application in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. A sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstruction of physical constraints, characterized by: The following steps are involved: A sparse time-frequency analysis network based on a self-supervised generative adversarial network is constructed using a deep learning framework. The sparse time-frequency analysis network includes a generator, a supervisor, and a reconstruction module, specifically including: Based on the image segmentation model, a generator is constructed that can map one-dimensional seismic signals into two-dimensional time-frequency spectrum. According to the inverse transform formula of short-time Fourier transform, a reconstruction module with physical meaning is constructed. Based on the method of measuring sparsity and building a supervisor with convolutional neural network, The reconstruction module has physical meaning and is used to constrain the generator to generate a correct sparse time-frequency spectrum. The regulator is used to monitor whether the sparse time-frequency spectrum generated by the generator is close to the real sparse time-frequency spectrum. Calculating the sparse time-frequency spectrum of synthetic seismic data to construct a data set for the supervisor; Selecting a portion of actual seismic data and merging it with the synthetic seismic data to serve as a data set for the generator; The generator loss function and the supervisor loss function are designed separately, including: specifying optimization constraints for the supervisor, The reconstruction module uses the inverse short-time Fourier transform to construct the reconstructed physical constraints. Using the supervisor to provide sparsity constraints, Constructing the generator loss function using the reconstruction physical constraint and the sparse constraint; Train a sparse time-frequency analysis network model based on a self-supervised generative adversarial network; The trained sparse time-frequency analysis network model based on self-supervised generative adversarial network is used for three-dimensional post-stack seismic data to analyze and interpret the geological reservoir structure.

2. The sparse time-frequency analysis method according to claim 1, characterized in that: The step of calculating the sparse time-frequency spectrum of the synthetic seismic data and constructing the data set of the supervisor comprises: Synthesize seismic data using convolution models; Calculate the sparse time-frequency spectrum corresponding to the synthetic seismic data; The sparse time-frequency spectrum set of the synthetic seismic data is used as a training regulator to construct a data set of the regulator.

3. The sparse time-frequency analysis method according to claim 2, characterized in that: The step of selecting a portion of actual seismic data and merging the synthesized seismic data as a data set for the generator includes: Select some actual seismic data for normalization; The normalized real earthquake data and synthetic earthquake data are used to construct the generator's dataset.

4. The sparse time-frequency analysis method according to claim 3, characterized in that: The steps of training the sparse time-frequency analysis network model based on the self-supervised generative adversarial network include: The regulator is trained using the sparse time-frequency spectrum set, so that the regulator in the sparse time-frequency analysis network based on the self-supervised generative adversarial network can distinguish between the real sparse time-frequency spectrum and the sparse time-frequency spectrum generated by the generator; A dataset of a mixture of real and synthetic earthquake data is constructed to serve as the generator for the sparse time-frequency analysis network based on the self-supervised generative adversarial network. The generator is trained using a suitable learning rate, number of iterations and loss function to obtain a trained sparse time-frequency analysis network model based on a self-supervised generative adversarial network.

5. The sparse time-frequency analysis method according to claim 4, characterized in that: It also includes dividing the data set of the generator into a training set, a validation set and a test set, with a division ratio of 8:1:

1.

6. The sparse time-frequency analysis method according to claim 5, characterized in that: The step of applying the trained sparse time-frequency analysis network model based on the self-supervised generative adversarial network to three-dimensional post-stack seismic data to analyze and interpret the geological reservoir structure includes: The actual earthquake data described in the trained sparse time-frequency analysis network model based on the self-supervised generative adversarial network are predicted as a whole, and data volumes of different frequencies are calculated; The different frequency data volumes are fused to obtain the RGB fusion result; The geological reservoir structure is interpreted through the RGB fusion result analysis.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstruction of physical constraints are implemented as described in any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the computer program implements the steps of the sparse time-frequency analysis method based on self-supervised generative adversarial learning and reconstruction of physical constraints as described in any one of claims 1 to 6.

Images