Blind deconvolution model, training method, blind deconvolution method, device and medium
By adopting U-Net network and closed-loop deep learning network framework in deconvolution technology, the inversion of sparse coefficients and seismic wavelets is achieved, and the problems of poor adaptability and amplitude retention in the existing technology are solved, manual intervention is reduced, and processing effect is improved.
Patent Information
- Application Number
- CN202311664317.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-06
- Publication Date
- 2025-06-06
Smart Images
Figure CN120106140A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of geophysical exploration technology, and in particular to an intelligent blind deconvolution technology based on a U-Net network, which realizes the inversion of sparse coefficients and seismic wavelets by using a closed-loop deep learning network framework, and can be used for resolution improvement processing and surface consistency deconvolution processing in the process of seismic data processing in petroleum geophysical exploration. Background Art
[0002] Deconvolution is an important technology in seismic data processing. It can be used to improve the resolution of seismic data, correct surface factors in seismic data, achieve surface consistency processing, and eliminate short-period ringing and other multiple wave interference. Deconvolution technology is based on the convolution model and infers the formation reflection coefficient from the observed reflected seismic signal. Since the seismic wavelet is unknown and the adjacent reflected seismic waves overlap, the seismic wavelet is difficult to extract accurately. Therefore, conventional deconvolution is mostly performed under certain assumptions, and such deconvolution processing has the risk of poor adaptability and amplitude preservation.
[0003] Blind deconvolution is a seismic data processing method that uses independent component analysis to perform blind deconvolution based on the randomness and non-Gaussianity of the reflection coefficient sequence. This method weakens the prior conditions for wavelets and reflection coefficients, and overcomes the dependence of traditional deconvolution methods on the assumption of minimum phase wavelets and Gaussian white noise reflection coefficients. Blind deconvolution does not require other assumptions, and takes both seismic wavelets and reflection coefficients as inversion objects. By alternately solving them, the purpose of inverting wavelets and reflection coefficients simultaneously is achieved, which can effectively alleviate the problems of inaccurate wavelet estimation and large noise influence.
[0004] Deep learning algorithms can automatically mine data features, have powerful feature extraction and nonlinear mapping capabilities, and continuously optimize network models through large amounts of data. Once network training is completed, it can be used for batch data processing. Seismic data processing based on artificial intelligence has broad application prospects. However, there is little labeled professional data in the field of seismic data processing, and most network models built purely by data drive have problems of weak applicability and poor stability. Closed-loop neural networks use closed-loop consistency loss functions to constrain the forward and inversion processes in network models, and can introduce physical equation constraints, so that unlabeled data can participate in network model training and improve the promotion and application capabilities of network models.
[0005] Based on the above problems, an intelligent blind deconvolution technology was developed by combining the idea of blind deconvolution with a closed-loop neural network framework. This technology reduces the requirements for the accuracy of seismic wavelet extraction, reduces manual intervention, improves the adaptability and intelligence of deconvolution technology, and better serves high-resolution seismic data processing and surface consistency deconvolution. Summary of the invention
[0006] The purpose of the present invention is to combine the idea of blind deconvolution with a closed-loop deep learning network framework to form an intelligent blind deconvolution technology, while realizing the inversion of sparse coefficients and seismic wavelets, reducing manual intervention, and introducing physical equation constraints in the construction of neural networks to reduce multiple solutions, improve network stability, ensure processing accuracy, and provide more reliable data for high-fidelity seismic imaging and interpretation.
[0007] To achieve the above object, the present invention provides a blind deconvolution model, comprising:
[0008] Neural network, used to predict sparse coefficients;
[0009] A convolutional network is used to perform one-dimensional convolution on the sparse coefficients predicted by the neural network and the convolution kernel representing the seismic wavelet to obtain synthetic seismic data.
[0010] Furthermore, the neural network is a U-Net network model, the input is seismic data, and the output is sparse coefficients;
[0011] The convolution network includes a one-dimensional convolution layer, the length of the convolution kernel is consistent with the length of the initial wavelet, and the weight of the convolution kernel is the amplitude of the initial wavelet.
[0012] Furthermore, the L2 norm of the synthetic seismic data and the real seismic data is defined as the loss function, and the Adam algorithm is used to update the network model by minimizing the loss function.
[0013] According to another aspect of the present invention, a method for training a blind deconvolution model is provided, wherein the training of the blind deconvolution model is performed using a step-by-step alternating strategy:
[0014] First, the weight of the one-dimensional convolution kernel of the convolution network is fixed, and the weight of the neural network model is updated by minimizing the loss function. After the network is trained for N cycles, the weight of the neural network model is fixed and the one-dimensional convolution kernel weight of the convolution network is updated. After the network is trained for M cycles, the above training strategy is repeated until the loss function value tends to stabilize.
[0015] According to another aspect of the present invention, there is provided a blind deconvolution method based on artificial intelligence, using the blind deconvolution model, comprising:
[0016] Inputting the seismic data into the blind deconvolution model to obtain sparse coefficients;
[0017] When processing to improve resolution, the sparse coefficients output by the neural network are low-pass filtered to obtain synthetic seismic data after high-resolution processing;
[0018] When performing surface consistent deconvolution processing, statistical wavelets are first obtained from the estimated wavelets on each seismic data, and then convolved with the sparse coefficients predicted by the neural network to obtain synthetic seismic data after surface consistent deconvolution processing.
[0019] According to another aspect of the present invention, there is provided a blind deconvolution method based on artificial intelligence, comprising:
[0020] Initial wavelet selection: select the corresponding type of wavelet according to the actual seismic data characteristics;
[0021] Network model construction: Build a U-Net network model, with seismic data as input and sparse coefficients as output. The U-Net network output layer is followed by a one-dimensional convolution layer, with the convolution kernel length being consistent with the initial wavelet length, and the convolution kernel weight being the amplitude of the initial wavelet.
[0022] Network model training, using the output after the one-dimensional convolution layer and the L2 norm of the actual seismic data as the loss function, selecting the optimizer and network training parameters, and adopting a step-by-step alternating strategy to perform network model training;
[0023] Blind deconvolution processing is performed to input the seismic data into the above-mentioned trained network model to obtain synthetic seismic data.
[0024] Furthermore, the network model training includes:
[0025] The weight of the one-dimensional convolution kernel representing the seismic wavelet is fixed, and the U-Net model weight is updated by minimizing the loss function. After the network is trained for N cycles, the U-Net model weight is fixed and the one-dimensional convolution kernel weight is updated. After the network is trained for M cycles, the above training strategy is repeated until the loss function value tends to be stable.
[0026] Furthermore, the blind deconvolution process includes:
[0027] High-resolution processing: low-pass filtering the sparse coefficients output by the U-Net network to obtain synthetic seismic data after high-resolution processing;
[0028] Surface consistent deconvolution processing is used to perform statistical processing on the seismic wavelets extracted from different seismic channels to obtain statistical wavelets, and then the statistical wavelets are convolved with the sparse coefficients predicted by the U-Net network to obtain reconstructed seismic records and synthetic seismic data processed by surface consistent deconvolution processing.
[0029] According to another aspect of the present invention, there is provided an electronic device, the electronic device comprising:
[0030] A memory storing executable instructions;
[0031] A processor runs the executable instructions in the memory to implement the artificial intelligence-based blind deconvolution method.
[0032] According to another aspect of the present invention, there is provided a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any one of the artificial intelligence-based blind deconvolution methods.
[0033] The present invention provides a blind deconvolution technology based on artificial intelligence. Compared with the prior art, the present invention has the following beneficial effects: intelligent blind deconvolution is realized through a closed-loop neural network and the idea of blind deconvolution, manual intervention is reduced, and in particular, the requirements for seismic wavelet extraction accuracy in traditional methods are avoided, the technology has a certain amplitude preservation, and can be used for surface consistency deconvolution and high-resolution processing. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] 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 in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.
[0035] Figure 1 The figure is a flow chart of the blind deconvolution method based on artificial intelligence according to the present invention.
[0036] Figure 2 Schematic diagram of a blind deconvolution model according to an embodiment of the present invention.
[0037] Figure 3 Schematic diagram of a U-Net network model according to an embodiment of the present invention.
[0038] Figure 4 The figure is a flowchart of the artificial intelligence-based blind deconvolution technology according to an embodiment of the present invention.
[0039] Figure 5 This is a test result diagram for simulated seismic data according to an embodiment of the present invention. Figure 5 a is the real earthquake data, Figure 5 b is the sparse coefficient inversion result, Figure 5 c is the seismic data output by the network, Figure 5 d is the seismic wavelet inversion result. DETAILED DESCRIPTION
[0040] The preferred embodiments of the present invention will be described in more detail below. Although the preferred embodiments of the present invention are described below, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.
[0041] The present invention relates to an intelligent blind deconvolution technology based on a U-Net network, and uses a closed-loop deep learning network framework and the idea of blind deconvolution to build an intelligent deconvolution network model. A step-by-step alternating strategy is used to update the weights of the U-Net network model and the one-dimensional convolution kernel that characterizes the seismic wavelet, and simultaneously realizes the inversion of sparse coefficients and seismic wavelets, avoiding the requirements for seismic wavelet extraction accuracy in traditional methods and reducing manual intervention. Physical equation constraints reduce multi-solutions, improve network stability, and ensure processing accuracy. It can be used for surface consistency deconvolution and high-resolution processing, and provide more reliable data for seismic high-fidelity imaging and interpretation.
[0042] Embodiment 1
[0043] Figure 1 FIG. 1 is a flow chart of a blind deconvolution method based on artificial intelligence according to the present invention, as shown in FIG. Figure 1 As shown, this embodiment provides a blind deconvolution method based on artificial intelligence, including:
[0044] Initial wavelet selection: select the corresponding type of wavelet according to the actual seismic data characteristics;
[0045] Network model construction: Build a U-Net network model, with seismic data as input and sparse coefficients as output. The U-Net network output layer is followed by a one-dimensional convolution layer, with the convolution kernel length being consistent with the initial wavelet length, and the convolution kernel weight being the amplitude of the initial wavelet.
[0046] Network model training, using the output after the one-dimensional convolution layer and the L2 norm of the actual seismic data as the loss function, selecting the optimizer and network training parameters, and adopting a step-by-step alternating strategy to perform network model training;
[0047] Blind deconvolution processing is performed to input the seismic data into the above-mentioned trained network model to obtain synthetic seismic data.
[0048] Specifically, initial wavelet selection: the initial wavelet is usually selected as a Ricker wavelet with a main frequency of f, where the main frequency f is determined by the seismic data to be processed. The corresponding type of wavelet can also be selected based on the characteristics of the actual seismic data.
[0049] Specifically, the network model is built: a U-Net network model is built, the input is seismic data, the output is sparse coefficients, the U-Net network output layer is followed by a one-dimensional convolution layer, the convolution kernel length is consistent with the initial wavelet length, and the weight of the convolution kernel is the amplitude of the initial wavelet.
[0050] Specifically, network model training: the output after the one-dimensional convolution layer and the L2 norm of the actual seismic data are used as the loss function, the appropriate optimizer and network training parameters are selected, and the network model training is carried out using a step-by-step alternating strategy. First, the weight of the one-dimensional convolution kernel that represents the seismic wavelet is fixed, and the U-Net model weight is updated by minimizing the loss function. After the network is trained for N cycles, the U-Net model weight is fixed, and the one-dimensional convolution kernel weight is updated. After the network is trained for M cycles, the above training strategy is repeated until the loss function value tends to be stable, and usually N is greater than M.
[0051] Specifically, the blind deconvolution process may include high-resolution oriented processing and surface consistent deconvolution process.
[0052] Towards high-resolution processing: After completing the above three steps on the seismic profile data, the sparse coefficients output by the neural network are low-pass filtered to obtain high-resolution processed seismic data.
[0053] Surface consistent deconvolution processing: After completing the above three steps on the common shot point records or common detection point records, statistical processing is performed on the seismic wavelets extracted by the neural network in different seismic channels to obtain statistical wavelets, and then the statistical wavelets are convolved with the sparse coefficients predicted by the neural network to obtain reconstructed seismic records, thereby realizing seismic data after surface consistent deconvolution processing.
[0054] Embodiment 2
[0055] Figure 2 Schematic diagram of a blind deconvolution model according to an embodiment of the present invention. Figure 2 As shown, this embodiment provides a blind deconvolution model, including:
[0056] Neural network, used to predict sparse coefficients;
[0057] A convolutional network is used to perform one-dimensional convolution on the sparse coefficients predicted by the neural network and the convolution kernel representing the seismic wavelet to obtain synthetic seismic data.
[0058] Specifically, the neural network is a U-Net network model, the input is seismic data, and the output is a sparse coefficient; the convolution network includes a one-dimensional convolution layer, the convolution kernel length is consistent with the initial wavelet length, and the weight of the convolution kernel is the amplitude of the initial wavelet.
[0059] The U-Net network is used to predict the sparse coefficient from seismic data. The network structure is as follows: Figure 3 As shown in the figure, the network consists of four convolutional layers and four deconvolutional layers, and uses skip-layer connections to maintain the original data characteristics. In addition, in order to ensure the sparsity of the network output results, a classification module is added to the network output part, and the classification results are used to constrain the sparse coefficients predicted by the neural network to improve the network performance. The sparse coefficients output by the neural network are convolved one-dimensionally with the convolution kernel representing the seismic wavelet to obtain synthetic seismic data.
[0060] In this embodiment, the loss function is defined as the L2 norm of the synthetic seismic data and the real seismic data. The Adam algorithm is used to minimize the loss function to achieve the optimization of the inversion network model and the continuous updating of the seismic wavelet.
[0061] Embodiment 3
[0062] This embodiment provides a training method for a blind deconvolution model, which adopts a step-by-step alternating strategy to train the blind deconvolution model.
[0063] Specifically, the output after the one-dimensional convolution layer and the L2 norm of the actual seismic data are used as the loss function, and the appropriate optimizer and network training parameters are selected, and the network model training is carried out using a step-by-step alternating strategy. First, the weight of the one-dimensional convolution kernel that represents the seismic wavelet is fixed, and the U-Net model weight is updated by minimizing the loss function. After the network is trained for N cycles, the U-Net model weight is fixed, and the one-dimensional convolution kernel weight is updated. After the network is trained for M cycles, the above training strategy is repeated until the loss function value tends to be stable, and usually N is greater than M.
[0064] The output of the trained neural network is the inverted sparse coefficient, and the convolution kernel used for synthetic seismic data calculation in the corresponding convolution network is the inverted seismic wavelet. This technology realizes the simultaneous inversion of sparse coefficients and seismic wavelets.
[0065] Embodiment 4
[0066] Figure 4 FIG. 1 is a flowchart of a blind deconvolution technique based on artificial intelligence according to an embodiment of the present invention. Figure 4 As shown, this embodiment is aimed at seismic deconvolution. In view of the problems that conventional deconvolution is difficult to achieve amplitude-preserving processing and the difficulty in accurately obtaining wavelets, an intelligent blind deconvolution technology based on U-Net is proposed. The neural network model is used to simultaneously realize the inversion of sparse coefficients and seismic wavelets, and physical equation constraints are introduced to reduce multiple solutions, thereby realizing intelligent seismic deconvolution processing.
[0067] The physical equation on which seismic deconvolution is based is the convolution model, that is, seismic data is the convolution of sparse coefficients and seismic wavelets (Formula 1). This process is similar to the convolution layer in a convolutional neural network. Therefore, the seismic wavelet is regarded as a convolution kernel and added to the construction of the seismic data deconvolution network. The process of continuous training and optimization of the network model is the process of updating and solving the seismic wavelet.
[0068] s(t)=r(t)*w(t) (1)
[0069] Among them, s(t) is the seismic record, r(t) is the sparse coefficient, and w(t) is the seismic wavelet.
[0070] The U-Net network is used to predict the sparse coefficient from seismic data. The network structure is as follows: Figure 3 As shown in Figure 1, the network consists of four convolutional layers and four deconvolutional layers, and uses skip-layer connections to maintain the original data features. In addition, in order to ensure the sparsity of the network output results, a classification module is added to the network output part, and the classification results are used to constrain the sparse coefficients predicted by the neural network to improve network performance.
[0071] The sparse coefficients output by the U-Net network are convolved in one dimension with the convolution kernel representing the seismic wavelet to obtain synthetic seismic data. In this embodiment, the loss function is defined as the L2 norm of the synthetic seismic data and the real seismic data. The Adam algorithm is used to minimize the loss function to achieve the optimization of the blind deconvolution model and the continuous updating of the seismic wavelet.
[0072] When processing to improve resolution, the sparse coefficients output by the U-Net network are low-pass filtered to obtain seismic data after high-resolution processing; when processing for surface consistent deconvolution, the statistical wavelet is obtained from the estimated wavelet on each seismic data, and convolved with the sparse coefficients predicted by the U-Net network to achieve surface consistent deconvolution processing.
[0073] Embodiment 5
[0074] Figure 5 This is a test result diagram for simulated seismic data according to an embodiment of the present invention. Figure 5 a is the real earthquake data, Figure 5 b is the sparse coefficient inversion result, Figure 5 c is the seismic data output by the network, Figure 5 d is the seismic wavelet inversion result.
[0075] In this embodiment, the artificial intelligence-based deconvolution method of the present invention was tested on simulated single-shot seismic records. Figure 5 b and 5d show the sparse coefficients predicted by the trained network model and the seismic wavelet represented by the trained one-dimensional convolution kernel, respectively. Figure 5a and 5c are the real seismic data and the seismic data output by the network, respectively. The two are basically consistent, and the number of main event axes in the sparse coefficient profile is consistent with the original seismic data. This application shows that the intelligent blind deconvolution method can effectively estimate the sparse coefficients and obtain seismic wavelets, which can be used for subsequent surface consistency processing.
[0076] Embodiment 6
[0077] This embodiment provides an electronic device, the electronic device comprising:
[0078] A memory storing executable instructions;
[0079] A processor, wherein the processor runs the executable instructions in the memory to implement the artificial intelligence-based blind deconvolution method, the method comprising:
[0080] Initial wavelet selection: select the corresponding type of wavelet according to the actual seismic data characteristics;
[0081] Network model construction: Build a U-Net network model, with seismic data as input and sparse coefficients as output. The U-Net network output layer is followed by a one-dimensional convolution layer, with the convolution kernel length being consistent with the initial wavelet length, and the convolution kernel weight being the amplitude of the initial wavelet.
[0082] Network model training, using the output after the one-dimensional convolution layer and the L2 norm of the actual seismic data as the loss function, selecting the optimizer and network training parameters, and adopting a step-by-step alternating strategy to perform network model training;
[0083] Blind deconvolution processing is performed to input the seismic data into the above-mentioned trained network model to obtain synthetic seismic data.
[0084] Embodiment 7
[0085] This embodiment provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the blind deconvolution method based on artificial intelligence is implemented. The method includes:
[0086] Initial wavelet selection: select the corresponding type of wavelet according to the actual seismic data characteristics;
[0087] Network model construction: Build a U-Net network model, with seismic data as input and sparse coefficients as output. The U-Net network output layer is followed by a one-dimensional convolution layer, with the convolution kernel length being consistent with the initial wavelet length, and the convolution kernel weight being the amplitude of the initial wavelet.
[0088] Network model training, using the output after the one-dimensional convolution layer and the L2 norm of the actual seismic data as the loss function, selecting the optimizer and network training parameters, and adopting a step-by-step alternating strategy to perform network model training;
[0089] Blind deconvolution processing is performed to input the seismic data into the above-mentioned trained network model to obtain synthetic seismic data.
[0090] 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).
[0091] In summary, the present invention uses a closed-loop deep learning network framework and blind deconvolution ideas to build an intelligent blind deconvolution network model. The step-by-step alternating strategy is used to update the weights of the U-Net network model and the one-dimensional convolution kernel that characterizes the seismic wavelet, while realizing the inversion of sparse coefficients and seismic wavelets, avoiding the requirements for seismic wavelet extraction accuracy in traditional methods and reducing manual intervention. Physical equation constraints reduce multi-solutions, improve network stability, and ensure processing accuracy. It can be used for surface consistency deconvolution and high-resolution processing, providing more reliable data for seismic high-fidelity imaging and interpretation.
[0092] 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 blind deconvolution model, It is characterized in that include: Neural network, used to predict sparse coefficients; A convolutional network is used to perform one-dimensional convolution on the sparse coefficients predicted by the neural network and the convolution kernel representing the seismic wavelet to obtain synthetic seismic data.
2. The blind deconvolution model according to claim 1, It is characterized in that The neural network is a U-Net network model, the input is seismic data, and the output is sparse coefficients; The convolution network includes a one-dimensional convolution layer, the length of the convolution kernel is consistent with the length of the initial wavelet, and the weight of the convolution kernel is the amplitude of the initial wavelet.
3. The blind deconvolution model according to claim 2, It is characterized in that The L2 norm of the synthetic seismic data and the real seismic data is defined as the loss function, and the Adam algorithm is used to update the network model by minimizing the loss function.
4. A method for training a blind deconvolution model as claimed in any one of claims 1 to 3, It is characterized in that The blind deconvolution model is trained using a step-by-step alternating strategy: First, the weight of the one-dimensional convolution kernel of the convolution network is fixed, and the weight of the neural network model is updated by minimizing the loss function. After the network is trained for N cycles, the weight of the neural network model is fixed and the one-dimensional convolution kernel weight of the convolution network is updated. After the network is trained for M cycles, the above training strategy is repeated until the loss function value tends to stabilize.
5. A blind deconvolution method based on artificial intelligence, It is characterized in that Using the blind deconvolution model described in any one of claims 1 to 3, comprising: Inputting the seismic data into the blind deconvolution model to obtain sparse coefficients; When processing to improve resolution, the sparse coefficients output by the neural network are low-pass filtered to obtain synthetic seismic data after high-resolution processing; When performing surface consistent deconvolution processing, statistical wavelets are first obtained from the estimated wavelets on each seismic data, and then convolved with the sparse coefficients predicted by the neural network to obtain synthetic seismic data after surface consistent deconvolution processing.
6. A blind deconvolution method based on artificial intelligence, It is characterized in that include: Initial wavelet selection: select the corresponding type of wavelet according to the actual seismic data characteristics; Network model construction: Build a U-Net network model, with seismic data as input and sparse coefficients as output. The U-Net network output layer is followed by a one-dimensional convolution layer, with the convolution kernel length being consistent with the initial wavelet length, and the convolution kernel weight being the amplitude of the initial wavelet. Network model training, using the output after the one-dimensional convolution layer and the L2 norm of the actual seismic data as the loss function, selecting the optimizer and network training parameters, and adopting a step-by-step alternating strategy to perform network model training; Blind deconvolution processing is performed to input the seismic data into the above-mentioned trained network model to obtain synthetic seismic data.
7. The artificial intelligence-based blind deconvolution method according to claim 6, It is characterized in that The network model training includes: The weight of the one-dimensional convolution kernel representing the seismic wavelet is fixed, and the U-Net model weight is updated by minimizing the loss function. After the network is trained for N cycles, the U-Net model weight is fixed and the one-dimensional convolution kernel weight is updated. After the network is trained for M cycles, the above training strategy is repeated until the loss function value tends to be stable.
8. The artificial intelligence-based blind deconvolution method according to claim 6, It is characterized in that The blind deconvolution process comprises: High-resolution processing: low-pass filtering the sparse coefficients output by the U-Net network to obtain synthetic seismic data after high-resolution processing; Surface consistent deconvolution processing is used to perform statistical processing on the seismic wavelets extracted from different seismic channels to obtain statistical wavelets, and then the statistical wavelets are convolved with the sparse coefficients predicted by the U-Net network to obtain reconstructed seismic records and synthetic seismic data processed by surface consistent deconvolution processing.
9. An electronic device, It is characterized in that The electronic device comprises: A memory storing executable instructions; A processor, wherein the processor runs the executable instructions in the memory to implement the artificial intelligence-based blind deconvolution method according to any one of claims 6 to 8.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the artificial intelligence-based blind deconvolution method according to any one of claims 6 to 8 is implemented.