A seismic facies identification method and system based on cross-correlation convolutional network
By combining cross-correlation convolutional networks with generative adversarial networks and PSP-Unet neural networks, the problems of poor model recognition performance for images of different sizes and insufficient data samples in seismic phase identification are solved, and more efficient seismic phase identification is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA NAT OFFSHORE OIL CORP
- Filing Date
- 2023-09-20
- Publication Date
- 2026-05-19
AI Technical Summary
Existing AI-based seismic phase recognition methods have poor performance in recognizing images of different sizes. The limited number of seismic data samples leads to low accuracy in training models. Conventional neural networks lack linear feature extraction and have slow convergence speed, and they cannot converge with noisy data.
A seismic phase identification method based on cross-correlation convolutional networks is adopted. By combining a committee mechanism model with generative adversarial networks and PSP-Unet neural networks, the seismic dataset is expanded, seismic features are extracted, and accurate identification is performed.
It improves the accuracy and convergence speed of seismic facies identification, enhances the extraction of seismic features at different scales, and solves the problems of timeliness and accuracy in seismic facies identification.
Smart Images
Figure CN117348088B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a seismic phase identification method and system based on cross-correlation convolutional networks, belonging to the field of geophysical exploration technology that combines artificial intelligence algorithms. Background Technology
[0002] Seismic facies are the sum of sedimentary facies manifested on a seismic profile. They are seismic characteristics formed by the sedimentary environment and refer to seismic reflection units within a certain area. The seismic attribute parameters of a seismic reflection unit differ from those of adjacent units, and it can represent the lithology of the sediments that produced the reflection. Seismic facies analysis, based on the division of seismic sequences, utilizes differences in seismic parameter characteristics to divide the seismic sequence into different seismic facies zones, and then makes inferences about lithofacies and sedimentary environments. Existing seismic facies interpretation processes are very complex, and their accuracy depends heavily on the professional knowledge and experience of the interpreters. Different interpreters often arrive at interpretations that differ, and even the same interpreter may arrive at interpretations at different times. Furthermore, seismic facies interpretation involves a large amount of data, requiring significant time and effort, and lacks timeliness.
[0003] With the development of artificial intelligence, the application of AI algorithms for seismic facies identification has become widespread. For example, semantic segmentation techniques from computer vision are combined with seismic data analysis for seismic facies identification. However, existing AI-based seismic facies identification methods have several problems: current models typically perform well only on images of a single size, while their performance is poor on other sizes; the limited number of seismic data samples in seismic facies identification leads to low accuracy in the trained models; conventional neural network models lack the ability to extract linear features and have slow convergence speeds; and pooling layers in neural network models can cause convergence failures with noisy data. Summary of the Invention
[0004] To address the aforementioned problems, the purpose of this invention is to provide a seismic phase identification method and system based on cross-correlation convolutional networks that can expand seismic phase training data.
[0005] To achieve the above objectives, the present invention proposes the following technical solution: a seismic facies identification method based on a cross-correlation convolutional network, comprising the following steps: obtaining well logging data and seismic data; inputting the well logging data into a committee mechanism model to perform well logging lithofacies identification on the well logging data; performing morphological processing on the seismic data; inputting the well logging lithofacies identification results and the morphologically processed seismic data into a generative adversarial network (GAN) to expand the sample dataset through the GAN; and inputting the original seismic data and the expanded sample dataset into a PSP-Unet neural network model to generate seismic facies identification results.
[0006] Furthermore, the committee mechanism model consists of three expert models: K-nearest neighbor, BP neural network, and support vector machine. The optimal solution of the weights of the three expert models is randomly searched using the Naive Bayes algorithm. Based on the optimal solution, the output results of the three expert models are weighted and summed to obtain the output result of the committee mechanism model.
[0007] Furthermore, the method for morphological processing of the seismic data is as follows: extracting the texture features of the morphological processing model, convolving the seismic profile using structuring elements, smoothing the boundaries of larger features in the morphological processing model, and separating features at smaller features; inputting the seismic data into the morphological processing model, and performing an opening operation on the seismic data through the mutual coupling operation of erosion and dilation to obtain morphologically processed seismic data.
[0008] Furthermore, the generative adversarial network includes a generative model and a discriminative model. Random noise is input into the generator to generate a noise signal. The noise signal and seismic data are input into the discriminative model, which is used to determine the authenticity of the seismic signal. Both the generative model and the discriminative model are composed of several convolutional networks with fully connected layers removed.
[0009] Furthermore, the output layer of the generative model uses the Tanh function as the activation function, and the remaining layers in the generative adversarial network all use the ReLU function as the activation function. The generative adversarial network is a conditional generative adversarial network, which learns a structured loss function to penalize the construction of output nodes.
[0010] Furthermore, the conditional generative adversarial network includes: a first generator, a second generator, a first discriminator, and a second discriminator. The first generator concatenates seismic information with random noise to generate a first vector, which is then input into the first generator to generate a seismic phase for judgment. The known seismic phase is concatenated with random noise, and the resulting second vector, along with the judged seismic phase, is input into the first discriminator for discrimination. The second vector is then input into the second generator to generate seismic data, and the generated seismic data, along with the first vector, is input into the second discriminator for discrimination.
[0011] Furthermore, the PSP-Unet neural network model includes a cross-correlation U-Net convolutional network model and a pyramid pooling model connected in sequence. Seismic data is input into the cross-correlation U-Net convolutional network model, and the output of the cross-correlation U-Net convolutional network model is input into the pyramid pooling model to generate seismic phase prediction results. The output layer of the cross-correlation U-Net convolutional network model and the output layer of the pyramid pooling model are connected across layers.
[0012] Furthermore, the cross-correlation U-Net convolutional network model incorporates a cross-correlation function into the downsampling layer of the U-Net convolutional network model.
[0013] Furthermore, the pyramid pooling model includes several pyramid pooling modules of different scales. The output of the cross-correlation U-Net convolutional network model is preprocessed through a deep neural network, and features are extracted using a network expansion strategy. Global priors are used to classify features at different levels. Low-dimensional features are upsampled using 1×1 convolutional layers, and the size of the features obtained through bilinear interpolation is compared with the initial features. Figure 1 Finally, all pyramid pooling feature outputs are concatenated into the final complete output using a fully connected layer.
[0014] This invention also discloses a seismic facies identification system based on a cross-correlation convolutional network, comprising: a data acquisition module for acquiring well logging data and seismic data; a well logging lithofacies identification module for inputting the well logging data into a committee mechanism model to perform well logging lithofacies identification on the well logging data; a morphological processing module for performing morphological processing on the seismic data; a dataset expansion module for inputting the well logging lithofacies identification results and the morphologically processed seismic data into a generative adversarial network to expand the sample dataset through the generative adversarial network; and an identification result generation module for inputting the original seismic data and the expanded sample dataset into a PSP-Unet neural network model to generate seismic facies identification results.
[0015] The present invention has the following advantages due to the adoption of the above technical solutions:
[0016] 1. This invention integrates a pyramid scene parsing network into the U-Net network to distinguish seismic facies labels at different scales, improving the accuracy of seismic facies identification. It enhances the extraction of small-scale seismic features. By maintaining the weights of global features and taking into account large-scale facies features, it can produce more accurate seismic facies segmentation results.
[0017] 2. This invention utilizes a conditional generative adversarial network (GAN) structure to learn the distribution characteristics of data in the earthquake training set, thereby expanding the earthquake facies training set data. Morphological processing effectively enhances the features of the earthquake data. Adding a block-based GAN to the structure allows for the extraction of more global features from earthquake profiles, ensuring that the generated sample data conforms to the distribution patterns of earthquake data, thus completing the expansion of the earthquake facies training set data.
[0018] 3. In this invention, the expert members introduced by the committee mechanism model can possess both linear and nonlinear features in the well logging data. The committee mechanism model consists of three expert models: K-nearest neighbor method, BP neural network and support vector machine model. Among them, the K-nearest neighbor algorithm reduces the influence of non-feature parameters in the well logging curve by increasing the K value, the BP neural network is a shallow structure with a single hidden layer, and the support vector machine model avoids the classification model from getting trapped in local minima by randomly selecting and calculating weights using the least squares method, thereby increasing the classification accuracy and convergence speed of the committee mechanism model.
[0019] 4. This invention uses batch normalization in both the generative and discriminative models, which helps the deep convolutional network partially converge, but it can cause sampling oscillations. By replacing spatial pooling layers with convolutional layers, the invention avoids permanently discarding values at certain locations in the image and instead uses a self-learning downsampling method of the deep convolutional network, thereby solving the problem of generative adversarial networks failing to converge when generating noisy data. Attached Figure Description
[0020] Figure 1 This is a flowchart of a seismic phase identification method based on a cross-correlation convolutional network according to an embodiment of the present invention;
[0021] Figure 2 This is a schematic diagram of the structure of a committee mechanism model in one embodiment of the present invention;
[0022] Figure 3 This is a logging lithofacies identification map obtained by acquiring logging data from a specific well, namely well 03-4, in one embodiment of the present invention;
[0023] Figure 4 This is a schematic diagram of morphologically processed seismic data in one embodiment of the present invention; Figure 4 (a) is seismic data that has undergone corrosion treatment. Figure 4 (b) is the expanded seismic data;
[0024] Figure 5 A schematic diagram of the structure of a generative adversarial network in one embodiment of the present invention;
[0025] Figure 6 A schematic diagram of the structure of a conditional generative adversarial network in one embodiment of the present invention;
[0026] Figure 7 A schematic diagram of the structure of the PSP-Unet neural network model in one embodiment of the present invention;
[0027] Figure 8 This is a schematic diagram of the structure of a cross-correlation U-Net convolutional network model in one embodiment of the present invention;
[0028] Figure 9This is a schematic diagram of the pyramid pooling model in one embodiment of the present invention;
[0029] Figure 10 This is a well-connected profile and corresponding logging lithofacies distribution map of wells F03-2 and F06-1 in one embodiment of the present invention. Detailed Implementation
[0030] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention is described in detail through specific embodiments. However, it should be understood that the specific embodiments are provided only for a better understanding of the present invention and should not be construed as limiting the present invention. In the description of the present invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0031] To address the limitations of existing artificial intelligence algorithms, such as the accuracy of models being highly dependent on the amount of training data (e.g., models with limited seismic data exhibit lower accuracy), and the significant size variations in seismic images leading to poor recognition performance for smaller images, this invention proposes a seismic facies identification method and system based on a cross-correlation convolutional network. Addressing the limitations of the U-Net network in seismic facies identification, a two-dimensional expression for the cross-correlation function is derived, and pyramid pooling layers are incorporated to mitigate these limitations. Furthermore, a morphological generative adversarial network (GAN) method is used to expand the seismic dataset. Digital morphological processing is applied to the model data to extract features. The GAN structure is improved, and the model data is compared to generate the optimal loss function combination. The GAN is then applied to a real-world work area to expand the seismic dataset, providing training data for seismic facies classification. Finally, well logging curves are automatically classified using a committee machine. The well logging facies classification results are combined with well sequence logging to create a seismic lithofacies model, which is then input into the GAN to learn distribution patterns and further expand the seismic lithofacies. It expands the dataset, increasing the accuracy and convergence speed of the various models involved. The invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0032] Example 1
[0033] This implementation discloses a seismic phase identification method based on cross-correlation convolutional networks, such as... Figure 1 As shown, it includes the following steps:
[0034] S1 obtained well logging data and seismic data.
[0035] S2 inputs the logging data into the committee mechanism model to perform logging lithofacies identification on the logging data.
[0036] like Figure 2As shown, the expert models of the committee mechanism model can accommodate both linear and nonlinear characteristics. The committee mechanism model consists of three expert models: K-nearest neighbor, BP neural network, and support vector machine. The K-nearest neighbor algorithm reduces the influence of non-feature parameters in the logging curve by increasing the K value, and the BP neural network is a shallow structure with a single hidden layer. The support vector machine model avoids the classification model from getting trapped in local minima by randomly selecting weights and using least squares to calculate weights, thereby increasing the classification accuracy and convergence speed of the committee mechanism model. A weighted average method is used as the combination strategy for the committee mechanism model. Specifically, the optimal solution for the weights of the three expert models is randomly searched using the Naive Bayes algorithm. Based on the optimal solution, the outputs of the three expert models are weighted and summed to obtain the output of the committee mechanism model.
[0037] Figure 3 This is a well logging lithofacies identification map obtained by acquiring well logging data from a specific well, namely well 03-4, in one embodiment of the present invention; the gamma value corresponding to the depth of the river delta sediment changes from low to high, reflecting the difference between coarse-grained sandstone and fine-grained sandstone, and the distribution of gas-bearing sandstone in the well logging lithofacies classification is relatively accurate.
[0038] S3 performs morphological processing on seismic data.
[0039] like Figure 4 As shown, the method for morphological processing of seismic data is as follows: extract the texture features of the morphological processing model, use the structuring element to convolve the seismic profile, eliminate the influence of small features in the morphological processing model, smooth the boundaries of larger features in the morphological processing model, and separate features at small features; input the seismic data into the morphological processing model, and perform an opening operation on the seismic data through the mutual coupling operation of erosion and dilation to obtain morphologically processed seismic data. This seismic data has more obvious geological features, clearer inner layer interfaces of the seismic profile, and more obvious edge features at faults.
[0040] S4 inputs well logging lithofacies identification results and morphologically processed seismic data into a generative adversarial network (GAN) to expand the sample dataset.
[0041] To address the issue that generative adversarial networks (GANs) tend to generate useless noisy data and cannot consistently generate images that conform to the distribution patterns of real samples, the convolutional neural network (CNN) structure can be extended to GANs. By simply deepening the convolutional layers, this network structure can successfully generate higher-quality data. This network structure eliminates the fully connected layers of the convolutional network, connecting the highest-level convolutional features with the input and output of the generative and discriminative models. For example... Figure 5As shown, the Generative Adversarial Network (GAN) comprises a generative model and a discriminative model. Random noise is input into the generator to produce a noise signal. This noise signal, along with seismic data, is input into the discriminative model, which determines the authenticity of the seismic signal. Both the generative and discriminative models consist of several convolutional networks with fully connected layers removed. The output layer of the generative model uses the Tanh function as the activation function, while the remaining layers in the GAN use the ReLU function. Batch normalization is used in both the generative and discriminative models to aid partial convergence of the deep convolutional network, but this can cause sampling oscillations. By replacing spatial pooling layers with convolutional layers, the system no longer discards values at certain locations in the image. Instead, it uses a self-learning downsampling method employed by the deep convolutional network, thus addressing the convergence problem caused by the GAN generating noisy data.
[0042] like Figure 6 As shown, the generative adversarial network is a conditional generative adversarial network (CGAN), typically described as a classification problem of individual samples. The output space is considered "unstructured," with each output point independent of all neighboring samples in the input data. CGAN learns a structured loss function to penalize the construction of output nodes. Various loss functions exist for CGAN, such as conditional random fields, cross-correlation simulation metrics, feature matching, nonparametric loss functions, and loss functions based on matching covariance statistics. The CGAN structure can penalize any structure that differs from the output and label, thereby optimizing the loss function. Based on the CGAN structure, the characteristics of the data distribution in the seismic training set are learned to expand the seismic facies training set data. Morphological processing effectively enhances the seismic data features. Adding a block-based CGAN to the CGAN structure can extract more global features from seismic profiles, making the generated sample data conform to the distribution patterns of seismic data, thus completing the expansion of the seismic facies training set data.
[0043] The conditional generative adversarial network includes: a first generator, a second generator, a first discriminator, and a second discriminator. Seismic information is concatenated with random noise to generate a first vector, which is then input into the first generator to generate a seismic phase for judgment. The known seismic phase is concatenated with random noise, and the resulting second vector, along with the judged seismic phase, is input into the first discriminator for discrimination. The second vector is input into the second generator to generate seismic data, and the generated seismic data, along with the first vector, is input into the second discriminator for discrimination.
[0044] S5 inputs the raw seismic data and the expanded sample dataset into the PSP-Unet neural network model to generate seismic facies identification results.
[0045] like Figure 7As shown, the PSP-Unet neural network model consists of a cross-correlation U-Net convolutional network model and a pyramid pooling model connected sequentially. Seismic data is input into the cross-correlation U-Net convolutional network model, and the output of the cross-correlation U-Net convolutional network model is input into the pyramid pooling model to generate seismic facies prediction results. The output layers of the cross-correlation U-Net convolutional network model and the pyramid pooling model are connected across layers. Integrating a pyramid scene resolution network into the U-Net network can distinguish seismic facies labels at different scales, improving the accuracy of seismic facies identification. It can enhance the extraction of small-scale seismic features. By maintaining the weights of global features and taking into account large-scale facies features, it can produce more accurate seismic facies segmentation results.
[0046] Considering the characteristics of lateral correlation variation in seismic phases, and referencing the U-Net++ loss function, a cross-correlation function reflecting lateral correlation is incorporated into the U-Net network. This helps the U-Net network encoder extract features for seismic phase classification. In signal processing, cross-correlation represents the similarity between two signals. The cross-correlation of two signals f(t) and g(t) is defined as:
[0047]
[0048] Among them, C fg (τ) is the cross-correlation between f(t) and g(t), and τ is the time shift. The above equation, after Fourier transform, can be expressed as:
[0049]
[0050] Where F(ω) is the Fourier transform of f(t), G(ω) is the Fourier transform of g(t), ω is the frequency, and * is convolution. Similar to convolution, the cross-correlation function is a sliding product between two sequences, but unlike convolution, the cross-correlated product sequences are directly multiplied by sliding and then summed without being flipped.
[0051] like Figure 8 As shown, the cross-correlation U-Net convolutional network model incorporates the cross-correlation function into the downsampling layer of the U-Net convolutional network model. By considering the correlation of seismic facies information between adjacent seismic data, the learning requirements of the upsampling layer are reduced, which helps the U-Net network to better extract seismic facies features from the seismic data.
[0052] Residual networks have a higher receptive capacity than the input image, but convolutional neural networks have a much lower receptive capacity, making it difficult for some network structures to fully integrate the previous global image. However, for complex scene images, this strategy cannot extract the necessary information. In this regard, separating global background information from secondary regional background information helps distinguish various categories. By fusing information from different sub-regions through these receptive fields, a more powerful feature representation can be obtained. Therefore, it is necessary to expand the receptive field to improve the accuracy and speed of feature extraction. In this embodiment, the above functions are achieved through a pyramid pooling model. The pyramid pooling model can extract contextual information between features and uses a hierarchical global prior method, incorporating information between different scales and feature maps, to pre-construct global feature outputs on the last layer of the deep neural network structure. Figure 9 As shown, the Pyramid Pooling Module (PPM) includes several pyramid pooling modules of different scales. The output of the cross-correlation U-Net convolutional network model is preprocessed through a deep neural network, and features are extracted using a network expansion strategy. Global priors are used to classify features at different levels. Low-dimensional features are upsampled using 1×1 convolutional layers, and the size of the features obtained through bilinear interpolation is compared with the initial features. Figure 1 This ensures that the weights of the global features are preserved. Finally, all pyramid pooling feature outputs are concatenated into a final complete output through a fully connected layer. In this embodiment, the pyramid pooling module includes four pyramid pooling modules of different scales, with sizes of 1×1, 2×2, 3×3, and 6×6, respectively. Each of these pyramid pooling modules is concatenated with an upsampled convolutional layer as the output layer of the pyramid pooling model. However, this is not a limitation; the number of different pyramids and the size of each level in the pyramid pooling model can be adjusted according to the specific input data.
[0053] Based on the well logging lithofacies sequence division results and the known seismic stratigraphic information of the F3 block, the seismic lithofacies are divided as follows: Figure 10 As shown in the figure, the well profiles of wells F03-2 and F06-1 and the corresponding well logging facies distribution are shown, confirming that the distribution of gas-bearing sandstone facies and sandstone facies within this stratum is accurate. The results indicate that the well logging facies classification results are consistent with the actual geological conditions.
[0054] Example 2
[0055] Based on the same inventive concept, this embodiment discloses a seismic phase identification system based on a cross-correlation convolutional network, comprising:
[0056] The data acquisition module is used to acquire well logging data and seismic data;
[0057] The well logging lithofacies identification module is used to input well logging data into the committee mechanism model and perform well logging lithofacies identification on the well logging data;
[0058] The morphological processing module is used to perform morphological processing on seismic data;
[0059] The dataset expansion module is used to input well logging lithofacies identification results and morphologically processed seismic data into the generative adversarial network, and expand the sample dataset through the generative adversarial network.
[0060] The identification result generation module is used to input the original seismic data and the expanded sample dataset into the PSP-Unet neural network model to generate seismic facies identification results.
[0061] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0062] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific embodiments of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention. The above content is only a specific embodiment of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.
Claims
1. A seismic phase identification method based on cross-correlation convolutional networks, characterized in that, Includes the following steps: Obtain well logging data and seismic data; The well logging data is input into the committee mechanism model to perform well logging lithofacies identification. The seismic data was then subjected to morphological processing. Well logging lithofacies identification results and morphologically processed seismic data are input into a generative adversarial network (GAN) to expand the sample dataset. The raw seismic data and the expanded sample dataset are input into the PSP-Unet neural network model to generate seismic facies identification results. The generative adversarial network is a conditional generative adversarial network, which learns a structured loss function to penalize the construction of output nodes. The conditional generative adversarial network includes: a first generator, a second generator, a first discriminator, and a second discriminator. The first vector is generated by concatenating seismic information with random noise and input into the first generator to generate the seismic phase for judgment. The known seismic phase is concatenated with random noise, and the resulting second vector is input together with the identified seismic phase into the first discriminator for discrimination. The second vector is then input into the second generator to generate seismic data, and the generated seismic data is input together with the first vector into the second discriminator for discrimination. The PSP-Unet neural network model includes a cross-correlation U-Net convolutional network model and a pyramid pooling model connected in sequence. The cross-correlation U-Net convolutional network model is a downsampling layer of the U-Net convolutional network model with the cross-correlation function added.
2. The seismic phase identification method based on cross-correlation convolutional networks as described in claim 1, characterized in that, The committee mechanism model consists of three expert models: K-nearest neighbor, BP neural network, and support vector machine. The optimal solution for the weights of the three expert models is randomly searched using the Naive Bayes algorithm. Based on the optimal solution, the outputs of the three expert models are weighted and summed to obtain the output of the committee mechanism model.
3. The seismic phase identification method based on cross-correlation convolutional networks as described in claim 1, characterized in that, The method for performing morphological processing on the seismic data is as follows: Extract texture features from the morphological processing model, use structuring elements to convolve the seismic profile, smooth the boundaries of larger features in the morphological processing model, and separate features at smaller features. The seismic data is input into the morphological processing model, and an opening operation is performed on the seismic data through the mutual coupling operation of erosion and dilation to obtain morphologically processed seismic data.
4. The seismic phase identification method based on cross-correlation convolutional networks as described in claim 1, characterized in that, The generative adversarial network includes a generative model and a discriminative model. Random noise is input into the generator to generate a noise signal. The noise signal and seismic data are input into the discriminative model, which is used to determine the authenticity of the seismic signal. Both the generative model and the discriminative model are composed of several convolutional networks with fully connected layers removed.
5. The seismic phase identification method based on cross-correlation convolutional networks as described in claim 4, characterized in that, The output layer of the generative model uses the Tanh function as the activation function, while the remaining layers in the generative adversarial network all use the ReLU function as the activation function.
6. The seismic phase identification method based on cross-correlation convolutional networks as described in claim 1, characterized in that, Seismic data is input into the cross-correlation U-Net convolutional network model, and the output of the cross-correlation U-Net convolutional network model is input into the pyramid pooling model to generate seismic phase prediction results. The output layer of the cross-correlation U-Net convolutional network model and the output layer of the pyramid pooling model are connected across layers.
7. The seismic phase identification method based on cross-correlation convolutional networks as described in claim 6, characterized in that, The pyramid pooling model includes several pyramid pooling modules of different scales. The output of the cross-correlation U-Net convolutional network model is preprocessed by a deep neural network, and features are extracted by a network expansion strategy. Features at different levels are classified by global prior. Low-dimensional features are upsampled using a 1×1 convolutional layer, and the size of the features obtained by bilinear interpolation is consistent with the initial feature map. Finally, all pyramid pooling feature outputs are concatenated into the final complete output through a fully connected layer.
8. A seismic phase identification system based on a cross-correlation convolutional network, characterized in that, include: The data acquisition module is used to acquire well logging data and seismic data; The well logging lithofacies identification module is used to input the well logging data into the committee mechanism model and perform well logging lithofacies identification on the well logging data; The morphological processing module is used to perform morphological processing on the seismic data; The dataset expansion module is used to input well logging lithofacies identification results and morphologically processed seismic data into a generative adversarial network, and expand the sample dataset through the generative adversarial network. The identification result generation module is used to input the raw seismic data and the expanded sample dataset into the PSP-Unet neural network model to generate seismic facies identification results. The generative adversarial network is a conditional generative adversarial network, which learns a structured loss function to penalize the construction of output nodes. The conditional generative adversarial network includes: a first generator, a second generator, a first discriminator, and a second discriminator. The first vector is generated by concatenating seismic information with random noise and input into the first generator to generate the seismic phase for judgment. The known seismic phase is concatenated with random noise, and the resulting second vector is input together with the identified seismic phase into the first discriminator for discrimination. The second vector is then input into the second generator to generate seismic data, and the generated seismic data is input together with the first vector into the second discriminator for discrimination. The PSP-Unet neural network model includes a cross-correlation U-Net convolutional network model and a pyramid pooling model connected in sequence. The cross-correlation U-Net convolutional network model is a downsampling layer of the U-Net convolutional network model with the cross-correlation function added.