A multi-scale seismic facies intelligent identification method based on combination of stacked generalized ensemble and convolutional neural network

Abstract

The application discloses a multi-scale seismic facies intelligent identification method based on a combination of a stacked generalization ensemble and a convolutional neural network, and comprises the following steps: S1, obtaining three-dimensional post-stack seismic data, and preprocessing the data to construct learning samples for intelligent modeling; S2, using k-fold cross-validation to train a convolutional neural network model with different convolution kernels to construct a first layer of the stacked generalization ensemble; S3, integrating the convolutional neural network containing multi-scale information in a learning manner to obtain the output of the first layer of the stacked generalization ensemble; S4, constructing new data samples from the first layer output and corresponding original seismic facies labels for training of a meta-model in the integration, and obtaining a final stacked generalization ensemble model; and S5, inputting test data to obtain a seismic facies identification result containing multi-scale information. The method has better pattern difference recognition capability, thereby helping to improve the accuracy of seismic facies identification.

Description

Technical Field

[0001] This invention belongs to the field of seismic reservoir prediction and relates to a multi-scale intelligent identification method for seismic facies based on stacked generalization ensemble and convolutional neural network. In particular, it relates to using stacked generalization ensemble to construct a multi-scale convolutional neural network to make up for the shortcomings of traditional convolutional neural networks in pattern recognition ability due to fixed convolution kernels, thereby realizing intelligent identification and classification of seismic facies at multiple scales. Background Technology

[0002] Seismic data is a crucial source of reservoir geological information. As a direct representation of sedimentary facies on seismic profiles, effective identification of seismic facies provides vital guidance for subsequent reservoir development. Traditional seismic facies identification is achieved manually or semi-automatically by professional interpreters based on personal and regional experience and subjective visual perception. This process relies heavily on the interpreter's knowledge and experience, resulting in a large workload, low efficiency, and significant subjectivity. In recent years, deep learning technology has made significant progress in fields such as computer vision. Based on the similarity between seismic data interpretation and computer vision, many researchers have used convolutional neural networks capable of automatically extracting useful features to classify and predict seismic facies, greatly improving the efficiency of seismic facies identification.

[0003] However, in existing intelligent seismic phase identification methods, the convolutional kernels of convolutional neural networks are fixed, which to some extent limits their ability to capture pattern differences at multiple scales. Therefore, based on the above research, this invention proposes a novel method for constructing multi-scale neural networks. This method uses multiple conventional convolutional neural networks with different kernels as basis functions, and integrates these basis functions through stacking generalization to obtain the perceptual information of the convolutional neural network at different scales. This integrated information is then input into the lower-level meta-model to achieve the final intelligent seismic phase identification. Summary of the Invention

[0004] This invention provides a multi-scale intelligent seismic phase identification method based on a combination of stacked generalization ensemble and convolutional neural network, comprising the following steps: S1, Obtain three-dimensional post-stack seismic data and preprocess the data to construct learning samples for intelligent modeling; S2, utilizing k Cross-validation training yields convolutional neural network models with different convolutional kernels, which are then used to construct the first layer of a stacked generalization ensemble. S3, Select a simple machine learning algorithm to integrate the convolutional neural network containing multi-scale information in a learning manner to obtain the output of the first layer in the stacked generalization ensemble; S4. New data samples are constructed using the first layer output and the corresponding original seismic phase labels, which are used for training the ensemble model to obtain the final stacked generalized ensemble model. S5, input test data to obtain seismic facies identification results containing multi-scale information.

[0005] Further, obtaining three-dimensional post-stack seismic data in step S1 and preprocessing the data to construct learning samples for intelligent modeling specifically includes the following steps: S11, Preprocess the acquired raw seismic data to obtain three-dimensional post-stack seismic data; after denoising the data, normalize its amplitude to [-1,1]; S12, perform non-repeating random sampling of the index of the minority class to obtain multiple profiles along the main survey line and the connecting survey line; S13, the cross-section is cut out in an overlapping manner according to a certain step size and a certain window size, and the index of the minority class is randomly sampled without repetition, so as to achieve random cutting of the cross-section with the same window size. S14, repeat the above process for the seismic phase label data corresponding to the profile, and combine the label slices with the slices obtained from the profile to obtain the final learning sample; S15, group the learning samples according to a certain ratio to obtain the training dataset. and test dataset .

[0006] Furthermore, in step S2, the following steps are utilized: k Cross-validation training yields convolutional neural network models with varying kernel sizes. The specific steps for constructing the first layer of a stacked generalization ensemble include: S21, adopts k Layered cross-validation is used to perform training data Grouping yields groups with similar proportions and sizes. k Subset ; S22, sequentially k In the data of fold k -10% off data Provided for learning by convolutional neural networks, where Indicates the number of folds; S23: For convolutional neural networks with different kernel sizes, repeat step S22 to obtain multiple convolutional neural network models with the same framework structure but different kernel sizes. ,in Indicates the first Each convolutional neural network (reflected in the different convolutional kernels) Indicates the first The convolutional neural network is based on the first one except the second one. Hypotheses or mappings learned from data other than folded data, i.e., convolutional neural network models.

[0007] Further, in step S3, a simple machine learning algorithm is selected to integrate the convolutional neural network containing multi-scale information in a learning manner to obtain the output of the first layer in the stacked generalization ensemble. This specifically includes the following steps: S31, the trained convolutional neural network model is used for The remaining data in the fold cross-validation that did not participate in model training ( Based on the prediction, we obtain:

[0008] in Indicates the first The model for the first Prediction of folded data; S32, merge the above datasets folded together, and average the outputs of each convolutional neural network model to obtain the output of the first layer of the stacked generalization ensemble, expressed as:

[0009] in, Indicates the number of models. Indicates the number of folds.

[0010] Further, step S4 involves constructing new data samples using the original seismic phase labels corresponding to the first-layer output, which are then used for training the ensemble middle-component model to obtain the final stacked generalized ensemble model. This process specifically includes the following steps: S41, using the original seismic phase label to... The data is labeled to obtain the training data for the second layer model of the stacked generalization ensemble. ; S42, the trained convolutional neural network models are used to predict the test data, and the prediction results are averaged to obtain...

[0011] in, Indicates the first A convolutional neural network model, This represents the original test data; S43, using the original seismic phase labels The data was labeled to obtain the test data for the second layer model in the stacked generalization ensemble. ; S44, Select the learning algorithm for building the second-layer model in the ensemble, and provide new training data for it to learn, to obtain the meta-model in the stacked generalization ensemble.

[0012] Furthermore, step S5, which involves inputting test data to obtain seismic facies identification results containing multi-scale information, specifically includes the following steps: S51 performs the same preprocessing on the test data as on the training data; S52, the preprocessed test data is processed in the same way as the training data; S53, the truncated test data is input into the stacked generalization ensemble model. This data stream flows through the meta-model via convolutional neural network models with different convolutional kernels, and the final output is the predicted seismic facies identification result.

[0013] Compared with the prior art, the method proposed in this invention has the following advantages: A stacked generalization ensemble strategy is employed to combine convolutional neural network models with varying kernel sizes. This allows the pattern differences of each seismic facies at different scales to be collectively fed into the second-layer learner within the ensemble, represented by the outputs of each convolutional neural network model, for further learning. This combination of scale information features and relearning enhances the proposed method's ability to identify pattern differences, thereby improving the accuracy of seismic facies identification. Attached Figure Description

[0014] Figure 1 This is a flowchart of a multi-scale intelligent seismic phase identification method based on stacked generalization ensemble and convolutional neural network according to the present invention; Figure 2 This is a schematic diagram of seismic data and seismic phase data extraction in an embodiment of the present invention; Figure 3 This is an embodiment of the present invention. Figure 2 The learning sample images obtained by cropping as shown are (a) cropping of seismic data and (b) cropping of seismic phases. Figure 4 This is a diagram of the convolutional neural network model established in this embodiment of the invention; Figure 5 This is a diagram of a stacked generalization ensemble model established based on multiple convolutional kernel scales in an embodiment of the present invention; Figure 6 This is a comparison diagram of the seismic facies identification results of various models on a two-dimensional profile along the main lateral line in this embodiment of the invention. Figure 6 (a) Figure 6 (b) Figure 6 (c) and Figure 6 (d) shows the prediction results of convolutional neural network models with kernel sizes of 3, 5, 7, and 9, respectively. Figure 6 Figure 6 (e) represents the prediction results of the model established in this invention. Figure 6 (f) represents the actual seismic phase; Figure 7 This is a comparison diagram of seismic facies identification results of various models on a two-dimensional profile in the direction of the connecting lateral line in the embodiments of the present invention. Figure 7 (a) Figure 7 (b) Figure 7 (c) and Figure 7 (d) shows the prediction results of convolutional neural network models with kernel sizes of 3, 5, 7, and 9, respectively. Figure 7 (e) represents the prediction results of the model established in this invention. Figure 7 (f) represents the actual seismic phase. Detailed Implementation

[0015] The following is in conjunction with the appendix Figure 1-7 The present invention provides a more detailed description of a multi-scale intelligent seismic phase identification method based on stacked generalization ensemble and convolutional neural network, so that those skilled in the art can better understand the present invention.

[0016] like Figure 1 As shown, this invention provides a multi-scale intelligent seismic phase identification method based on a combination of stacked generalization ensemble and convolutional neural network. The method includes the following steps: S1, Obtain three-dimensional post-stack seismic data and preprocess the data to construct learning samples for intelligent modeling; S2, utilizing k Cross-validation training yields convolutional neural networks with different kernels, which are then used to construct the first layer of a stacked generalization ensemble. S3, Select a simple machine learning algorithm to integrate the convolutional neural network containing multi-scale information in a learning manner to obtain the output of the first layer in the stacked generalization ensemble; S4. New data samples are constructed using the first layer output and the corresponding original seismic phase labels, which are used for training the ensemble model to obtain the final stacked generalized ensemble model. S5, input test data to obtain seismic facies identification results containing multi-scale information.

[0017] In specific implementation, obtaining three-dimensional post-stack seismic data and preprocessing the data to construct learning samples for intelligent modeling specifically includes the following steps: S11, Preprocess the acquired raw seismic data to obtain three-dimensional post-stack seismic data; after denoising the data, normalize its amplitude to [-1,1]; S12, perform non-repeating random sampling of the index of the minority class to obtain multiple profiles along the main survey line and the connecting survey line; S13, as Figure 2As shown in (a), the cross-section is randomly cropped in a regular manner with a step size of 10 and a window size of (31, 51), while the index of the minority class is randomly sampled without repetition, so as to achieve random cropping of the cross-section with the same window size. S14, as Figure 2 As shown in (b), the above process is repeated for the seismic phase label data corresponding to the profile, and the label slices are combined with the slices obtained from the profile to obtain the final result. Figure 3 Learning samples; S15, group the learning samples according to a certain ratio to obtain the training dataset. and test dataset .

[0018] In specific implementation, the aforementioned utilization k The process of training convolutional neural networks with varying kernel sizes using cross-validation to construct the first layer of a stacked generalization ensemble specifically includes the following steps: S21, uses 5-fold hierarchical cross-validation on the training data. Grouping yields five subsets of similar size representing each type of instance. ; S22, sequentially divide the data into 4-fold data from the 5-fold data. Used for learning convolutional neural networks, where Indicates the number of folds; S23: For convolutional neural networks with kernel sizes of 3, 5, 7, and 9, repeat step S22 to obtain multiple convolutional neural network models with the same framework structure but different kernel sizes. ,in Indicates the first Each convolutional neural network (reflected in the different convolutional kernels) Indicates the first The convolutional neural network is based on the first one except the second one. Hypotheses or mappings learned from data other than folded data are also called convolutional neural network models; like Figure 4 As shown, the convolutional neural network constructed in this embodiment contains three convolutional layers, namely Conv1, Conv2, and Conv3, and two fully connected layers, namely FC1 and FC2. A max-pooling layer is added after Conv1, and the activation function used in the network is ReLU. To prevent overfitting, batch regularization and Dropout regularization are used in each layer. Furthermore, one-third of the training samples are used as validation data to implement an early stopping mechanism during model training.

[0019] In specific implementation, the method of integrating the convolutional neural network containing multi-scale information through learning to obtain the output of the first layer in the stacked generalization ensemble specifically includes the following steps: S31, the trained convolutional neural network is used for the remaining fold of data in the 5-fold cross-validation that did not participate in model training. Based on the prediction, we obtain: (1) in Indicates the first The model for the first Prediction of folded data; S32, merge the above datasets folded together, and average the outputs of each convolutional neural network model to obtain the output of the first layer of the stacked generalization ensemble, expressed as: (2) in, Indicates the number of models. Indicates the number of folds; In specific implementation, the step of constructing new training samples based on the original seismic phase labels corresponding to the first-layer output, for training the integrated middle-order model, and obtaining the final stacked generalized ensemble model specifically includes the following steps: S41, using the original seismic phase label to... The data is labeled to obtain the training data for the second layer model of the stacked generalization ensemble. ; S42, the trained convolutional neural network models are used to predict the test data, and the prediction results are averaged to obtain... (3) in, Indicates the first A convolutional neural network model, This represents the original test data; S43, using the original seismic phase labels The data was labeled to obtain the test data for the second layer model in the stacked generalization ensemble. ; S44, Select the learning algorithm for building the second-layer model in the ensemble, and provide new training data to it for learning, to obtain the meta-model in the stacked generalization ensemble.

[0020] When building the meta-model of stacked generalization ensemble, the logistic regression algorithm was selected as the meta-learner, and the training data... It serves as the learning sample. The two-layer structure model constructed from the multi-scale convolutional neural network model and the meta-model is as follows. Figure 5 The stacked generalization ensemble model described above.

[0021] In specific implementation, the input of test data to obtain seismic facies identification results containing multi-scale information includes the following steps: S51 performs the same preprocessing on the test data as on the training data; S52, the preprocessed test data is processed in the same way as the training data; S53, the truncated test data is input into the stacked generalization ensemble model. This data stream flows through the meta-model via convolutional neural network models with different convolutional kernels, and the final output is the predicted seismic facies identification result.

[0022] like Figure 6 The diagram shows a comparison between the seismic facies identification results of the various convolutional neural network models and the stacked generalization ensemble model constructed in this invention on a two-dimensional profile along the main seismic line in the test data and the actual seismic facies. Comparing the positions within the red, yellow, and orange boxes reveals that the model constructed in this invention integrates the advantages of convolutional neural network models at various scales while avoiding their corresponding disadvantages, exhibiting the best identification ability both within and at the boundaries of seismic facies.

[0023] like Figure 7 The diagram shows a comparison between the seismic facies identification results of the various convolutional neural network models and the stacked generalization ensemble model established in this invention for a certain two-dimensional profile of the connecting survey line in the test data and the actual seismic facies. The red arrows indicate the locations of incorrectly identified seismic facies. The comparison results further illustrate that integrating convolutional neural networks with different convolutional kernels through stacked generalization helps improve the accuracy of seismic facies identification.

Claims

1. A multi-scale intelligent seismic facies identification method based on stacked generalization ensemble and convolutional neural network, characterized in that, Includes the following steps: S1, Obtain three-dimensional post-stack seismic data and preprocess the data to construct learning samples for intelligent modeling; S2, utilizing k Cross-validation training yields convolutional neural network models with different convolutional kernels, which are then used to construct the first layer of a stacked generalization ensemble. S3, the convolutional neural network containing multi-scale information is integrated in a learning manner to obtain the output of the first layer in the stacked generalization integration; S4. New data samples are constructed using the first layer output and the corresponding original seismic phase labels, which are used for training the ensemble model to obtain the final stacked generalized ensemble model. S5, input test data to obtain seismic facies identification results containing multi-scale information.

2. The multi-scale seismic facies intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 1, characterized in that, Step S1 involves obtaining three-dimensional post-stack seismic data and preprocessing the data to construct learning samples for intelligent modeling. This specifically includes the following steps: S11, Preprocess the acquired raw seismic data to obtain three-dimensional post-stack seismic data; after denoising the data, normalize its amplitude to [-1,1]; S12, perform non-repeating random sampling of the index of the minority class to obtain multiple profiles along the main survey line and the connecting survey line; S13, the cross-section is cut out in an overlapping manner according to a certain step size and a certain window size, and the index of the minority class is randomly sampled without repetition, so as to achieve random cutting of the cross-section with the same window size. S14, repeat the above process for the seismic phase label data corresponding to the profile, and combine the label slices with the slices obtained from the profile to obtain the final learning sample; S15, group the learning samples according to a certain ratio to obtain the training dataset. and test dataset .

3. The multi-scale seismic facies intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 2, characterized in that, The cross-section described in step S13 is randomly cut in a regular manner with a step size of 10 and a window size of (31, 51).

4. The multi-scale seismic phase intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 1, characterized in that, In step S2, utilize k Cross-validation training yields convolutional neural network models with varying kernel sizes. The specific steps for constructing the first layer of a stacked generalization ensemble include: S21, adopts k Layered cross-validation is used to perform training data Grouping yields groups with similar proportions and sizes. k Subset ; S22, sequentially k In the data of fold k -10% off data Provided for learning by convolutional neural networks, where Indicates the number of folds; S23: For convolutional neural networks with different kernel sizes, repeat step S22 to obtain multiple convolutional neural network models with the same framework structure but different kernel sizes. ,in Indicates the first A convolutional neural network, Indicates the first The convolutional neural network is based on the first one except the second one. Hypotheses or mappings learned from data other than folded data, i.e., convolutional neural network models.

5. The multi-scale seismic facies intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 4, characterized in that, The convolutional kernel sizes of the convolutional neural networks described in step S23 are 3, 5, 7, and 9, respectively.

6. The multi-scale seismic facies intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 1, characterized in that, Step S3, which integrates the convolutional neural network containing multi-scale information in a learning manner to obtain the output of the first layer in the stacked generalization ensemble, specifically includes the following steps: S31, the trained convolutional neural network model is used for The remaining data in the fold cross-validation that did not participate in model training ( Based on the prediction, we obtain: in Indicates the first The model for the first Prediction of folded data; S32, merge the above datasets folded together, and average the outputs of each convolutional neural network model to obtain the output of the first layer of the stacked generalization ensemble, expressed as: in, Indicates the number of models. Indicates the number of folds.

7. The multi-scale seismic facies intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 6, characterized in that, The steps described in step S31 It is 5.

8. The multi-scale seismic facies intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 1, characterized in that, Step S4 involves constructing new data samples using the original seismic phase labels corresponding to the first-layer output, which are then used for training the ensemble middle-component model to obtain the final stacked generalized ensemble model. This process specifically includes the following steps: S41, using the original seismic phase label to... The data is labeled to obtain the training data for the second layer model of the stacked generalization ensemble. ; S42, the trained convolutional neural network models are used to predict the test data, and the prediction results are averaged to obtain... in, Indicates the first A convolutional neural network model, This represents the original test data; S43, using the original seismic phase labels The data was labeled to obtain the test data for the second layer model in the stacked generalization ensemble. ; S44, Select the learning algorithm for building the second-layer model in the ensemble, and provide new training data for it to learn, to obtain the meta-model in the stacked generalization ensemble.

9. The multi-scale seismic phase intelligent identification method based on stacked generalization ensemble and convolutional neural network according to claim 1, characterized in that, Step S5 involves inputting test data to obtain seismic facies identification results containing multi-scale information, specifically including the following steps: S51 performs the same preprocessing on the test data as on the training data; S52, the preprocessed test data is processed in the same way as the training data; S53, the truncated test data is input into the stacked generalization ensemble model. This data stream flows through convolutional neural network models with different convolutional kernels and then through the meta-model. The final output is the predicted seismic facies identification result.

Images