Geological Feature Constrained ACGAN Seismic Facies Identification Method

Abstract

The present invention discloses a method for identifying seismic facies with geological feature constraints for ACGAN, which is applied to the field of seismic data processing. Aiming at the problems existing in the prior art that overfitting phenomenon will occur when the number of samples is insufficient or the sample category distribution is unbalanced, and it is still very difficult and expensive to obtain seismic facies label data; the present invention first uses the convolutional layer of an unsupervised convolutional neural network to extract the deep features of unlabeled actual seismic data, replaces the noise input of the original ACGAN with it, and adds label information to the features as the input of the generator, so as to realize the constraint on the process of generating seismic data by ACGAN, which is more helpful for generating data that meets the geological sedimentation law; the method of the present invention can enhance the dataset with insufficient original sample quantity and unbalanced sample category distribution, thus greatly improving the identification effect of seismic facies.

Description

Technical Field

[0001] The present invention belongs to the field of seismic data processing, and particularly relates to a seismic facies identification technique. Background Art

[0002] A seismic facies is a seismic reflection unit in a three-dimensional space defined by specific seismic reflection parameters. It is the seismic response of a specific sedimentary facies or geological body. The seismic reflection characteristics of this unit in the three-dimensional space are different from those of its adjacent units, representing the lithological combination, bedding, and sedimentary characteristics of the sediments that generate its reflection, providing a reference for analyzing geological conditions and predicting oil and gas reservoirs. The identification of seismic facies is of great significance for oil and gas exploration.

[0003] Initially, seismic facies were manually divided by the "phase plane method", which was carried out by observing and describing the reflection characteristics (amplitude, frequency, and continuity) on seismic profiles. In the face of complex seismic profile situations, it is often affected by subjectivity, so interpreters are required to have high professional qualities. With the progress of seismic exploration technology and computers, there are more and more means for identifying and describing seismic facies, and the application scope is getting wider and wider. The identification methods of seismic facies can be divided into waveform classification methods, seismic attribute feature mapping methods, and phase division methods based on seismic geomorphology. In addition, unsupervised methods such as self-organizing feature map neural networks (SOM), K-means clustering analysis methods, and sparse autoencoders (SAE) have been widely used in seismic facies identification. These methods can extract geological feature information from massive seismic data. However, unsupervised methods are difficult to determine the number of clusters, and their results still need to be interpreted or calibrated by seismic experts.

[0004] With the breakthrough progress of deep learning technology, deep learning algorithms have been widely applied in seismic facies classification, greatly improving the work efficiency of interpreters. The neural network method for seismic facies classification has strong adaptive ability, fault tolerance and large data volume calculation ability. Methods such as competitive neural network, convolutional autoencoder, convolutional neural network, and semi-supervised generative adversarial network have been successively applied to seismic facies identification. Through deep learning, the features of seismic facies can be better extracted, achieving good recognition results. However, deep learning methods rely on a large number of labeled sample data. When the sample quantity is insufficient or the sample category distribution is unbalanced, overfitting will occur, and it is still very difficult and expensive to obtain seismic facies label data. Currently, the data sources mainly include two parts. One part comes from the dataset obtained by drilling and formation parameter measurement. However, affected by exploration and development costs, these data far from meet the requirements of deep learning algorithms for complex geological structures. Synthetic geological models are also a method for obtaining data, but the generated models cannot fully simulate all underground situations. Therefore, how to solve the situation of insufficient sample data in actual data and enhance the dataset is an urgent problem to be solved.

[0005] In the traditional image segmentation process, in order to enable the network to learn more features, data augmentation methods are often used to increase the sample quantity, including rotation, translation, scaling, random occlusion, horizontal flipping, noise perturbation, etc. Due to the unique directions, textures and other features of specific geological structures, formation sediments and lithofacies in seismic data, rotating, translating, flipping, etc. of seismic data will change the geological meaning depicted by seismic facies, and traditional methods are not suitable for enhancing seismic data samples and labels.

[0006] Generative Adversarial Networks (GAN) can generate data consistent with the original data distribution from random noise vectors and have excellent image simulation ability. Since its proposal, GAN has often been used to expand the dataset. Auxiliary Classifier GAN (ACGAN) has improved the original GAN and can generate data of specific categories, thereby enhancing the sample data.

[0007] The success of applying machine learning algorithms depends on the completeness of the labeled dataset. A large and feature-complete labeled dataset can ensure the effectiveness of machine learning algorithms, prevent overfitting, and improve the generalization ability of the algorithms. At the same time, if the sample distribution of the data is unbalanced, it will lead to too few features in the category with a small sample size, and it is difficult to mine and identify patterns from it. Even if a classification model is obtained, it is prone to the problem of overfitting due to over-reliance on limited data samples. When the model is applied to new data, the accuracy and robustness of the model will be very poor. Affected by exploration and development costs, the seismic response datasets corresponding to different geological structure models calibrated by drilling and formation parameter measurements are far from meeting the needs of training machine learning algorithms for exploring and developing oil and gas reservoirs with complex geological structures. And due to geological structure characteristics, the distribution of seismic facies is unbalanced in the same work area. Considering sedimentological geological laws, existing data augmentation methods in the field of image processing such as rotation and flipping are not suitable for seismic dataset augmentation. Summary of the Invention

[0008] The present invention improves the ACGAN by considering geological features as constraints, generates seismic data to expand the original sample dataset, and improves the recognition ability of CNN for seismic facies.

[0009] The technical solution adopted by the present invention is: a geological feature-constrained ACGAN seismic facies recognition method, which is applied to the situation where the number of fault data in actual seismic data is much less than the number of non-fault data; including:

[0010] S1. Normalize the seismic data of the seismic facies to be recognized to obtain a training set and a test set;

[0011] S2. Expand the training set using the improved ACGAN;

[0012] S3. Train the CNN network according to the expanded training set;

[0013] S4. Input the test set into the trained CNN network to obtain the seismic facies recognition result.

[0014] The beneficial effects of the present invention: The present invention first improves the ACGAN network based on geological features as constraints according to the characteristics of seismic data and the requirements of seismic facies recognition, and then uses the ACGAN to generate seismic data of specific categories and corresponding labeled datasets, so as to enhance the original unbalanced data and insufficient sample data. Finally, the effect of seismic facies recognition after data augmentation is tested on a Convolutional Neural Network (CNN); the method of the present invention has the following advantages:

[0015] 1. For data with unbalanced sample classes, specific classes of data can be generated to enrich the features of the original data and avoid overfitting of the network during seismic facies recognition;

[0016] 2. For data with a limited number of samples, the geological feature-constrained ACGAN proposed by the present invention is used to generate the dataset and the label set simultaneously, greatly improving the accuracy of seismic facies recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the existing GAN network structure;

[0018] Figure 2 It is a schematic diagram of the existing ACGAN network structure;

[0019] Figure 3 It is a structural diagram of the geological feature-constrained ACGAN seismic facies recognition network of the present invention;

[0020] Figure 4 It is a flow chart of the geological feature-constrained ACGAN data augmentation CNN seismic facies recognition method of the present invention;

[0021] Figure 5 It is a process of normalizing seismic data;

[0022] Among them, (a) is the original data, and (b) is the normalized data;

[0023] Figure 6 It is a schematic diagram of using a convolutional layer to extract features from a large amount of unlabeled seismic data;

[0024] Figure 7 It is the generated fault model;

[0025] Among them, (a) is the used fault model, and (b) is the model profile along the crosslines direction;

[0026] Figure 8 It is a schematic diagram of cutting the fault model into 64×16 data along the crossline direction;

[0027] Among them, (a) is the original non-fault data, and (b) is the original fault data;

[0028] Figure 9 It is the data generated by ACGAN;

[0029] Among them, (a) is the generated non-fault data, and (b) is the generated fault data;

[0030] Figure 10 It is the accuracy of the test set when different amounts of fault data are added;

[0031] Figure 11 The accuracy rate and loss function curves of the network when the fault data are 20:1 and 1:1;

[0032] Among them, (a) is the accuracy rate curve of the network, and (b) is the loss function curve of the network;

[0033] Figure 12 The confusion matrix for the classification of the test set;

[0034] Among them, (a) is before the fault data augmentation, and (b) is after the fault data augmentation;

[0035] Figure 13 The 10 types of seismic facies data of F3;

[0036] Figure 14 The ten types of seismic facies data generated by ACGAN;

[0037] Figure 15 The accuracy rate of the test set when different amounts of generated data are added;

[0038] Figure 16 The accuracy rate and loss function curves of CNN;

[0039] Among them, (a) is the accuracy rate curve, and (b) is the loss function curve;

[0040] Figure 17 The t-SNE visualization result;

[0041] Among them, (a) is the original test set, (b) is the test set features extracted by the network before data augmentation, and (c) is the test set features extracted by the network after data augmentation. Specific implementation manners

[0042] To facilitate those skilled in the art to understand the technical content of the present invention, the prior art related to the present invention will be described first:

[0043] 1. ACGAN network architecture

[0044] The structure of GAN is as Figure 1As shown, it consists of two parts: a discriminator and a generator. The purpose of the discriminator is to try to correctly determine whether the input data comes from the real data distribution or the generated distribution; while the purpose of the generator is to try to learn a distribution consistent with the real data distribution and generate samples that are indistinguishable from the real ones. The core idea is to use the adversarial training strategy between the discriminator and generator networks, forcing the generated distribution of the network to approximate the real data distribution infinitely. The training process of GAN is divided into two stages: the first stage trains the discriminator, and the second stage trains the generator. After training the discriminator, true and false information is passed to the generator, and the generator continuously optimizes the network (essentially updates the parameters) according to the authenticity of the information (essentially gradient information), trying to generate high-quality samples to "fool" the discriminator. Thus, such an adversarial strategy is produced: when training the discriminator, the discriminator determines that the generated samples are maximized; while when training the generator, the discriminator determines that the generated samples are minimized. Through such a training strategy, both the discriminator and the generator are continuously improving their discrimination ability and generation ability until the discriminator can no longer determine whether the samples generated by the generator come from the real data distribution or the generated distribution.

[0045] The structure of ACGAN is as Figure 2 shown. Based on the original GAN, auxiliary classification information is introduced, and pictures with category information can be generated. Specifically, a category vector is added to the input noise as the input of the generator, so that the generated fake data also has label information. A category discrimination layer is added to the discriminator to judge both the authenticity and category of the input pictures at the same time. The final loss function includes the judgment of authenticity L s and the judgment of category L c in two parts.

[0046] L S = E[logP(S = real|X real )] + E[logP(S = fake|X fake )] (1)

[0047] L C = E[logP(C = c|X real )] + E[logP(C = c|X fake )] (2)

[0048] The training purpose of the discriminator is to maximize L C + L S , that is, to judge the authenticity of the data and make a correct prediction of the category of the data at the same time; the training purpose of the generator is to maximize L C - L S , hoping that the generated pictures can deceive the discriminator and make it judge wrongly, and at the same time hoping that the generated pictures can be predicted as the correct category.

[0049] In the prior art, ACGAN uses noise as input and tries to learn the distribution of real data as much as possible to generate data consistent with the original data distribution. For seismic data generation, ACGAN with noise as input does not consider the actual geological features and may generate data that violates the laws of stratigraphic sedimentology. Seismic data contains rich information such as strata, geology, lithology, and structure, and there are certain differences in the statistical laws of these information in different work areas. A large amount of unlabeled seismic data can be obtained from actual seismic exploration work areas. If the features contained in these unlabeled seismic data can be used as prior knowledge to guide the data generation process of the ACGAN network, the effectiveness and rationality of the generated data will be greatly improved.

[0050] The present invention first performs a convolution operation on seismic data using a two-dimensional convolutional layer to obtain a feature image, replaces the original noise input with it, and adds randomly generated label information to the features as the input of the generator, thereby realizing the constraint on the process of generating seismic data by ACGAN, and thus constructing an ACGAN model suitable for seismic data features and seismic facies requirements. The overall network structure diagram of this article is as Figure 3 shown:

[0051] The method flow framework for seismic facies recognition with data augmentation in the present invention is as Figure 4 shown. The specific implementation steps are as follows:

[0052] A. Data processing

[0053] The input data required by the present invention is cut by sliding seismic data. The value ranges of seismic data at different positions vary greatly, and the network may not be able to learn the most essential change features of the data and cause misclassification. In order to eliminate the influence of data dimensions and facilitate network solution and optimization, it is necessary to normalize the original seismic data. The normalization method can maintain the features of the data and only transform the data into different numerical ranges, so it has no impact on the recognition task of the present invention.

[0054] Mean Normalization is a common data normalization method in machine learning, which can scale the numerical range to the interval [-1, 1]. The normalized data can be obtained by the following formula:

[0055]

[0056] Where: is x i the normalized data; u, x max and x min are the mean, maximum value, and minimum value of the original data respectively. The process of normalizing seismic data is as Figure 5 shown.

[0057] The normalized data obtained from step A is divided into two parts. One part is added with two types of labels according to the presence or absence of faults and used as the labeled real data to input into the discriminator. The other part is not added with labels, but directly uses the convolutional layer to extract features, and the features are used as the input of the generator.

[0058] B. Feature extraction

[0059] The present invention uses a convolutional layer to extract features from a large amount of unlabeled seismic data. The calculation formula of two-dimensional convolution is as follows:

[0060]

[0061] Among them, matrix A is the original data, that is, unlabeled actual seismic data, and its corresponding number of rows and columns are M r and M c ; matrix B is the convolution kernel, and its corresponding number of rows and columns are N r and N c . m and n are the coordinates along the x-axis and y-axis in the matrix to be convolved, and s - m and t - n are the coordinates along the x-axis and y-axis corresponding to the convolution kernel. C is the extracted feature, and s and t satisfy 0 ≤ s < M r +N r -1, 0 ≤ t < M c +N c -1. As Figure 6 shown, the convolution window slides on the original data, and the original data taken is multiplied point by point with the convolution kernel to obtain the corresponding output. Compared with noise, the extracted feature data retains the significant features of the original seismic data, can constrain the generated seismic data, and is more helpful for generating data that meets the geological sedimentation law.

[0062] C. Data augmentation

[0063] The features extracted in the previous step are fused with the label information and used as the input of the generator in ACGAN. Through the training of ACGAN, seismic data approaching the statistical distribution law of real data is generated. When the accuracy of the discriminator of ACGAN in judging the authenticity of data approaches 50%, it indicates that the generator has generated relatively real seismic data, which is sufficient to deceive the discriminator. Select the labels of the generated seismic data set after the accuracy reaches 50% and add them to the training set of CNN to enhance the original data set.

[0064] Source of label information: For the fault model data, the faults in the model are artificially generated, so their positions are known. The fault data is cut, and corresponding labels can be formulated according to whether there are faults in it; for F3 data, its data and labels are divided manually according to horizons, and this data has been open-sourced and can be obtained from the Internet.

[0065] The fusion in step C is specifically as follows: Through the concatenate function in the neural network, the features and labels are concatenated together to make their size consistent with that of the real seismic data, which can be used as the input of the discriminator.

[0066] D. Training the model

[0067] Use the 2D CNN model built by the present invention to train and test the seismic data before and after augmentation respectively to realize the identification of seismic facies. The 2D CNN model built by the present invention is Figure 3 the convolutional neural network in, specifically including: a convolutional layer, a normalization layer, a pooling layer, a dropout layer, and a Softmax classification layer.

[0068] The application effect of the method of the present invention will be described below in combination with model data and actual data.

[0069] 1) Model data

[0070] In actual seismic data, the number of fault data is much less than that of non-fault data. During the training of the neural network, it is considered that when the ratio of the number of fault data to the number of non-fault data is less than 1:10, it is in an unbalanced distribution state, that is, the number of fault data is much less than that of non-fault data. If CNN is directly used for seismic facies identification, it will cause the network to overfit to non-fault data and cannot accurately identify fault data. In this embodiment, on the generated fault model as Figure 7 shown, the process of enhancing fault data through the improved ACGAN is simulated under the condition of unbalanced sample distribution.

[0071] The seismic event is the connection line of the extreme values with the same vibration phase of each seismic trace on the seismic record, and it has a certain thickness along the time axis. In order to finely divide the fault while including the complete seismic event, in this embodiment, the fault model is cut into 64×16 data along the crossline direction, as Figure 8 shown. Select 10,000 non-fault data and 500 fault data as the training set of CNN. At this time, the ratio of fault data to non-fault data is 20:1. 250 fault data and 250 non-fault data are used as the test set. At the same time, 30,000 data are selected to extract features, and 500 fault data and 500 non-fault data are used as the training set of ACGAN to generate data. The generated data is shown in Figure 9 .

[0072] By adding different amounts of generated data, the ratio of fault to non-fault data in the training set can be changed. In this embodiment, in Figure 10 the classification accuracy of CNN for the test set under different training sets is plotted. At the same time, inFigure 11 The accuracy and loss function curves of the network are shown when the fault data is 20:1 and 1:1. From the comprehensive results, when the proportion of each type of data is close to 1:1, the network shows better convergence and accuracy. To further measure the recognition of the network for fault data, in this embodiment, Figure 12 the confusion matrix of the network classification results before and after data augmentation is drawn. The confusion matrix is an index to evaluate the accuracy of the classification model, which can separately count the number of misclassified and correctly classified categories of the classification model and display them in tabular form. The diagonal position of the table represents the classification accuracy of this type of seismic facies. It can be seen that with the addition of more generated data, the classification accuracy of the original fault data with a small quantity has been improved. It shows that the method of the present invention can enrich the original data, add more features, and make it achieve a better recognition effect.

[0073] 2) Actual data

[0074] In this embodiment, the method proposed by the present invention is applied to actual seismic data to verify its feasibility. The F3 work area is a block in the Dutch area of the North Sea, which is publicly available on https: / / terranubis.com / datainfo / F3-Demo-2020 by the Dutch government. Reinaldo et al. reinterpreted the F3 data to generate nine horizons, and thus divided it into ten types of seismic facies, which are respectively marked as numbers 1, 2, 3, ……, 10 in Figure 13 and 14 . After interpretation, the size of the seismic facies data is 64×25, as shown in Figure 13 and this data is published on https: / / doi.org / 10.5281 / zenodo.1422787.

[0075] This method uses the F3 seismic facies data to simulate the situation of insufficient sample quantity. In this embodiment, 200 seismic data of each type in the inline direction are selected as the training set of the network, and at the same time, 15,000 data are selected as unlabeled data for feature extraction. Data is generated through the improved ACGAN. When the discrimination accuracy of the discriminator for the authenticity of the data is close to 50%, it is considered that the generated data is close to the real data distribution and can be used to expand the data set. The generated pictures are as shown in Figure 14 .

[0076] By adding different amounts of generated data, the number of training sets of the CNN is expanded to {4000, 6000, 8000, 10000, 15000}, and the classification effect of the network is tested on the same test set with 50 data of each type of seismic facies. Figure 15The classification accuracy results of the test set are shown under different training sets. It can be observed that as more generated data is added, the accuracy of the test set continuously increases. When the number of generated data added is 8000, the accuracy of the test set tends to be stable.

[0077] The accuracy and loss function curves of the CNN are plotted for the original training set without added generated data and the training set with 8000 added generated data, as Figure 16 shown. After adding 8000 generated data, the convergence speed of the CNN is accelerated. In the original case, the network reaches the convergence state after about 200 iterations, while after adding the generated data, it can reach the convergence state after about 100 iterations. And the accuracy of the training set has increased from 85% to 94%. To further analyze the classification results of the trained CNN model on the test set, in this embodiment, t-Distributed Stochastic Neighbor Embedding (t-SNE) is used as a measurement criterion. t-SNE is an embedding model that can map data in a high-dimensional space to a low-dimensional space and preserve the local characteristics of the data set. After the t-SNE transformation, if the data is separable in the low-dimensional space, then the data is separable. The t-SNE visualization images of the test set features extracted by the CNN before and after data augmentation are shown in Figure 17 It can be seen that after adding the generated data, the visualization results show better separation ability. This shows that the seismic facies recognition effect of the network can be improved by using the data augmentation method of the present invention.

[0078] Those of ordinary skill in the art will realize that the embodiments described herein are to assist the reader in understanding the principles of the present invention and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, various changes and modifications can be made to the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. Geological feature-constrained ACGAN seismic facies identification method, characterized in that Applied to the situation where the number of fault data in actual seismic data is much less than that of non-fault data; the implementation process includes: S1. Normalize the seismic data to obtain a training set and a test set; S2. Use the improved ACGAN to expand the training set; the improved ACGAN described in step S2 is specifically: extract the features of the unlabeled seismic data to replace the noise input in the ACGAN; the training process of the improved ACGAN is specifically: S21. Mark the seismic data with faults in the result of normalizing the seismic data as labeled seismic data; those without faults are marked as unlabeled seismic data; S22. Extract the features of the unlabeled seismic data; S23. Train the ACGAN. Specifically: after fusing the features extracted in step S22 with randomly generated label information, use it as the input of the generator in the ACGAN; use the labeled seismic data as the input of the discriminator in the ACGAN; S24. Select the generated labeled seismic data set after the discriminator accuracy reaches 50% to expand the training set of the CNN network; S3. Train the CNN network according to the expanded training set; S4. Input the test set into the trained CNN network to obtain the seismic facies recognition result.

2. The geological feature-constrained ACGAN seismic facies identification method according to claim 1, wherein The extraction of the features of the unlabeled seismic data is specifically: use a two-dimensional convolutional layer to perform a convolutional operation on the unlabeled seismic data to obtain a feature image.

3. The method for identifying seismic facies of ACGAN constrained by geological features according to claim 2, wherein In the training set of the expanded CNN network in step S3, the ratio of labeled real data to unlabeled seismic data is 1:1.

Images