Unsupervised seismic facies analysis method based on lognormal mixture variational autoencoder

Abstract

The application discloses an unsupervised seismic facies analysis method based on a lognormal mixture variational autoencoder, and comprises the following steps: obtaining prestack seismic data, performing data preprocessing, making a data set, constructing a deep clustering model of the LMVAE, iteratively training the data set by using the deep clustering model of the LMVAE, completing unsupervised prestack seismic data reflection mode analysis, predicting a seismic facies category, and generating a prestack seismic facies map. The method models the lognormal mixture probability of the seismic data in the feature, solves the limitation of the asymmetric data in the deep feature space distribution in the seismic reflection mode analysis, and simultaneously, for the purpose of simplifying the model solution, an inference model using a reparameterization skill for direct optimization is constructed, the inference difficulty problem of the deep generation model under a complex latent structure is overcome, the accuracy of the seismic facies map is improved, and thus strong technical support is provided for the prestack seismic data reflection mode analysis.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of seismic exploration, and particularly relates to an unsupervised seismic facies analysis method based on a lognormal mixed variational autoencoder. BACKGROUND

[0002] Fossil fuels have been the main component of world energy and play a key role in daily life, such as providing the needs of clothing, food, shelter and other aspects. Although China has made significant progress in oil resource exploration in recent years, making the oil and gas production continue to grow, but still cannot meet the demand of economic growth. This is mainly because the simple seismic exploration technology used in the early stage has been unable to meet the requirements of modern exploration. Those oil and gas that can be found by simple exploration technology have been exhausted, and the remaining oil and gas exploration is more difficult. Therefore, modern exploration technology has become the main means of oil and gas exploration and development, and its main principle is to obtain multi-dimensional and large amount of seismic signals by artificial excitation of seismic waves, and to analyze the structure and characteristics of underground rock layers by using the reflection and refraction phenomena generated by the seismic signals when propagating in the underground rock. By dividing the seismic signals into different reflection patterns, the identification of oil reservoirs is completed.

[0003] Seismic reflection pattern analysis usually includes two important steps, namely feature extraction and clustering. Early researchers optimized the two steps separately. The purpose of feature extraction is to reduce the dimension and feature representation of high-dimensional data to retain the effective features in the data, so as to improve the accuracy of reflection pattern analysis. Subsequently, clustering algorithms are applied to cluster the features to generate facies maps. However, the isolated optimization of feature extraction and clustering may lead to suboptimal results and increase the complexity of analysis. With the development of deep learning, it is proposed to combine feature learning and clustering into a unified framework, which can directly cluster the original seismic images. This method is called deep clustering.

[0004] Deep clustering is a new clustering method that can directly cluster the original seismic image. This method combines feature learning and clustering into a unified framework to learn a more beneficial feature representation for clustering. Deep clustering mainly includes methods based on different network structures such as feedforward neural network, auto-encoder network, variational auto-encoder network and generative adversarial network. Among them, the document“Xie.J, Girshick.R, Farhadi.A. Unsupervised deep embedding for clustering analysis. Proceedings of the 33rd International Conference on Machine Learning, 2016: 478-487.” can extract the embedded features of the input by applying the neural network in deep learning and design the clustering layer by using KL divergence (Kullback-Leibler Divergence), which can complete feature extraction and clustering at the same time. In order to more accurately identify the ancient river channel features, the document“Duan.Y, Zheng.X, Hu.L, et al. Seismic facies analysis based on deep convolutional embedded clustering. Geophysics, 2019, 84(6): 87-97.” optimizes the KL clustering loss function, which significantly improves the performance of the algorithm.

[0005] The existing deep clustering model normal mixture variational auto-encoder (GMVAE) assumes that the seismic data in the deep feature space (i.e. latent space projection) conforms to the Gaussian mixture distribution, but the latent space projection of the seismic data is not necessarily a symmetric distribution like the Gaussian distribution, so the GMVAE cannot accurately represent the asymmetrically distributed seismic data in the deep feature space, and the deep clustering model has the deficiency in processing complex potential structure data, especially in the inference in the complex potential structure model. SUMMARY

[0006] To solve the above technical problems, the present application provides an unsupervised seismic facies analysis method based on a lognormal mixture variational auto-encoder. By introducing a variational auto-encoder, using a lognormal mixture distribution instead of a Gaussian mixture distribution, a deep clustering model based on a lognormal mixture variational auto-encoder (LMVAE) is constructed, which more accurately analyzes the asymmetrically distributed seismic data in the deep feature space, realizes the deep clustering of prestack seismic data, and thus completes the reflection pattern analysis. In order to simplify the model solution, a reasoning model using a reparameterization technique for direct optimization is constructed, thereby providing strong technical support for prestack seismic data reflection pattern analysis.

[0007] The technical scheme of the present application is: an unsupervised seismic facies analysis method based on a lognormal mixture variational autoencoder, and the specific steps are as follows:

[0008] S1. Obtain prestack seismic data, perform data preprocessing, and make a data set;

[0009] S2. Construct a deep clustering model of the LMVAE, and use the deep clustering model of the LMVAE to iteratively train the data set;

[0010] S3. Based on step S2, complete unsupervised prestack seismic data reflection mode analysis, predict seismic facies categories, and generate a prestack seismic facies map.

[0011] Further, the step S1 is specifically as follows:

[0012] According to the actual work area data horizon and the addition of time window data, four-dimensional prestack seismic data are obtained, and using a deep learning method, the four-dimensional seismic data are split into two-dimensional seismic data according to CDP for processing as a training set input into the network.

[0013] Further, the step S2 is specifically as follows:

[0014] S21. Construct a deep clustering model network of the LMVAE;

[0015] The deep clustering model of the LMVAE is composed of a generation model and an inference model, the network structure of the deep clustering model of the LMVAE is divided into an inference network structure module and a generation network structure module, the inference model network is composed of two convolutional layers, the convolution kernel size is 3x3, the deep features are extracted from the input prestack seismic data by using the two convolutional layers, then the features are input into a fully connected layer to generate latent variables u, w and z, each convolutional layer adopts a LeakyReLU activation function, and a batch normalization layer is added after the convolutional layer to speed up the training convergence, and a maxpool layer is introduced to select the most effective features between the two convolutional layers.

[0016] The generation model adopts a network structure basically symmetrical to the inference model, uses two transpose convolutional layers and a LeakyReLU activation function to infer the observed data x.

[0017] Wherein, x represents a CDP set of prestack seismic data; the kernel size of the first transpose convolutional layer is the same as that of the first convolutional layer of the inference model, which is used to map the latent variables back to the feature space; the kernel size of the second transpose convolutional layer is the same as that of the second convolutional layer of the inference model, which is used to convert the data in the feature space into the original data. Like the inference model, the generation model also introduces a batch normalization layer and an up-sampling layer.

[0018] S22. Use the LMVAE deep clustering model to iteratively train the dataset;

[0019] Generative model p of deep clustering model of LMVAE β,θ (x,u,w,z) represents the following:

[0020] p β,θ (x,u,w,z)=p(w)p(z)p β (u|z,w)p θ (x|u) (1)

[0021]

[0022] z~Mult(π) (3)

[0023]

[0024]

[0025] Where β and θ represent neural network parameters, K represents the pre-defined number of seismic phase classifications, and z k Let represent the k-th element of the latent variable z, and its magnitude represents the mixing weight of the k-th log-normal distribution in the mixture log-normal distribution. Let p(w) represent the probability density of the latent variable w, and p(z) represent the probability density of the latent variable z. β (u|z,w) represents the probability density of latent variable u based on latent variables z and w, given by a neural network with parameter β, p θ (x|u) represents the probability density of latent variable w based on latent variable u, given by a neural network with parameter θ. x represents the CDP set of pre-stack seismic data. Let represent a Gaussian distribution, Mult represent a multivariate Bernoulli distribution, and the latent variable z represent a vector following Mult and sampled from the mixture probability π, i.e., the seismic phase category corresponding to the CDP set. u represents a continuous latent variable following a log-normal mixture distribution under z and w conditions, i.e., the projection of the CDP set points into the latent feature space; I represents the identity matrix. The neural network is parameterized using θ and β respectively to obtain... and Let represent the mean and variance of the k-th log-normal distribution in a mixture of log-normal distributions based on w and β. Let represent the mean and variance in a mixed log-normal distribution based on u and θ.

[0026] Generative models are trained using variational inference objectives, namely, the lower bound of logarithmic evidence. The expression is as follows:

[0027]

[0028] where q(u, w, z | x) denotes the mean-field variational family, E[·] denotes expectation.

[0029] Setting the mean-field variational family q(u, w, z | x) as a proxy for the posterior, the posterior is decomposed as:

[0030]

[0031] where i denotes the index of the data point, φ u and φ w denote the parameters of the neural network, denotes the latent variable u i , w i The best estimate of the variational posterior is given by the neural network with parameters φ u and φ w , p β (z i |u i , w i ) denotes the probability density of the latent variable z i based on the latent variables u i , w i , given by the neural network with parameters β.

[0032] Simplify the notation by omitting i, and only consider one data point at a time, use to identify the network φ u and φ w to parameterize each variational factor, output the parameters of the variational distribution, and specify their form as log-normal posteriors, derive the posterior p β (z | u, w) of z as follows:

[0033]

[0034] where z j denotes the j-th element of the latent variable z, and π j is the value of p(z j = 1), π k is the value of p(z k = 1).

[0035] When , select m samples from the input data x, and then generate new samples from the latent variables u, w, and z generated by the generative model Train the model by the lower bound of the log probability:

[0036]

[0037] where, denotes the value of the rth iteration, ε denotes the preset convergence precision, KL denotes the KL divergence, and the terms in ELBO are respectively referred to as a reconstruction term, a conditional prior term, a prior term of w, and a prior term of z, λ, γ, and η denote preset balance parameters, and in addition to λ, the weights of different priors are controlled by using γ and η. The value of the rth iteration is basically unchanged, that is, convergence.

[0038] Finally, the network is iteratively optimized by using a Monte Carlo sampling technique and an adam optimization algorithm until the loss value converges.

[0039] Further, in the step S2, to avoid excessive regularization in the allocation of the LMVAE cluster, the following is specifically implemented:

[0040] The prior term of z in formula (9) is modified as:

[0041]

[0042] Wherein, τ represents a preset constant value.

[0043] Further, the step S3 is specifically implemented as follows:

[0044] Based on the step S2, the category of the data point is inferred by inferring which latent distribution pattern the hidden layer is generated from, thereby realizing deep clustering, completing pre-stack seismic data reflection pattern analysis, outputting a predicted seismic facies category z, and generating a pre-stack seismic facies map.

[0045] Advantages of the present application: the method of the present application obtains pre-stack seismic data, performs data preprocessing, makes a data set, constructs a deep clustering model of the LMVAE, iteratively trains the data set by using the LMVAE deep clustering model, completes unsupervised pre-stack seismic data reflection pattern analysis, predicts a seismic facies category, and generates a pre-stack seismic facies map. The method of the present application models the lognormal mixture probability of the characteristics of the seismic data, solves the limitations of the asymmetrically distributed data in the deep feature space in the seismic reflection pattern analysis, at the same time, for the purpose of simplifying the model solution, a reasoning model using a reparameterization technique for direct optimization is constructed, the problem of difficult reasoning of the deep generative model under a complex latent structure is overcome, the accuracy of the seismic facies map is improved, thereby providing strong technical support for the pre-stack seismic data reflection pattern analysis. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of an unsupervised seismic facies analysis method based on a lognormal mixture variational autoencoder of the present application.

[0047] Figure 2 A network structure diagram of a deep clustering model based on the LMVAE in the embodiment of the present application.

[0048] Figure 3 This is a schematic diagram of the LMVAE inference model in an embodiment of the present invention.

[0049] Figure 4 This is a schematic diagram of the LMVAE generation model in an embodiment of the present invention.

[0050] Figure 5 This is a statistical diagram of pre-stack seismic data characteristics in an embodiment of the present invention.

[0051] Figure 6 This is a diagram illustrating the seismic phase effect of the physical model in an embodiment of the present invention.

[0052] Figure 7 This is a seismic phase effect diagram of the ZJ work area in an embodiment of the present invention. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 The flowchart of an unsupervised seismic facies analysis method based on a log-normal hybrid variational autoencoder of the present invention is shown below. The specific steps are as follows:

[0055] S1. Obtain pre-stack seismic data, perform data preprocessing, and create a dataset;

[0056] S2. Construct a deep clustering model for LMVAE and use the LMVAE deep clustering model to iteratively train the dataset;

[0057] S3. Based on step S2, perform unsupervised pre-stack seismic data reflection mode analysis, predict seismic facies categories, and generate pre-stack seismic facies maps.

[0058] In this embodiment, step S1 is specifically as follows:

[0059] Four-dimensional pre-stack seismic data is obtained by adding time window data to the actual work area data. Using deep learning methods, the four-dimensional seismic data is split into two-dimensional seismic data according to CDP and input into the network as the training set.

[0060] In this embodiment, step S2 is specifically as follows:

[0061] S21. Construct a deep clustering model network for LMVAE;

[0062] like Figure 2As shown in the figure, the network structure diagram of the LMVAE-based deep clustering model, the LMVAE-based deep clustering model is composed of a generation model and an inference model, in order to extract the depth features of the prestack seismic data, the network reduces the dimension of the prestack seismic data through multiple convolutional layers, and has good feature extraction capability. At the same time, inspired by the GMVAE with a fully connected layer structure, the LMVAE with a multi-layer convolutional structure and a fully connected layer structure is applied to the reflection pattern analysis of the prestack seismic data.

[0063] The network structure of the LMVAE-based deep clustering model is divided into an inference network structure module and a generation network structure module, as shown in Figure 3 As shown in the figure, the inference model network is composed of two convolutional layers, and the convolution kernel size is 3*3, the two convolutional layers are used to extract deep features from the input prestack seismic data, and then the features are input into the fully connected layer to generate latent variables u, w and z, each convolutional layer adopts a LeakyReLU activation function, and a batch normalization layer is added after the convolutional layer to speed up the training convergence, and a maxpool layer is introduced to select the most effective features between the two convolutional layers.

[0064] In order to accurately reconstruct the original seismic data, as shown in Figure 4 As shown in the figure, the generation model adopts a network structure basically symmetrical to the inference model, uses two transpose convolutional layers and LeakyReLU activation functions to infer the observed data x.

[0065] Wherein, x represents a CDP (Common Depth Point) set of prestack seismic data; the kernel size of the first transpose convolutional layer is the same as that of the first convolutional layer of the inference model, which is used to map the latent variable back to the feature space; the kernel size of the second transpose convolutional layer is the same as that of the second convolutional layer of the inference model, which is used to convert the data in the feature space into the original data. Like the inference model, the generation model also introduces a batch normalization layer and an up-sampling layer to improve the stability and non-linear expression ability of the network.

[0066] S22, iteratively training the data set using the LMVAE deep clustering model;

[0067] The lognormal mixture variational autoencoder (LMVAE) and the ordinary variational autoencoder (VAE) are usually composed of a generation model and an inference model. However, the ordinary VAE assumes that the latent variable obeys a Gaussian distribution, and cannot directly obtain the classification of the latent variable, while the GMVAE selects a mixed Gaussian prior as the prior, and infers the category of the data point, which is equivalent to inferring which mode of the latent distribution it is generated from. Similar to the GMVAE generation model, the LMVAE model finds out the lognormal distribution of which mode the latent variable belongs to, which is equivalent to completing clustering.

[0068] Single latent variable conforming to the lognormal mixture distribution increases the difficulty of model training, and the deep clustering model of the LMVAE uses one or more latent variables conforming to the lognormal distribution and one latent variable conforming to the categorical distribution to jointly represent the lognormal mixture distribution of the latent space.

[0069] The generating model p of the deep clustering model of the LMVAE β,θ (x, u, w, z) is represented as follows:

[0070] p β,θ (x, u, w, z) = p(w)p(z)p β (u|z, w)p θ (x|u) (1)

[0071]

[0072] z ~ Mult(π) (3)

[0073]

[0074]

[0075] where β, θ represent neural network parameters, K represents the number of categories of seismic facies preset, z k represents the kth element of the latent variable z, and its size represents the mixing weight of the kth lognormal distribution in the mixed lognormal distribution, p(w) represents the probability density of the latent variable w, p(z) represents the probability density of the latent variable z, p β (u|z, w) represents the probability density of the latent variable u based on the latent variables z and w, which is given by a neural network with parameters β, p θ (x|u) represents the probability density of the latent variable w based on the latent variable u, which is given by a neural network with parameters θ, x represents a CDP set of prestack seismic data, N represents a Gaussian distribution, Mult represents a multinomial distribution, the latent variable z represents a vector subject to Mult and sampled from the mixing probability π, that is, the seismic facies category corresponding to the CDP set, u represents a continuous latent variable subject to the lognormal mixture distribution under the conditions of z and w, that is, the projection of the CDP set point in the latent feature space; I represents a unit matrix. The neural networks are parameterized by θ and β respectively to obtain and represents the mean and variance of the kth lognormal distribution in the mixed lognormal distribution based on w and β, represents the mean and variance in the mixed lognormal distribution based on u and θ.

[0076] The generating model is trained using the variational inference objective, that is, the lower bound of the log evidence (Evidence Lower Bound, ELBO), which is expressed as follows:

[0077]

[0078] where q(u, w, z|x) represents the mean-field variational family, and E[·] represents expectation.

[0079] Setting the mean-field variational family q(u, w, z|x) as a proxy for the posterior, the posterior is decomposed as:

[0080]

[0081] where i represents the index of the data point, and φ u and φ w represent the parameters of the neural network, represent the latent variables u i , w i The best estimate of the variational posterior is given by the neural network with parameters φ u and φ w , p β (z i |u i , w i ) represents the probability density of the latent variable z i based on the latent variables u i , w i , which is given by the neural network with parameters β.

[0082] Further simplify the symbols by omitting i, and only consider one data point at a time, and use to identify the network φ u and φ w Parameterize each variational factor, output the parameters of the variational distribution, and specify their form as a log-normal posterior, and derive the posterior p β (z|u, w) of z as follows:

[0083]

[0084] where z j represents the jth element of the latent variable z, and π j is the value of p(z j = 1), and π k is the value of p(z k = 1).

[0085] When , select m samples from the input data x, and then generate new samples from the latent variables u, w, and z generated by the generative model Train the model by the lower bound of the log probability:

[0086]

[0087] wherein, denotes the value of the rth iteration, ε denotes the preset convergence precision, KL denotes the KL divergence, the terms in ELBO are respectively called the reconstruction term, the conditional prior term, the prior term of w and the prior term of z, λ, γ, η denote the preset balance parameters, and the weights of different priors are controlled by using γ and η except λ. The value of is basically unchanged with iteration, that is, convergence.

[0088] Finally, the network is iteratively optimized by using the Monte Carlo sampling technique and the adam optimization algorithm until the loss value converges.

[0089] In the embodiment, in step S2, to avoid excessive regularization in the allocation of the LMVAE cluster, the following is specifically performed:

[0090] In some studies of VAE, it has been proved many times that the regularization term will affect the training of VAE. This is due to the strong influence of the prior, which leads to the obtained latent representation being too simplified and not being able to well represent the underlying structure of the data. There are currently two main methods to overcome this influence: one is to anneal the KL term during training, allowing the reconstruction term to train the autoencoder network, and then slowly incorporating the regularization of the KL term into the training. The other method is to use a threshold to achieve the purpose of modifying the objective function, when the KL term is lower than a certain threshold, the threshold removes the influence of the KL term.

[0091] The prior term of z in equation (9) is modified as follows:

[0092]

[0093] wherein τ denotes a preset constant value.

[0094] In the embodiment, step S3 is specifically performed as follows:

[0095] Based on step S2, the class of the data point is inferred by inferring which latent distribution pattern the hidden layer is generated from, thereby realizing deep clustering, completing the analysis of the reflection pattern of prestack seismic data, outputting the predicted seismic facies class z, and generating a prestack seismic facies map.

[0096] In this embodiment, the method of the present application also verifies that the latent space projection of pre-stack seismic data obeys a non-normal distribution, and the GMVAE selects a mixed Gaussian prior as the prior of the latent space projection as an intuitive extension of the unimodal Gaussian prior. However, in reality, the true seismic feature cluster may not follow a symmetric Gaussian distribution. This situation may be reflected in the pre-stack seismic reflection pattern analysis results, resulting in the predicted data missing some details, thereby ultimately affecting the performance of the cluster. Considering that the distribution of the pre-stack seismic data corresponding to the features may be non-Gaussian, a quantile-quantile plot is used to evaluate the probability distribution of these features. As shown in Figure 5 Figure 1 shows the pre-stack seismic data feature statistics, Figure 5 (a) represents a quantile-quantile plot of the pre-stack seismic data features and a Gaussian distribution sequence, Figure 5 (b) represents a quantile-quantile plot between the pre-stack seismic data features and a lognormal distribution sequence. By observing these figures, it is found that in some cases, the feature distribution of the pre-stack seismic data is closer to the lognormal distribution than to the Gaussian distribution. This finding suggests that when modeling the pre-stack seismic data, a distribution model that is more suitable for describing its feature distribution may need to be used, rather than being limited to the Gaussian distribution.

[0097] As shown in Figure 6 In this embodiment, a physical model experiment is performed, and the experimental data contains two simulated areas, a fracture-vug model area and a fault model area. The fracture-vug model area is located in the north, and the stratum thickness is about 120 meters. The top is distributed with fracture-vugs, including several groups of fracture-vugs with different lengths, widths, spacings, densities, inclinations, and azimuths.

[0098] The effects of the four methods on the physical model data will be compared below. In these methods, the existing feature extraction algorithm uses the SOM unsupervised clustering algorithm, and the number of classifications for all experiments is set to 3 (according to the actual geological division into three rock patterns), so that the results of the four methods can be compared. Figure 6 The root mean square amplitude attribute map of the physical model data in (a) is the comparison standard for the effect of the algorithm. The four methods are: LMVAE, whose processed seismic facies map is shown in Figure 6 (b), GMVAE, whose processed seismic facies map is shown in Figure 6 (c), DCAE feature extraction + SOM clustering, whose processed seismic facies map is shown in Figure 6 (d), and Gabor feature extraction + SOM clustering algorithm, whose processed seismic facies map is shown in Figure 6 (e).

[0099] In Figure 6 (a), for the interpretation of the root mean square amplitude, first observe the fracture-vug area. From the DCAE + SOM algorithm ( Figure 6 (d)) and the Gabor + SOM algorithm (Figure 6 In the comparison of (e), it can be seen that the DCAE+SOM algorithm can clearly identify the outline of the cavity region, and the outline is relatively complete. In contrast, the Gabor+SOM algorithm's effect is that the cavity region does not present an overall elliptical shape and is relatively incomplete, making it inferior to the DCAE+SOM algorithm. Next, observing the fault region, it can be found that while both the DCAE+SOM and Gabor+SOM algorithms can show the outline of the fault, neither can clearly identify the fault itself. Overall, the DCAE+SOM algorithm outperforms the Gabor+SOM algorithm, which also demonstrates the advantages of deep learning methods. However, the DCAE+SOM algorithm also has significant identification errors in the cavity and fault regions. For example, in the root mean square amplitude attribute of the physical model, the DCAE+SOM algorithm produces a cluttered background and incorrect classification results. Conversely, GMVAE ( Figure 6 (c)) and LMVAE( Figure 6 Method (b) produces a much cleaner background in both fracture and fault areas. Figure 6 As can be seen, the GMVAE method is significantly superior to the DCAE+SOM and Gabor+SOM methods because the fracture-cavity shape is a more complete ellipse, and the continuity and clarity of the fault are better. However, compared with the LMVAE method, the GMVAE method still has obvious shortcomings. In the fracture-cavity region, the background of the LMVAE method is cleaner, meaning the classification result is more accurate, and the fracture-cavity shape is more complete; in the fault region, the background of the LMVAE method is also cleaner, and the fault is presented more continuously and clearly. In summary, the seismic facies results of the LMVAE method are closer to the physical model data and are superior to the other four comparative methods.

[0100] like Figure 6 As shown, in this embodiment, in order to further verify the effectiveness of the method of the present invention in practical applications, pre-stack seismic data tests were conducted in actual work areas ZJ and LZB, respectively.

[0101] The actual ZJ seismic data is 4D pre-stack seismic data located in the ZJ seismic area of ​​the Sichuan Basin in southwestern China, covering an area of ​​approximately 40 square kilometers. The following section compares the effectiveness of four methods on the ZJ seismic data. Figure 7 (a) The root mean square amplitude attribute map of the ZJ pre-stack seismic data serves as the benchmark for comparing algorithm performance. Among these methods, the existing feature extraction algorithm employs the SOM unsupervised clustering algorithm, setting the number of classifications for all experiments to 4 (derived based on experience and the Davidson-Bolding index, a classification fitness metric). The four methods are: LMVAE, whose processed seismic phase map is shown below. Figure 7 (b) GMVAE, the seismic phase diagram obtained from its processing is as follows:Figure 7 (c), DCAE feature extraction + SOM clustering, the processed seismic facies map is as follows Figure 7 (d), Gabor feature extraction + SOM clustering algorithm, the processed seismic facies map is as follows Figure 7 (e).

[0102] In Figure 7 (a), the interpretation of the root mean square amplitude. First, observe the small fault area, through the comparison of the DCAE + SOM method ( Figure 7 (d)) and the Gabor + SOM method ( Figure 7 (e)), it can be found that the DCAE + SOM algorithm can better identify the contour of the small fault area, and the contour is relatively complete, while the Gabor + SOM method also identifies the contour, but the background is cluttered, and there are many classification errors. Secondly, for large fault areas, the DCAE + SOM method can see the contour of the large fault, but it is not very clear, while the Gabor + SOM method basically cannot see the contour of the river channel, and the overall background is more cluttered, and the classification is wrong. In summary, the DCAE + SOM method is better than the Gabor + SOM method, but there are still some obvious recognition errors, especially in the feature recognition of large fault areas. On the contrary, the GMVAE method ( Figure 7 (c)) and the LMVAE method ( Figure 7 (b)) have cleaner backgrounds in small fault areas and large fault areas, and the fault features are more continuous. From Figure 7 Figure 7 it can be seen that the GMVAE method is significantly better than the DCAE + SOM and Gabor + SOM methods, but compared with the LMVAE method, there are still some deficiencies. Specifically, in the small fault area, the background of the LMVAE method is cleaner, the classification result is more correct, and the small fault is more continuous than the GMVAE method; in the large fault area, the LMVAE method can clearly identify the large fault and correctly classify it, while the GMVAE method can only identify the fault contour, but it is not clear and continuous. In summary, the seismic facies result of the LMVAE method is closer to the actual ZJ work area data, so the method of the present application is better than the other three comparison methods.

[0103] In summary, the method of the present application obtains clearer and more accurate seismic facies maps in physical model data and actual work area data, and compared with other comparison methods, the clustering effect shows obvious improvement. Therefore, the main contribution of the method of the present application is to propose a new deep clustering model to effectively solve the problem of asymmetric distribution of deep feature space of seismic data, and simplify the solving process of the inference model, thereby providing strong technical support for prestack seismic data reflection pattern analysis.

[0104] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and understanding of the principles of the application and should not be construed as limiting the scope of the application to such specifically enumerated embodiments. This application can be modified and varied in various ways, and it is therefore to be understood that the application can be practiced otherwise than specifically set forth herein. Any modifications, equivalent substitutions, improvements, etc., made to the application in accordance with the spirit and principles of the application are to be encompassed within the scope of the application as set forth in the following claims.

Claims

1. An unsupervised seismic facies analysis method based on a log-normal hybrid variational autoencoder, the specific steps of which are as follows: S1. Obtain pre-stack seismic data, perform data preprocessing, and create a dataset; S2. Construct a deep clustering model for LMVAE and use the LMVAE deep clustering model to iteratively train the dataset; The deep clustering model of the LMVAE is composed of a generation model and an inference model; the generation model of the deep clustering model of the LMVAE is represented as follows: （1）； （2）； （3）； （4）； （5）； wherein, , denote the neural network parameters, denote the number of classes pre-set for seismic facies, denote the k-th element of latent variable z, whose size represents the mixing weight of the k-th lognormal distribution in the mixture lognormal distribution, denote the probability density of latent variable w, denote the probability density of latent variable z, denote the probability density of latent variable u based on latent variables z, w, which is given by a neural network with parameters , denote the probability density of latent variable w based on latent variable u, which is given by a neural network with parameters , denote the CDP set of pre-stack seismic data, denote the Gaussian distribution, denote the multivariate Bernoulli distribution, latent variable denote the vector sampled from the mixture probability , i.e., the seismic facies class corresponding to the CDP set, denote the continuous latent variable of the lognormal mixture distribution under the conditions of and , i.e., the projection of the CDP set point in the latent feature space; denote the identity matrix; the neural networks are parameterized by and , , and , , denote the mean and variance of the k-th lognormal distribution in the mixture lognormal distribution based on and , , denote the mean and variance in the mixture lognormal distribution based on and , S3. Based on step S2, perform unsupervised pre-stack seismic data reflection mode analysis, predict seismic facies categories, and generate pre-stack seismic facies maps.

2. The method of claim 1, wherein the method is based on a lognormal mixture variational autoencoder. The specific steps of S1 are as follows: Four-dimensional pre-stack seismic data is obtained by adding time window data to the actual work area data. Using deep learning methods, the four-dimensional seismic data is split into two-dimensional seismic data according to CDP and input into the network as the training set.

3. The method of claim 1, wherein the method is based on a lognormal mixture variational autoencoder. The specific details of step S2 are as follows: S21. Construct a deep clustering model network for LMVAE; The network structure of the deep clustering model of the LMVAE is divided into an inference network structure module and a generation network structure module, the inference model network is composed of two convolutional layers, the convolution kernel size is , deep features are extracted from the input pre-stack seismic data by the two convolutional layers, then the features are input into a fully connected layer to generate latent variables , and , a LeakyReLU activation function is used for each convolutional layer, a batch normalization layer is added after the convolutional layer to speed up the training convergence, and a maxpool layer is introduced to select the most effective features between the two convolutional layers; The generative model employs a network structure substantially symmetrical to the inference model, using two transpose convolutional layers and LeakyReLU activation functions to infer the observation data ; wherein, represents a CDP set of pre-stack seismic data; the kernel size of the first transpose convolution layer is the same as the first convolution layer of the inference model, which is used to map the latent variable back to the feature space; the kernel size of the second transpose convolution layer is the same as the second convolution layer of the inference model, which is used to convert the data in the feature space to the original data; like the inference model, the generation model also introduces a batch normalization layer and an up-sampling layer; S22. Use the LMVAE deep clustering model to iteratively train the dataset; The generative model is trained with a variational inference objective, i.e., a lower bound on the log evidence , which is expressed as follows: （6）； wherein denotes the family of mean field variational, denotes expectation; Setting the mean-field variational family As a proxy for the posterior, the posterior is then decomposed as: （7）； wherein, denotes an index over data points, and denotes a parameter of a neural network, denotes a latent variable the best estimate of the variational posterior given by the neural network with parameters and denotes a latent variable the probability density based on the latent variable given by the neural network with parameters ;​ Simplified notation will omit, consider only one data point at a time, with the recognition network and Parametrizing each variational factor, outputting the parameters of the variational distribution, and specifying their form as a log-normal posterior, derive the posterior as follows: （8）； wherein represents a latent variable the jth element in is the value of is the value of When the input data is selected a sample, then according to the generated model generated latent variables , and , generate a new sample , by the lower bound of the logarithmic probability to train the model: （9）； wherein, denotes the value of the rth iteration, denotes a pre-set convergence precision, denotes divergence, the terms in ELBO are called reconstruction term, conditional prior term, prior term of and prior term of , , denotes a pre-set balance parameter, in addition to , the weights of different priors are controlled by and ; the value of is basically unchanged with iteration, i.e. convergence; Finally, the network is iteratively optimized using Monte Carlo sampling techniques and the Adam optimization algorithm until the loss value converges.

4. The method of claim 3, wherein the method is based on a lognormal mixture variational autoencoder. In step S2, to avoid excessive regularization in the allocation of the LMVAE cluster, the specific steps are as follows: Modifying the prior term in equation (9): ​ （10）； wherein represents a constant value set in advance.

5. The method of claim 1, wherein the method is based on a lognormal mixture variational autoencoder. Step S3 is as follows: Based on step S2, the deep clustering is realized by inferring the data points are generated by which potential distribution pattern through the hidden layer, and then inferring the category of the data points, the pre-stack seismic data reflection pattern analysis is completed, and the predicted seismic facies category is output , and a pre-stack seismic facies map is generated.

Images