A prestack seismic facies analysis method for decoupling specific physical properties

Abstract

The present invention discloses a prestack seismic facies analysis method for decoupling specific physical properties, which is applied to the field of seismic exploration. Aiming at the problem that the existing seismic facies analysis methods are difficult to separate the influences of different geological and geophysical factors; the present invention considers taking the seismic, logging data and geological characteristic laws of the actual work area as constraints, based on the deep learning theory, and using the decoupled representation learning method as a means to decouple and separate the specific geological and geophysical factors that are coupled and interact with each other in the complex prestack seismic signal, so as to improve the accuracy and reliability of using artificial intelligence methods for prestack seismic facies analysis.

Description

Technical Field

[0001] The present invention belongs to the field of oil and gas exploration, and particularly relates to a prestack seismic facies analysis technique. Background Art

[0002] Seismic exploration is one of the most effective means to solve oil and gas exploration problems. The seismic signals obtained through seismic exploration provide a necessary basis for the identification of oil and gas reservoirs. Seismic facies analysis technology is one of the important technologies for seismic signal analysis in oil and gas exploration. Seismic facies analysis is an effective technology for interpreting reservoir distribution, sedimentary facies, etc. based on seismic reflection data. It mainly identifies and maps seismic facies units according to a series of characteristic parameters of seismic wave reflections and interprets the reservoir types, sedimentary facies, sedimentary systems, etc. represented by these seismic facies. In the initial period, seismic facies analysis was mainly carried out by the "facial method" of professional personnel. This method is often affected by subjectivity and requires interpreters to have a high professional quality in oil and gas geology and geophysics. With the rapid development of computer technology and pattern recognition theory, a series of supervised classification methods such as Bayesian algorithm, neural network algorithm or unsupervised classification methods such as Self-organization mapping (SOM), K-means clustering have been widely used in seismic facies analysis. A large number of practical applications have shown that the existing method technologies have significantly improved the efficiency and accuracy of seismic facies analysis.

[0003] The seismic facies analysis technology based on post-stack seismic data has achieved good application effects in reservoir prediction, sedimentary facies inference, etc. However, using the stacking processing technology to obtain post-stack seismic data is also likely to cause the loss of information about lithology, fluid, etc., and the distortion of local signals contained in prestack seismic data. Prestack seismic data is the reflected seismic signals received by surface geophones at different shot-receiver offsets and different observation azimuths for the same reflection position. Compared with post-stack seismic data, prestack seismic data contains more abundant information such as lithology, fluid types, porosity, and anisotropy, which can be used for more refined oil and gas reservoir interpretation and description. In order to meet the requirements of more refined reservoir description, fluid identification, sedimentary facies inference, etc., extracting the characteristics of relevant geological factors directly from prestack seismic data has become an inevitable development direction. With the rapid development of computer hardware technology and related intelligent algorithms in the fields of artificial intelligence and machine learning in recent years, carrying out seismic facies analysis based on prestack seismic data has received close attention from geologists and geophysicists, and a large amount of time and effort has been invested in the research and development of related technologies.

[0004] Existing pre-stack seismic facies analysis methods usually process the pre-stack seismic signals of the points to be analyzed by regarding them as an image. Based on this idea and technical methods such as classification and clustering, a series of pre-stack seismic facies analysis methods suitable for different geological conditions have been formed, such as pre-stack seismic facies analysis based on self-organizing feature mapping neural network, unsupervised seismic facies analysis based on deep convolutional auto-encoder, and pre-stack seismic facies analysis based on waveform sparse representation. As we all know, an image is often composed of the superposition and coupling of specific physical attributes such as color, texture, direction, spatial relationship, and azimuth, that is, the specific physical attributes do not exist independently in the presented image space. Geological and geophysical experts have clearly stated that pre-stack seismic signals contain rich information reflecting lithology, fluid properties, and anisotropy. Moreover, the characteristics generated by geological factors such as lithology, fluid properties, stratigraphic structure, and anisotropy contained in pre-stack seismic signals are similar to the specific physical attributes contained in the image, that is, the characteristics presented by various geological factors in pre-stack seismic data are mostly mutually influential and non-independent. For example, changes in lithology will cause the amplitude to change with the offset, and differences in fluid properties will also cause corresponding changes in the amplitude with the offset. In addition, geological factors such as reservoir thickness, gas-bearing property, and porosity also have significant differences in their contributions to the characteristics of pre-stack seismic signals. For example, changes in lithology and porosity play a major role in the waveform structure, while changes in fluid properties and water saturation have relatively weak effects on pre-stack seismic signals. In existing pre-stack seismic facies analysis methods, whether it is the use of simple classification algorithms or the application of deep learning, all perform classification, clustering, or regression processing on all the features extracted from pre-stack seismic signals, without effectively analyzing which features correspond to which geological factors, and it is difficult to effectively separate different geological factors hidden in pre-stack seismic data, resulting in the lack of targeted effectiveness of the seismic facies analysis results obtained based on these features, and the accuracy of reservoir description and sedimentary facies plane distribution inference needs to be further improved. The pre-stack seismic facies analysis technology based on all features developed based on pre-stack seismic data is difficult to reflect the due advantages of both pre-stack seismic data and the technology itself, thus making it difficult to play an ideal role in the application in the fields of reservoir description, sedimentary facies inference, etc. Although the pre-stack seismic facies analysis method based on deep learning can reveal the deep features in pre-stack seismic data, these deep features lack corresponding physical meanings, making the results obtained by using pre-stack seismic facies analysis not satisfactory to geological and geophysical experts. For example, for thin interbedded hydrocarbon reservoirs, the reflected waves from different interfaces interfere with each other, affecting the variation of the seismic reflection amplitude of a single interface with the incident angle. The thin layer tuning effect causes changes in the seismic reflection amplitude and frequency, seriously affecting the accuracy and reliability of the seismic facies map obtained by using pre-stack seismic facies analysis technology, and further having an adverse impact on the further interpretation by geological and geophysical experts.

[0005] Decoupled representation learning is a method that can separate different factors that are coupled with each other in data and meet certain conditions. In essence, the decoupled representation learning method combines the advantages of traditional mathematical modeling and machine learning modeling, and has great advantages in model interpretability, object generation and operation, and zero-shot learning. To solve the problems existing in the above prestack seismic facies analysis technology, based on the theory of decoupled representation learning, the present invention invents a prestack seismic facies analysis method oriented to the decoupling of specific physical attributes. The inventive method decouples and separates the seismic signal characteristics generated by different geological factors that are coupled and interact with each other in complex prestack seismic data for specific physical attributes that are coupled with each other in practical problems, and then processes them using relevant methods, which can effectively solve the problem of the coupled influence of multiple geological factors in prestack seismic facies analysis and improve the accuracy and reliability of the application of prestack seismic facies analysis technology in many aspects such as reservoir description, lithology identification, and fluid prediction.

[0006] For complex prestack seismic signals, the actual seismic records are generated by the mutual influence and coupling of many geological and geophysical factors such as reservoir thickness, gas content, porosity, etc. The characteristics presented by different geological and geophysical factors in the prestack seismic signals are coupled with each other, and it is easy to interfere with each other during analysis. The obtained seismic facies maps have strong multi-solution properties, which affect the application effect of the final seismic facies maps. On the one hand, the existing seismic facies analysis methods are difficult to separate the influences of different geological and geophysical factors. The traditional seismic facies analysis method based on amplitude analysis is also difficult to achieve ideal results under the condition of the mutual influence of multiple factors. On the other hand, many deep features extracted by deep learning are often difficult to explain their physical meanings. For example, for thin interbedded hydrocarbon reservoirs, the reflected waves from different interfaces interfere with each other, changing the variation of the seismic reflection amplitude of a single interface with the incident angle. Under the condition of thin interbeds, the reflection interfaces of geological bodies do not necessarily correspond to the seismic records, and the variation law is complex. If the thin layer formation attributes affecting the seismic records can be separated from other physical factors, the above problems can be effectively solved. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes a prestack seismic facies analysis method oriented to the decoupling of specific physical attributes. Based on the theory of decoupled representation learning and using deep learning methods, a new decoupled autoencoder network is built to realize the decoupling of specific physical attributes that are coupled with each other in complex prestack seismic signals.

[0008] The technical solution adopted by the present invention is: a prestack seismic facies analysis method oriented to the decoupling of specific physical attributes, including:

[0009] S1. Generate a training data set using a forward algorithm based on the Aki-Richards equation and the convolution model, based on the elastic parameters and formation structure characteristics of the logging data in the actual work area.

[0010] S2. Build a decoupled autoencoder network based on the convolutional autoencoder and the decoupling process A: Swap(·), and use the training data set generated in step S1 to train the decoupled autoencoder network. Train the decoupled autoencoder network.

[0011] S3. Import the actual work area data into the trained decoupled autoencoder network, extract the required decoupled sub-attributes, and perform clustering analysis on the decoupled sub-attributes to obtain the seismic facies analysis results.

[0012] Advantages of the present invention: Based on the theory of decoupled representation learning, the present invention proposes a new pre-stack seismic facies analysis method for decoupling specific physical attributes, which can achieve feature decoupling with clear geological and geophysical meanings for complex pre-stack seismic signals. Experiments on synthetic data and actual data prove that using this feature decoupling model can effectively separate the physically coupled and interacting attributes in the data, significantly improving the accuracy and effectiveness of pre-stack seismic facies analysis for specific applications, as well as the interpretability of deep learning models. Through the implementation of the invention technology, the characteristics generated by the geological and geophysical factors hidden in the pre-stack seismic signals are greatly revealed, and the accuracy of reservoir description and sedimentary facies spatial distribution inference can be greatly improved. Description of the Drawings

[0013] Figure 1 is the basic structure of the autoencoder;

[0014] Figure 2 is an example of a convolutional autoencoder;

[0015] Figure 3 is a schematic diagram of a set of data with different feature coupling relationships;

[0016] Figure 4 is a schematic diagram of the decoupled autoencoder network;

[0017] Figure 5 is the flow chart of the method of the present invention;

[0018] Figure 6 is the flow chart of network training;

[0019] Figure 7 is a schematic diagram of the elastic parameters of the test formation model;

[0020] Among them, (a) is the density parameter, (b) is the P-wave velocity parameter, and (c) is the S-wave velocity parameter;

[0021] Figure 8 are pre-stack data with three different randomly selected characteristic coupling relationships;

[0022] Among them, (a) is the pre-stack data waveform of water-bearing sandstone with a formation thickness of 5 m, (b) is the pre-stack waveform data of water-bearing sandstone with a bottom layer thickness of 20 m, and (c) is the pre-stack waveform data of gas-bearing sandstone with a formation thickness of 20 m;

[0023] Figure 9 is a decoupled autoencoder network model;

[0024] Figure 10 is a schematic diagram of the training model loss function;

[0025] Figure 11 is the clustering result of the decoupled sub-properties of the Kmeans decoupled autoencoder;

[0026] Among them, (a) is the effect of 5-clustering on the extracted thickness property, and (b) is the effect of 2-clustering on the extracted elastic parameter property;

[0027] Figure 12 is the clustering result of the depth property of the Kmeans ordinary autoencoder;

[0028] Among them, (a) is the effect of 5-clustering on the extracted depth property, and (b) is the effect of 2-clustering on the extracted depth property;

[0029] Figure 13 is the root mean square amplitude map of the target formation;

[0030] Figure 14 is a schematic diagram of the change of the loss function trained with actual data;

[0031] Figure 15 is the effect comparison between the decoupled sub-property clustering and the ordinary depth property clustering of the present invention;

[0032] Among them, (a) is the decoupled sub-property clustering map of the present invention, and (b) is the ordinary depth property clustering map. Detailed implementation manners

[0033] To facilitate the understanding of the technical content of the present invention by those skilled in the art, the following technologies are described first:

[0034] 1. Convolutional autoencoder network

[0035] The autoencoder is a typical unsupervised deep learning model, aiming to learn the abstract features of the input samples by reconstructing the input samples through the network output. As a neural network used in unsupervised learning, the input X of the autoencoder and the expected output X RAll are unlabeled samples, while the output h of the hidden layer is the abstract feature representation of the samples. The autoencoder first inputs the sample X, converts it into an efficient abstract feature representation h through the encoder, and then outputs the reconstruction X of the original sample through the decoder. R .

[0036] An autoencoder usually consists of two parts: an encoder and a decoder. The encoder maps high-dimensional input samples to low-dimensional abstract feature representations, achieving sample compression and dimensionality reduction; the decoder then converts the abstract feature representations into the desired output, achieving the reconstruction of the input samples, that is, it is expected that X≈X. R . The structure of the autoencoder is as Figure 1 shown.

[0037] Traditional autoencoders generally use fully connected layer neural networks. This method has no impact on the processing of one-dimensional signals, but for two-dimensional images or video signals, the fully connected layer will lose spatial relationship information. In image processing, a convolutional network is generally used as the encoder and a transposed convolutional network is used as the decoder. This type of autoencoder is called a convolutional autoencoder. Compared with the fully connected autoencoder, the convolutional autoencoder can effectively prevent a large amount of information loss caused by data slicing and data stacking. The structure of a simple convolutional autoencoder is as Figure 2 shown.

[0038] 2. Disentangled Representation Learning

[0039] Disentangled representation learning is a method of extracting the mutually coupled features in data and decomposing them into multiple independent factors. From the perspective of the modeling process of data generation, in representation learning, the generation process of real data x is usually modeled in two steps:

[0040] Step1: Sample the latent variable value z from the prior distribution p(z).

[0041] Step2: Sample the data observation value x from the conditional data generation distribution p(x|z).

[0042] Suppose the real data x can be regarded as generated by the mutual coupling of a series of physically semantically interpretable factors v1, v2,..., v n , through the action of a complex unknown non-linear system mapping function Sim(·), that is, x = Sim(v1, v2,..., v n ). The latent variable z in this model is the approximate representation of these physically interpretable factors v1, v2,..., v n . The conditional likelihood distribution p(x|z) is the approximation of the unknown non-linear system mapping function Sim(·) from a probabilistic perspective. On this basis, disentangled representation learning aims to learn a separable latent variable representation z = {z1, z2,..., z n}, and such that in this representation, each latent variable z k can effectively represent and control the corresponding generation factor v within the data k for effective characterization control.

[0043] Theoretically speaking, decoupled representation is to decompose the varying factors in the feature representation on a dataset that meets certain conditions. These varying factors are independent of each other, and each varying factor represents a certain physical semantic meaning. When a single varying factor changes, it will affect the change of a single factor in the generated data, while other factors remain unchanged. In the field of computer vision, decoupled representation learning has achieved a great deal of results in aspects such as image editing and generation, domain adaptation, video frame prediction, and 3D modeling.

[0044] The present invention considers taking the seismic, logging data, and geological feature laws of the actual work area as constraints, based on the deep learning theory, and using the decoupled representation learning method as a means to decouple and separate the specific geological and geophysical factors that are coupled and interact with each other in complex pre-stack seismic signals, so as to improve the accuracy and reliability of pre-stack seismic facies analysis using artificial intelligence methods.

[0045] The present invention uses a pair of data with a feature coupling relationship as input to train the deep feature decoupling network proposed by the present invention. Using the network constructed by the present invention, the physical relationships that interact with each other hidden in this set of data are separated, and pre-stack seismic facies analysis is carried out according to the extracted deep attributes, so as to eliminate the adverse effects of the mutual influence of different physical factors on the accuracy and reliability of pre-stack seismic facies analysis and realize the interpretability of the model. The present invention also gives an implementation method of this decoupled representation learning, that is, the decoupled convolutional autoencoder proposed by the present invention will be introduced next:

[0046] An ordinary convolutional autoencoder inputs a data X and outputs a reconstructed data X of X R , and trains the model by minimizing the gap between the input data X and the reconstructed data X R . In terms of composition, the convolutional autoencoder X→X R is composed of an encoder E:X→h and a decoder D:h→X R , where represents the front and back structural composition of the model. The network of the present invention uses a set of data from a space M with different attribute coupling relationships for network training and construction. Therefore, the network structure of the present invention can be regarded as composed of an encoder E:X×M→h×M and a decoder D:h×M→X R , where M is the space of data attribute coupling relationships. The data used by the present invention Where X contains (a batch of) training samples, and M represents the feature coupling relationship (sub)space between them. In the autoencoder, the deep attributes including the feature coupling relationship output by the encoder E are directly passed to the decoder In addition, the present invention also designs a decoupling process A to form The network training process is carried out to achieve feature decoupling while extracting deep attributes and obtain the decoupled sub-attributes of the data.

[0047] The implementation process of the present invention comprises the following steps:

[0048] 1. Construction of the training data and its attribute coupling relationship space (X, M)

[0049] To achieve decoupling of different physical properties, a single data input is difficult to meet our requirements for network input, because a single data cannot reflect the different feature coupling relationships between different data. The input of this invention is not a single data, but a set of multiple data with feature coupling relationships. Form a data pair. Thus, in the training network When , we want to be able to disentangle the potential output h of E so that we can use D to decode samples that were not provided to E.

[0050] Here, the present invention uses a pre-stack seismic data forward modeling method based on the elastic parameters and stratigraphic structural characteristics of the actual work area logging data and the Zoeppritz equation as the theoretical basis to achieve different specific physical properties. This method first extracts the target stratigraphic elastic parameters from the real data and uses Gaussian distribution to fit the data distribution of the elastic parameters. Monte Carlo sampling is used to obtain stratigraphic curves from the fitted elastic distribution. The desired stratigraphic model is obtained by constructing a geological model and simulating the stratigraphic bending trend. Using this stratigraphic model construction method, on the one hand, the model can be close to the actual data in terms of elastic parameter distribution; on the other hand, simulating stratigraphic bending and squeezing can also make the obtained stratigraphic model have basic stratigraphic sedimentology and geological tectonic theory support. Because the Zoeppritz equations are too complex to be applied in practice, the most widely used approximate equation, the Aki-Richards equation, and the convolution forward modeling method are used here to perform forward simulation of prestack seismic data with different specific physical properties, obtaining complex prestack seismic data containing multiple incident angles. For the specific process, please refer to: Aki KI, and Richards PGQuantitative seismology[M]. WHFreeman and Co, 1980: 135-137.

[0051] According to the application purpose of the technology, each set of training data of the present invention It is composed of 4 pre-stack seismic data. The 4 pre-stack seismic data are formed by coupling two different physical properties, but different data have different characteristic coupling relationships. The specific forms are as follows: Figure 3 As shown. Among them, x i Indicates different data, attribute mj indicates the jth value of the mth attribute, in this embodiment, j = 1, 2; specifically, in Figure 3 4 data are shown in , each data consists of two attributes, where data x1 is consistent with x3 in the first attribute and consistent with x4 in the second attribute; similarly, data x2 is consistent with x4 in the first attribute and consistent with x3 in the second attribute. Figure 3 It can be seen that different data have different feature coupling relationships, and there is a certain relationship between the attributes contained in one data and the attributes of other data, such as being the same in one attribute but different in another attribute, which provides a corresponding priori basis for the decoupling of data features.

[0052] The two attributes of this embodiment each include two value conditions, which can make the four pre-stack seismic data in the same set of training data have different characteristic coupling relationships and are mutually correlated in terms of attributes; in actual applications, the number of attributes can also be greater than two, and the value of each attribute can also be greater than two, thereby increasing the diversity of attribute matching coupling.

[0053] 2. Construction of the autoencoder network of decoupling process A

[0054] When training an autoencoder D(E(X,M)), we hope to unlock the potential output of E so that we can use D to decode samples that are not provided to E. D(E) outputs (inputs) multiple data into (from) the data space, where both networks can access the feature coupling relationship.

[0055] Before training, the present invention designs a decoupling process A, which aims to exchange different data attributes by exchanging the corresponding values in the depth attributes. To this end, the present invention redesigns the latent space Z = E (X, M) of the depth attributes extracted by the encoder E. The depth attribute extracted by the basic autoencoder is a one-dimensional vector The vector z = [g1, g2] is the concatenation of two vectors, g1 and g2 represent two different decoupled sub-attributes of the depth attribute z in the latent space. In the case of multiple depth attributes, we use g ij To represent the depth attribute z i Then, we define the Swap(·) operation to exchange different features to achieve the desired feature decoupling effect:

[0056] Swap k(z1,z2) = Swap k ([g 11 ,g 21 ,[g 12 ,g 22 ) = ([g 11 ,g 22 ,[g 12 ,g 21 )

[0057] To more clearly illustrate the help of the Swap(·) function for feature decoupling, we use the data set S = {x1, x2, x3, x4; m1, m2, m3, m4} for further explanation. Here, x1 to x4 are the four data included in this data set, and m1 to m4 are the four feature coupling relationships among this data set. Use the encoder E to encode the data therein to obtain the feature relationship in the latent space z1 = E(x1) = [g 11 ,g 21 , z2 = E(x2) = [g 12 ,g 22 , z3 = E(x3) = [g 13 ,g 23 , z4 = E(x4) = [g 14 ,g 24 . Assume that for the four data in S that have passed through the encoder E, the feature coupling relationships among them are m1: (g 11 = g 13 ), m2: (g 21 = g 24 ), m3: (g 12 = g 14 ), m4: (g 22 = g 23 ). It can be seen that data x3 and x4 are obtained by swapping the second attribute of data x1 and x2 respectively, that is, data x3 and x4 are cross-coupled by the corresponding physical attributes in x1 and x2. If the decoupled sub-attributes g1 and g2 obtained by the encoder E can achieve feature decoupling of this data set, then after inputting the extracted decoupled sub-attributes into the decoder D, the final output should be the reconstruction of data x3 and x4, that is Therefore, the loss function of the decoupling autoencoder used in the present invention is:

[0058]

[0059] Compiling the above various processes, the architecture of the network finally built by the present invention is as shown in Figure 4 Figure 4As shown. Of course, during the actual training of the network, z3 and z4 are not encoded by the encoder E, but are calculated using the Swap(·) function after two encoding processes of z1 = E(x1) and z2 = E(x2).

[0060] In summary of the forward modeling synthetic data construction and network building processes mentioned above, the actual application of the present invention can be finally divided into the following steps, and the flowchart of the overall method is as Figure 5 shown.

[0061] Step1: Data generation and predefined decoupled sub-properties

[0062] First, predefine the decoupled sub-properties to be decoupled according to the actual problem to be solved. Corresponding to the problem to be solved in this article, in the synthetic data verification stage, we define two decoupled sub-properties as elastic parameters and the thickness of the formation of interest such as underground channels or sand bodies; similarly, other physical properties such as rock porosity, rock water and gas saturation, and basic properties of rocks can also be decoupled accordingly; then, according to the seismic records and well logging data of the actual work area, use the forward modeling algorithm based on the Aki-Richards equation and the convolution model to generate the required training data set.

[0063] Step2: Construction and training of the decoupling autoencoder

[0064] Build a decoupling autoencoder network based on the convolutional autoencoder and the decoupling process A: Swap(·), and iteratively update the network parameters based on the gradient descent method. The specific training process of the network is as Figure 6 shown.

[0065] Encoding: The present invention uses multiple sets of prestack data with characteristic coupling relationships generated by the forward modeling method as the network input. For each set of data, the physical properties of each data have different characteristic coupling relationships with other data in the same set. Divide this set of data into two categories, so that the properties of the second category of data are cross-coupled by the properties of the data in the first category: use one category as the input of the encoder, and the second category is used for the construction of the loss function. For the training of a set of data, the present invention inputs x1 and x2 in the data set S = {x1, x2, x3, x4} into the encoder E successively to obtain their corresponding depth attributes z1 = E(x1) = [g 11 , g 21 , z2 = E(x2) = [g 12 , g 22 .

[0066] A: Swap(·) attribute decoupling:

[0067] We pre-define the depth attribute z obtained by the encoder E for the purpose of decoupling, and perform the exchange of the corresponding attribute vectors, that is, the A:Swap(·) calculation process mentioned above, to obtain the feature attribute z3 = [g 11 , g 22 , z4 = [g 12 , g 21 , so as to obtain the decoupled sub-attributes with a new feature coupling relationship.

[0068] Decoding: Input z3 and z4 obtained after exchanging the attribute vectors into the decoder D for decoding successively to obtain the reconstructed data According to the feature coupling relationship in the data group S, we hope that the network can achieve the expected feature decoupling process. Then, there should be z3 = E(x3) = [g 13 , g 23 = [g 11 , g 22 , z4 = E(x4) = [g 14 , g 24 = [g 12 , g 21 , that is That is also the loss function of the decoupling autoencoder:

[0069]

[0070] Among them, D and E represent the parameters of the decoder and encoder of the built network, M is the feature coupling space of the training data, and the number of i depends on the number of different features in the pre-defined depth attribute.

[0071] Finally, use the gradient descent method to update and iterate the network parameters by minimizing the value of the loss function.

[0072] Step3: Application of actual data

[0073] Import the actual work area data into the trained decoupling autoencoder network, extract the required decoupled sub-attributes, and perform clustering analysis on the decoupled sub-attributes to obtain the seismic facies analysis results that meet the feature requirements of geological and geophysical experts.

[0074] The following illustrates the effect of the present invention in combination with specific data:

[0075] 1. Verification of synthetic data

[0076] Here, we verify the validity of synthetic data according to the example mentioned above. First, based on the well logging data of the actual work area, the distribution of elastic parameters and formation thickness in the actual data is extracted, and then synthetic data is generated using the forward algorithm calculated based on the Aki-Richards equation and the convolution model. Here, we generated 40,000 groups of data as the training set. Each group of data contains 4 pre-stack data, with 48 time sampling points, a sampling rate of 1 ms, and each data contains seismic reflection information at 10 different incident angles. Considering the actual situation of the problem, the physical properties selected here are elastic parameters and formation thickness. It is predefined that the data we generate is mainly affected by these two physical factors. Among them, the elastic parameters are water-bearing sandstone and gas-bearing sandstone, and the formation thickness matching the actual distribution is coupled. The elastic parameters of the formation are set as Figure 7 shown.

[0077] In terms of test data, also based on the actual work area data, we synthesized 800 pre-stack data, and the data dimensions are the same as those of the training data. Among them, the data numbered 1-400 are water-bearing sandstone, and the data numbered 401-800 are gas-bearing sandstone, and the model has a total of five different formation thicknesses: 5 m, 10 m, 15 m, 20 m, and 25 m. We randomly selected three pre-stack data with different characteristic coupling relationships for display, as Figure 8 shown. Among them, Figure A is the waveform of the pre-stack data of the water-bearing sandstone with a formation thickness of 5 m, Figure B is the pre-stack waveform data of the water-bearing sandstone with a bottom layer thickness of 20 m, and Figure C is the pre-stack waveform data of the gas-bearing sandstone with a formation thickness of 20 m. It can be seen that the different pre-stack data generated by different characteristic coupling relationships are very different in terms of amplitude size and lateral variation law.

[0078] The actually used decoupled convolutional autoencoder network consists of a decoder and an encoder. The encoder is composed of multiple convolutional layers, pooling layers, and a final fully connected layer. After encoding the input data into a multi-dimensional attribute tensor, the fully connected layer is used to finally obtain a one-dimensional depth attribute. The decoder consists of corresponding fully connected layers, deconvolutional layers, and anti-pooling layers. The specific model parameters of the encoder are as Figure 9 shown, and the model parameters of the decoder are symmetrically corresponding to it.

[0079] Use the network we built to train and test with the synthetic data. The change of the training loss function is as Figure 10 shown, and the depth attributes extracted by the ordinary convolutional autoencoder are used for comparison. For unsupervised seismic facies analysis, usually after using a deep learning network to extract depth features, a clustering method is used to cluster the extracted features to obtain a seismic facies map, and the obtained seismic facies map is actually analyzed in combination with known information such as drilling, well logging, and mud logging.

[0080] Here, we select the Kmeans clustering algorithm to analyze the two depth attributes. Considering that our test model has a total of 2 elastic parameters and 5 formation thicknesses, we perform 5 - clustering on the extracted formation thickness attributes defined by us, 2 - clustering on the elastic parameter attributes, and also perform 2 - clustering and 5 - clustering on the depth attributes extracted from the network without the decoupling process trained with synthetic data for comparison. The final results are as Figure 11 、 Figure 12 shown. By comparing Figure 11 and Figure 12 , it can be seen that when clustering using the depth features extracted by our designed decoupling network, the two physical attributes can be well separated in the attribute domain, and the final clustering results are also similar to theirs; while the clustering results of the depth attributes extracted by the original network show irregularity and instability, without decoupling ability.

[0081] 2. Actual data

[0082] We select an actual work area in the Sichuan Basin for experimental verification of actual seismic data. The total area of this work area is about 1250 square kilometers, the depth of the target formation is about 1000 - 1150 meters, and the root - mean - square amplitude slice of the target formation is as Figure 13 shown. It can be seen from Figure 13 that there are a large number of channels developed in the target formation. There are obvious channels developed at point A and point B in Figure 13 ; and there may be thin - layer channels with a narrow transverse width and a longitudinal thickness lower than the tuning thickness developed at point C, but the channel development situation and channel trend at this point cannot be clearly distinguished from the root - mean - square amplitude map.

[0083] To more clearly analyze the channel development situation and channel trend of the target formation, here we select two attributes: formation elastic parameters and the presence or absence of channels with different thicknesses for decoupling. According to the method of the present invention, first, combined with the actual logging data of this work area, we use the forward modeling method to construct sufficient complex pre - stack seismic data. We have obtained a total of 40000 groups of pre - stack data. The elastic parameters of the data are consistent with the data distribution in the actual logging and drilling data, and channels with different thicknesses are embedded in some of the data. Each data contains the information of the amplitude versus offset at 10 incident angles.

[0084] We still build a decoupling auto - encoder network as shown in Figure 7 , and use the forward data we obtained for training. The change curve of its loss function is as Figure 14As shown, the network converges basically after 1000 rounds of training. After that, the encoder of the trained network is used to extract two decoupled sub-properties of the formation thickness and channel development of the actual data in this work area. Considering that this experiment is mainly to explore the channel development of the target formation, so here we select the decoupled sub-property of channel development for K-means clustering, and the number of clusters for clustering is set to 8. For comparison, we use the depth properties extracted by a common autoencoder also trained with this training data for K-means clustering, and the number of clustering clusters is 8. The comparison diagram of the two results is as Figure 15 shown. It can be seen that the clustering result using the decoupled sub-property of channel development is clearer and has better continuity. The main channels at points A and B are obvious, and the thin-layer channels at point C that are not obvious in the root-mean-square amplitude map can also clearly depict the channel trend, and can separate channels at different depths. Compared with Figure 13 the comparative analysis, it can be observed that compared with the root-mean-square amplitude map and the clustering map of the common autoencoder, the lateral continuity of the channel itself is clearer and more in line with the geological law of channel deposition. After careful analysis Figure 15 on the right, it can be seen that the development of the channel cannot be clearly judged from the clustering map of the common autoencoder, and the main channel is not clear, which is not conducive to further seismic facies analysis.

[0085] Model experiments and actual data tests prove that using the pre-stack seismic facies analysis technology for specific physical property decoupling proposed in the present invention significantly improves the accuracy and reliability of pre-stack seismic facies analysis for specific applications, and there are obvious advantages in network interpretability.

[0086] 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 it should be understood that the protection scope 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 modification, equivalent replacement, improvement, 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. A prestack seismic facies analysis method for decoupling specific physical properties, characterized in that Including: S1. Based on the elastic parameters and formation structure characteristics of the actual work area logging data, use the forward algorithm based on the Aki-Richards equation and the convolution model to generate several training data sets S2. Build a decoupled autoencoder network based on a convolutional autoencoder and a decoupling process A:Swap(·), and use the training data generated in step S1 to train the decoupled autoencoder network. The decoupled autoencoder network specifically includes: an encoder module, an A:Swap(·) attribute decoupling module, and a decoding module. S3. Import the actual work area data into the trained decoupled autoencoder network, extract the required decoupled sub-attributes, and perform clustering analysis on the decoupled sub-attributes to obtain the seismic facies analysis results.

2. The prestack seismic facies analysis method for decoupling specific physical properties according to claim 1, wherein Each training data set in step S1 includes 4 pre-stack seismic data. Each pre-stack seismic data is a combination of two attributes, and each attribute has two value cases. Denote the 4 pre-stack seismic data as x1, x2, x3, and x4 respectively. The first attribute of x1 is the same as the first attribute of x3, and the second attribute of x1 is the same as the second attribute of x4. The first attribute of x2 is the same as the first attribute of x4, and the second attribute of x2 is the same as the second attribute of x3.

3. A prestack seismic facies analysis method for decoupling specific physical properties according to claim 2, characterized in that, The two attributes are: elastic parameters and the thickness of the formation of interest.

4. A prestack seismic facies analysis method for decoupling specific physical properties according to claim 3, characterized in that, The training process of step S2 is specifically as follows: Divide the training data set into two categories. The attributes of the second category of data are cross-coupled from the attributes of the data in the first category. Use the first category of data as the input of the encoder, and the second category of data is used to construct the loss function; The first type of data obtains the attribute of the thickness of the formation of interest after passing through the encoder. The attribute of the thickness of the formation of interest obtained by the encoder passes through the A:Swap(·) calculation process to obtain the decoupled sub-attribute of the new feature coupling relationship. The decoupled sub-attribute of the new feature coupling relationship is input into the decoder to obtain the reconstructed data.

5. A pre-stack seismic facies analysis method for decoupling specific physical properties according to claim 4, characterized in that The expression of the loss function is: Among them, D represents the parameters of the decoder, E represents the parameters of the encoder, M is the feature coupling space of the training data, and the number of i depends on the number of different features in the predefined depth attributes.

6. The prestack seismic facies analysis method for decoupling specific physical properties according to claim 4, wherein Denote the first type of data in the two categories divided by the 4 pre-stack seismic data x1, x2, x3, x4 as x1, x2. Input x1 and x2 into the encoder module successively to obtain their corresponding depth attributes z1 = E(x1) = [g 11 , g 21 , z2 = E(x2) = [g 12 , g 22 ; Use the A:Swap(·) calculation process to swap the attribute vectors of z1 and z2, obtaining the characteristic attributes z3 = [g 11 , g 22 , z4 = [g 12 , g 21 ; The A:Swap(·) calculation process is as follows: Swap k (z1,z2) = Swap k ([g 11 ,g 21 ,[g 12 ,g 22 ) = ([g 11 ,g 22 ,[g 12 ,g 21 ) where, g 11 represents the first sub-property of z1, g 12 represents the second sub-property of z1, g 21 represents the first sub-property of z2, g 22 represents the second sub-property of z2.

Images