Method and System for Extracting Seismic Data Features Based on Manifold Learning Algorithm

Abstract

The present invention discloses a method and system for extracting seismic data features based on a manifold learning algorithm. By using the Locally Linear Embedding (LLE) algorithm to reduce the dimension of high-dimensional seismic attribute parameters, after dimension reduction, fewer attributes are used to replace the high-dimensional attributes, removing a lot of redundant information while retaining most of the information of the high-dimensional seismic attribute data, achieving lossless dimension reduction to obtain low-dimensional optimal attribute data, and performing feature extraction through a convolutional neural network, ultimately realizing reservoir prediction. The experimental results show that the texture features of the low-dimensional optimal attribute data can well reflect the texture features of the formation, providing strong support for subsequent seismic interpretation, reservoir prediction, etc. Analyzing seismic profiles by combining various attribute data feature extraction techniques is one of the research and development directions in future seismic exploration.

Description

Technical Field

[0001] The present invention belongs to the field of seismic data, and particularly relates to a method and system for extracting seismic data features based on a manifold learning algorithm. Background Art

[0002] In the late 1980s, the pattern recognition technology of manifolds has been mentioned in the IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI). In 2000, three articles were published in the American Science magazine, which studied manifold learning from a cognitive perspective and used "Manifold Learning" to emphasize the integrity of the cognitive process.

[0003] Cognitive scientists believe that under the influence of factors such as time and space, the same things will generate a low-dimensional manifold, and human cognitive ability is based on this. According to these two situations, assuming that the attributes of adjacent objects in the space where the object itself is located are the same, since the distance between the main bodies in the object space is not very close, different feature extraction methods are adopted in the space or feature space to provide two closest samples at the same time. The essence of manifold learning is that when the space containing data is a low-dimensional smooth manifold, its internal geometric structure or internal rules are obtained from these data. This indicates that compared with traditional dimensionality reduction algorithms, manifold learning can better reflect the essence of things, is more conducive to understanding and processing data, and can also well solve problems that could not be solved in machine learning before.

[0004] In recent years, certain achievements have been made in the research of manifold learning. LLE and ISOMAP are typical non-linear dimensionality reduction algorithms. The LLE method of Roweis and Saul can map high-dimensional data points into a global low-dimensional coordinate system while ensuring the mutual relationship of adjacent points. This method can not only well detect the non-linearity of data but also obtain features such as invariance of parallel motion and rotation. Tenenbaum et al. introduced the ISOMAP method into multi-dimensional scaling analysis (MDS) and high-order spaces. The approximate geodesic distance obtained by using the shortest minimum path can reflect the Euclidean distance of the internal manifold structure. Through experiments on faces and gestures, this method found the possible low-dimensional parameter space in the high-dimensional initial space. Doonoho et al. used the ISOMAP method to test these data. Through the analysis of these data, ISOMAP can accurately find the basic parameters of the image and thus obtain good results. Doonoho also gave an HLLE method, which can generalize the LLE method to manifolds and find the parameters with local potential isometric mapping.

[0005] In recent years, based on the LLE theory, Zhang Changshui et al. mapped from a low-dimensional embedding space to a high-dimensional space, and through the three-dimensional reconstruction experiment of multi-pose human faces, the effectiveness of the algorithm was proved. Zhan Dechuan and Zhou Zhihua proposed an input parameter range that can produce better visualization effects for the ISOMAP algorithm, thus reducing the sensitivity to noise. Zhao Lianwei et al. further improved the basic principle of ISOMAP, gave the proof of spatial isometric mapping, distinguished the embedding space dimension, the intrinsic dimension of high-order data, and the manifold dimension, and proved that when there is a circular manifold in the high-order data space, the manifold dimension is smaller than the embedding space dimension, and gave an efficient circulation search method. According to this method, an accurate low-dimensional parameter space can be obtained. In addition, Zhou Zhihua et al. also proposed a method that can show the relationship between the high-dimensional data in the observation space and the reduced low-dimensional data from two aspects: the magnification factor and the extension direction, and compared the performance of several manifold learning algorithms (ISOMAP and LLE) through experiments.

[0006] The digitization of seismic signals, the rapid development of new mathematical methods and computer technologies have provided an objective basis for the analysis of seismic characteristics. The digital recording of seismic data is a method that combines mathematics and computer technology, which can better analyze the rich information contained in seismic data. Computers have powerful computing capabilities, huge storage capacities and visualization technologies, which provide favorable conditions for the development of seismic data processing technologies.

[0007] In the 1980s, with the rapid growth of seismic attributes, some difficulties were encountered in the research of seismic attributes. Most seismic attributes have precise mathematical definitions (such as dominant frequency, complexity of K-L signals, etc.), but their geological meanings are mostly unclear. Due to the unclear geological meanings of seismic attributes, it is difficult for people to predict oil and gas reservoirs, which restricts the development of seismic exploration technology itself. The view of Lindseth (1982) is very representative: "Except for amplitude, these (seismic properties) are not universal and have not been widely used in interpretation. The superficial reason is based on many such facts that they cannot be directly related to geological conditions."

[0008] In the aspect of oil and gas exploration and development in China, the research on seismic attributes lags behind relatively. Based on absorbing and digesting foreign technologies, a set of seismic attribute analysis methods applicable to oil and gas exploration and development have been developed. Geological exploration workers in oil companies and scientific research institutions have carried out a large amount of excavation and utilization of seismic attributes. Zhang Yingbo has carried out microscopic analysis and development on reservoir physical parameters based on Biot's two-phase medium. Liu Qiying, Zhu Guangsheng, Chen Zunde, Du Shitong and others have all conducted in-depth research on seismic attributes and published relevant monographs. At present, the research on seismic attributes mainly focuses on the selection of seismic attributes and the identification of reservoirs. At present, the seismic attribute analysis technology has attracted the attention of many domestic experts, and the research on seismic attributes has been applied to many aspects such as structural interpretation, stratigraphic analysis, lithologic feature description, and reservoir monitoring, achieving good results. However, there are still many imperfections in this technology that need to be further studied by subsequent scholars.

[0009] Manifold learning algorithms have now been used in the processing of seismic data to reduce the dimension of the obtained seismic data attributes. The main manifold methods that have been studied include:

[0010] (1) ISOMAP (Isometric Mapping) method: The basic idea of ISOMAP is to use the local distance approximation method to find the geodesic distance between data points, and perform dimensionality reduction by comparing the geodesic of the original data with the spatial distance of the reduced-dimensional data. The ISOMAP algorithm reduces the dimension by performing dimensionality reduction on the geodesic between point pairs and uses the Multi Dimensional Scalling (MDS) method to obtain the optimal geometric structure, thereby accurately determining the possible data flow parameter space.

[0011] (2) LLE (Locally Linear Embedding) method: LLE is a manifold algorithm based on local linearity. Locally, it can be regarded as linear. It can select any point and express it through an approximate linear fitting. The core idea of the LLE method is to establish a local linear relationship between adjacent data in the original high-dimensional data and try to retain its local linear representation as much as possible, so as to achieve the purpose of dimensionality reduction.

[0012] (3) HLLE (Hessian Locally Linear Embedding) algorithm: HLLE is a very effective non-linear data dimensionality reduction method, which has similar implementation steps to the LLE algorithm, only different in finding the k nearest neighborhood points of a point in the first step. According to the defined distance, find the k nearest neighbors of the data point k, instead of the original LLE algorithm that finds the k nearest neighbors of the data point at the Euclidean distance. HLLE attempts to recover the generating coordinates of the manifold that is locally isometric to an open connected subset in the low-dimensional Euclidean space.

[0013] (4) LE Laplace Eigenmap Algorithm: Laplace Eigenmap is a spectrum-based manifold learning method. Its basic idea is to find the local attribute mapping of data points in an average sense. This method first constructs a weighted neighborhood map using the k-neighbor method or the neighborhood method, then uses the heat kernel to weight it and finally minimize it.

[0014] (5) LPP locality preserving projection algorithm: LPP is a linear estimate of the characteristic function of the Laplace-Beltrami operator. Its purpose is to maintain the corresponding adjacent relationship in the projected space and construct a new coordinate axis so that adjacent data points in the original data space can be projected onto a new coordinate axis. Finding this optimal coordinate axis becomes a measure of the quality of dimensionality reduction.

[0015] (6) MDS multidimensional scaling analysis: Multidimensional scaling analysis technology is a method that can compress data containing multiple variables into an intuitive spatial diagram and can spatially express the potential relationship between the variables. This method can query nominal and continuous parameters without satisfying multiple normal distribution assumptions. Therefore, when using factor analysis, minimum space analysis and other data for retrieval and analysis, the use of multidimensional scaling technology is a good choice. MDS43 analyzes similar data to discover hidden structural information from the data. The measure of similarity is generally expressed in Euclidean distance. The MDS algorithm aims to map data samples to a lower dimensional space while trying to ensure the spacing between samples.

[0016] (7) PCA algorithm: PCA is a major statistical method that can transform multiple index problems into a smaller comprehensive index. The basic idea of ​​principal component analysis is to transform the original variables and component-related random variables into new variables of uncorrelated components through orthogonal transformation, transform the covariance matrix of the original variables into a diagonal matrix from an algebraic perspective, from a geometric system to a new orthogonal system, and distribute the sample points in the maximum orthogonal direction.

[0017] (8) LTSA local tangent space arrangement algorithm: This algorithm realizes the local detail representation of the low-dimensional manifold by approximating the tangent space of each data point, and the local low-dimensional coordinates are solved by inputting the data points to the projection of the local tangent space, and the local tangent space arrangement is used to obtain the overall low-dimensional embedding coordinates.

[0018] (9) DM random walk diffusion mapping algorithm: The main idea is to expand the k nearest neighbor points, that is, all sample points have an impact on the dimensionality reduction result, but the magnitude of the impact mainly depends on the distance between the sample points. Sample points with a short distance have a greater impact on the dimensionality reduction result, while sample points with a long distance have a smaller impact on the dimensionality reduction result.

[0019] Many methods in cognitive science are studied with popular learning algorithms. Assuming that the data is a low-dimensional manifold and uniformly sampled in Euclidean space, manifold learning restores the low-dimensional manifold structure in high-order data, that is, searches for the low-dimensional manifold in a higher initial space, and then through the corresponding relationship of embedding, enables it to simplify or display the data. In recent years, the supervised LLE manifold learning method has also made certain progress and shown good results in pattern recognition. In terms of supervised feature extraction, it is considered that the extended characteristics of dimensionality reduction learning should be fully utilized. In the embedding space, it is necessary to truly reflect the structure of the manifold, and at the same time, separate the sub-manifolds composed of different categories as much as possible, increase the distance between different categories, and compress the data within the class as much as possible. This fully proves that manifold learning can discover the essence of substances from the observed phenomena and reveal its internal laws.

[0020] Generally speaking, seismic attributes and geological objects do not correspond one by one. In other words, there are many factors that affect the nature of earthquakes. In fact, many earthquake properties are determined by comprehensive factors such as geology and oil and gas, excluding the influence of the calculation of seismic attributes and the data itself. Therefore, there is a large degree of uncertainty in using seismic attributes to analyze reservoir characteristics. This is a complex problem for geophysical inversion. Seismic attribute technology methods overcome the uncertainty of seismic attributes, including comprehensively positioning the physical properties of rock masses and comprehensively evaluating multi-seismic attributes. Understanding the uncertainty and non-uniqueness of seismic attributes and their relationship with the quality of seismic data are also issues to be solved by seismic attribute technology.

[0021] As many as 200 kinds of seismic attribute parameters can be extracted from the collected seismic data. Although the increase in seismic attributes is helpful for structural interpretation and stratigraphic analysis, the geological meanings of most attributes are not clear, and it causes the problem of high-dimensional bottleneck of data. How to select the optimal seismic attributes for feature extraction and reservoir identification is the focus of current research on seismic attribute analysis technology.

[0022] Therefore, currently, it is necessary to solve the problem that the existing seismic attribute analysis technology cannot select the optimal seismic attributes. Summary of the Invention

[0023] The purpose of the present invention is to overcome the problem that the existing seismic attribute analysis technology cannot select the optimal seismic attributes, and provides a method and system for extracting seismic data features based on the manifold learning algorithm. The LLE algorithm is used to reduce the dimension of high-dimensional seismic attribute data, and after the obtained low-dimensional optimal attribute data is subjected to feature extraction by a convolutional neural network, reservoir prediction can be realized.

[0024] To achieve the above object, the present invention adopts the following technical solutions:

[0025] A method for extracting seismic data features based on a manifold learning algorithm, comprising the following steps:

[0026] Step 1: Preprocess the seismic data to obtain a high-dimensional data set;

[0027] Step 2: Normalize the high-dimensional data set to obtain high-dimensional seismic attribute data;

[0028] Step 3: Use the LLE algorithm to reduce the dimension of the high-dimensional seismic attribute data and map it into a low-dimensional space to obtain low-dimensional optimal attribute data;

[0029] Step 4: Use a convolutional neural network to extract seismic data features from the low-dimensional optimal attribute data.

[0030] Further, Step 3 is specifically: Obtain the neighbor points of the high-dimensional seismic attribute data sample points, calculate the local linear reconstruction weight matrix of the high-dimensional seismic attribute data sample points from the neighbor points, and use the local linear reconstruction weight matrix to map all high-dimensional seismic attribute data sample points onto the internal global coordinate low-dimensional vectors to output the low-dimensional optimal attribute data.

[0031] Further, the specific method for obtaining the neighbor points of the high-dimensional seismic attribute data sample points is: Use the KNN algorithm to calculate the neighbor points of each high-dimensional seismic attribute data sample point, and define the neighbor points that meet the Euclidean distance relative to the high-dimensional seismic attribute data sample point to be obtained as the neighbor points of the high-dimensional seismic attribute data sample point to be obtained. The Euclidean distance objective function is as follows:

[0032] δ E (x i ,x j )=[(x i -x j ) T (x i -x j )] 1 / 2

[0033] where x i is the high-dimensional seismic attribute data sample point, x j is the neighbor point, δ E (x i ,x j ) is the Euclidean distance objective function of x i and x j , and (x i -x j ) T is the transpose of the difference between the high-dimensional seismic attribute data sample point and the neighbor point.

[0034] Further, the specific method for calculating the local linear reconstruction weight matrix of the high-dimensional seismic attribute data sample points from the neighbor points is as follows: construct an error function and then minimize the error function to solve for the local linear reconstruction weight matrix.

[0035] Further, the error function is as follows:

[0036]

[0037] where minε(W) is the error function, N is the high-dimensionality, x i is the high-dimensional seismic attribute data sample point, x ij (j = 1, 2,..., k) are the neighbor points of x i , k is the number of neighbor points, w ij is the weight between x i and x ij ;

[0038] The process of minimizing the error function follows two constraint conditions:

[0039] The first constraint condition is: if x ij is not a neighbor point of x i , then w ij = 0;

[0040] The second constraint condition is: if x ij is a neighbor point of x i , then

[0041] The local linear reconstruction weight matrix w ij is as follows:

[0042]

[0043] where Q i is a k×k local covariance matrix:

[0044]

[0045]

[0046] where x ij , x im , x ip , x iq are the j, m, p, q neighbor points of x i respectively, and (x i - x ij ) T is the transpose of the difference between the high-dimensional seismic attribute data sample point and the neighbor point.

[0047] Furthermore, the conditions that the mapping needs to satisfy are as follows:

[0048]

[0049] where ε(Y) is the loss function value, Y = (y 1 , y 2 , …, y m ) is the low-dimensional feature vector, m is the number of non-zero eigenvalues, y i is the output vector of x i , x i is the high-dimensional seismic attribute data sample point, y ij (j = 1, 2, …, k) are the k nearest neighbor points of y i , and N is the number of high-dimensional seismic attribute data sample points.

[0050] Furthermore, the preprocessing in step one is specifically as follows: obtaining the seismic attribute profile of the prestack seismic data or the post-stack seismic data and grouping the seismic attribute profile according to seismic traces. The seismic attribute profile includes the instantaneous amplitude attribute profile, the instantaneous frequency attribute profile, the envelope peak instantaneous frequency attribute profile, and the dip stack peak amplitude attribute profile;

[0051] The formula for obtaining the seismic attribute profile of the prestack seismic data or the post-stack seismic data is as follows:

[0052] The calculation formula for the instantaneous amplitude is as follows:

[0053] A(t) = |s(t) + i·H[s(t)]|

[0054] where, is the analytic part of the signal calculated by wavelet transform, A(t) is the modulus of the instantaneous amplitude, i is the imaginary unit, s(t) is a trace of seismic data, S(b, a) is the wavelet transform of the seismic data, is the real part of the Fourier transform of the wavelet function g(t), a is the scale factor, b is the translation factor, and the wavelet transform of s(t) with respect to g(t) is defined as:

[0055]

[0056] In the formula, t, b ∈ R, a > 0; g(t) ∈ L 1 (R, dt) ∩ L 2 (R, dt), is the complex conjugate of g(t);

[0057] The envelope peak instantaneous frequency refers to the envelope peak instantaneous frequency of the constant-phase wavelet. The envelope peak instantaneous frequency of the constant-phase wavelet is the average frequency obtained by weighting the wavelet amplitude spectrum by the Fourier frequency, as shown in the following formula:

[0058]

[0059] where f p (τ) is the envelope peak instantaneous frequency of the constant phase wavelet, f is the frequency, and S(τ, f) is the source wavelet amplitude spectrum;

[0060] The calculation of the instantaneous frequency is performed using the derivative of the instantaneous phase as follows:

[0061]

[0062] where s(t) is the seismic signal, y(t) is the Hilbert transform of the seismic signal s(t), and f(t) is the instantaneous frequency;

[0063] The tilted stack peak amplitude profile is obtained by performing a local linear Radon transform on the seismic profile and then stacking the seismic traces near the reference trace along the selected slope. When the selected stacking slope is close to or equal to the slope of the event axis, the stacking value of the records in the time - distance domain along this line is the largest. The average value of the maximum stacking values is placed at the corresponding position in the time - distance domain to construct a super - trace gather to increase the signal - to - noise ratio. This profile is called the tilted stack peak amplitude profile.

[0064] Further, step three is only performed on the first trace of each group of seismic traces.

[0065] Further, the seismic data in step one is in sgy format.

[0066] A seismic data feature extraction system based on the manifold learning algorithm includes a pre - processing module: used to pre - process seismic data to obtain a high - dimensional data set; a normalization module: used to normalize the high - dimensional data set to obtain high - dimensional seismic attribute data; a dimensionality reduction module: used to reduce the dimension of the high - dimensional seismic attribute data using the LLE algorithm and map it into a low - dimensional space to obtain low - dimensional optimal attribute data; a feature extraction module: used to extract seismic data features from the low - dimensional optimal attribute data using a convolutional neural network.

[0067] Compared with the prior art, the present invention has the following beneficial technical effects:

[0068] Based on seismic data and combined with the latest research in the current interdisciplinary field, the present invention applies the manifold learning algorithm to the processing of seismic data. The local linear embedding algorithm (LLE) is used to reduce the dimension of the attribute data of the obtained high - dimensional seismic data to obtain the low - dimensional optimal attribute data after dimensionality reduction. The low - dimensional optimal attribute data is used to extract feature data through a convolutional neural network to achieve target recognition and search for target reservoirs, and finally realize reservoir prediction. The normalization process unifies the dimensions of different attributes, balances the contributions of each attribute feature, and avoids errors in calculating the distances between calculation samples due to excessive differences in parameter values.

[0069] By reducing the dimensionality of pre-stack / post-stack seismic high-dimensional attribute data, the present invention obtains the optimal k value for dimensionality reduction of seismic attribute data. After dimensionality reduction, fewer attributes are used to replace the high-dimensional attributes, removing a lot of redundant information. The LLE algorithm can not only reduce the dimensionality of the data structure but also help to discover the inherent dimensionality of the data structure. The results of dimensionality reduction of different seismic trace attribute data by the LLE algorithm are different. On the optimized result, LLE can use fewer attributes to replace the original attributes, and at the same time can retain most of the information of the original high-dimensional data set like PCA, achieving lossless dimensionality reduction to a certain extent, and its optimization effect is obviously better.

[0070] The LLE algorithm adopted by the present invention can transform high-dimensional data into an overall low-dimensional coordinate system, enabling it to maintain the original topological structure after dimensionality reduction. The LLE algorithm focuses on maintaining the local linear characteristics of the samples, abandoning the global optimal dimensionality reduction of all samples, and assuming that the sample set satisfies a linear relationship locally in order to reduce the computational complexity of dimensionality reduction.

[0071] The present invention extracts attributes such as instantaneous frequency, instantaneous amplitude, and envelope peak instantaneous frequency from seismic data to form high-dimensional seismic attribute data. Through the comparison of the results after dimensionality reduction, the optimal attributes after final dimensionality reduction are obtained. The optimal attributes are used to construct multi-layer convolutional kernels through a convolutional neural network to extract features from the target data. The experimental results show that the texture features of the low-dimensional optimal attribute data can well reflect the texture features of the formation, providing strong support for subsequent seismic interpretation, reservoir prediction, etc. Analyzing seismic profiles by combining various attribute data feature extraction techniques is one of the research and development directions in future seismic exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings in the specification are used to provide a further understanding of the present invention, and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0073] Figure 1 It is a diagram of the LLE algorithm.

[0074] Figure 2 It is a three-dimensional tensor representation of an image.

[0075] Figure 3 It is a standard schematic diagram of a three-dimensional image of the Swiss Roll dataset.

[0076] Figure 4 It is a result diagram of dimensionality reduction of the Swiss Roll using the LLE algorithm.

[0077] Figure 5 It is a schematic diagram of a common shot gather profile.

[0078] Figure 6 It is a schematic diagram of the instantaneous frequency profile.

[0079] Figure 7 It is a schematic diagram of the instantaneous amplitude profile.

[0080] Figure 8 It is a schematic diagram of the instantaneous frequency profile of the envelope peak.

[0081] Figure 9 It is a dimensionality reduction data graph of the first group of seismic data, where (a) is the dimensionality reduction data graph of the first group of seismic data when k = 20, (b) is the dimensionality reduction data graph of the first group of seismic data when k = 40, (c) is the dimensionality reduction data graph of the first group of seismic data when k = 60, and (d) is the dimensionality reduction data graph of the first group of seismic data when k = 80.

[0082] Figure 10 It is a dimensionality reduction data graph of the second group of seismic data, where (a) is the dimensionality reduction data graph of the second group of seismic data when k = 20, (b) is the dimensionality reduction data graph of the second group of seismic data when k = 40, (c) is the dimensionality reduction data graph of the second group of seismic data when k = 60, and (d) is the dimensionality reduction data graph of the second group of seismic data when k = 80.

[0083] Figure 11 It is the post-stack actual seismic data.

[0084] Figure 12 It is the image feature representation.

[0085] Figure 13 It is the histogram parameter for network training.

[0086] Figure 14 It is the feature extraction image. Detailed implementation manner

[0087] The present invention will be further described in detail below:

[0088] A method for extracting seismic data features based on the manifold learning algorithm includes the following steps:

[0089] Step 1: Preprocess the seismic data to obtain a high-dimensional data set;

[0090] Step 2: Normalize the high-dimensional data set to obtain high-dimensional seismic attribute data;

[0091] Step 3: Use the LLE algorithm to reduce the dimension of the high-dimensional seismic attribute data and map it into a low-dimensional space to obtain low-dimensional optimal attribute data;

[0092] Step 4: Use a convolutional neural network to extract seismic data features from the low-dimensional optimal attribute data.

[0093] Step 3 specifically includes: obtaining the neighbor points of the high-dimensional seismic attribute data sample points, calculating the local linear reconstruction weight matrix of the high-dimensional seismic attribute data sample points from the neighbor points, and using the local linear reconstruction weight matrix to map all high-dimensional seismic attribute data sample points onto the internal global coordinate low-dimensional vectors to output the low-dimensional optimal attribute data.

[0094] Specifically, obtaining the neighbor points of the high-dimensional seismic attribute data sample points means: using the KNN algorithm to calculate the neighbor points of each high-dimensional seismic attribute data sample point, and defining the neighbor points that meet the Euclidean distance relative to the sought high-dimensional seismic attribute data sample point as the neighbor points of the sought high-dimensional seismic attribute data sample point. The Euclidean distance objective function is as follows:

[0095] δ E (x i ,x j )=[(x i -x j ) T (x i -x j )] 1 / 2

[0096] where x i is the high-dimensional seismic attribute data sample point, x j is the neighbor point, δ E (x i ,x j ) is the Euclidean distance objective function between x i and x j , and (x i -x j ) T is the transpose of the difference between the high-dimensional seismic attribute data sample point and the neighbor point.

[0097] Specifically, calculating the local linear reconstruction weight matrix of the high-dimensional seismic attribute data sample points from the neighbor points means: constructing an error function and then minimizing the error function to solve for the local linear reconstruction weight matrix.

[0098] The error function is as follows:

[0099]

[0100] where minε(W) is the error function, N is the high-dimensionality, x i is the high-dimensional seismic attribute data sample point, x ij (j = 1, 2,..., k) are the neighbor points of x i , k is the number of neighbor points, and w ij is the weight between x i and x ij ;

[0101] The process of minimizing the error function follows two constraints:

[0102] The first constraint is that if x ij is not a neighbor of x i , then w ij = 0;

[0103] The second constraint is that if x ij is a neighbor of x i , then

[0104] the local linear reconstruction weight matrix w ij is as follows:

[0105]

[0106] where Q i is a k×k local covariance matrix:

[0107]

[0108]

[0109] where x ij , x im , x ip , x iq are the j, m, p, q neighbors of x i respectively, and (x i - x ij ) T is the transpose of the difference between the high - dimensional seismic attribute data sample point and its neighbor.

[0110] The conditions that the mapping needs to satisfy are as follows:

[0111]

[0112] where ε(Y) is the value of the loss function, Y = (y 1 , y 2 , …, y m ) is the low - dimensional sample set, m is the number of non - zero eigenvalues, y i is the output vector of x i , x i is the high - dimensional seismic attribute data sample point, y ij (j = 1, 2, ..., k) are the k neighbors of y i , and N is the number of data points.

[0113] In Step 1, the preprocessing is specifically as follows: Obtain the seismic attribute profiles of the pre-stack seismic data or the post-stack seismic data, and group the seismic attribute profiles according to seismic traces. The seismic attribute profiles include the instantaneous amplitude attribute profile, the instantaneous frequency attribute profile, the envelope peak instantaneous frequency attribute profile, and the tilted stack peak amplitude attribute profile.

[0114] The formula for obtaining the seismic attribute profiles of the pre-stack seismic data or the post-stack seismic data is as follows:

[0115] The calculation formula for the instantaneous amplitude is as follows:

[0116] A(t) = |s(t) + i·H[s(t)]|

[0117] Where, is the analytic part of the signal calculated by wavelet transform, A(t) is the modulus of the instantaneous amplitude, i is the imaginary unit, s(t) is a trace of seismic data, S(b,a) is the wavelet transform of the seismic data, is the real part of the Fourier transform of the wavelet function g(t), a is the scale factor, b is the translation factor, and the wavelet transform of s(t) with respect to g(t) is defined as:

[0118]

[0119] In the formula, t, b ∈ R, a > 0; g(t) ∈ L 1 (R, dt) ∩ L 2 (R, dt), is the complex conjugate of g(t).

[0120] According to Barnes' definition, the envelope peak instantaneous frequency (EPIF) of the constant phase wavelet is the average frequency weighted by the wavelet amplitude spectrum to the Fourier frequency, that is

[0121]

[0122] Where, f p (τ) is the EPIF, f is the frequency, and S(τ,f) is the source wavelet amplitude spectrum.

[0123] The instantaneous frequency calculation can directly adopt the definition of the instantaneous frequency, that is, the derivative of the instantaneous phase is as follows:

[0124]

[0125] Where, s(t) is the seismic signal, y(t) is the Hilbert transform of the seismic signal s(t), and f(t) is the instantaneous frequency.

[0126] The tilted stack peak amplitude profile is obtained by performing a local linear Radon transform on a seismic profile and then stacking several traces near the reference trace along the selected slope. When the selected stacking slope is close to or equal to the slope of the event, the stacking value along this line in the t-x domain (time-distance domain) is the largest. Placing the average of the maximum stacking values at the corresponding positions (τ j , x m ) in the t-x domain can construct a super gather to increase the signal-to-noise ratio, and this profile is called the tilted stack peak amplitude profile.

[0127] Step 3 is only performed on the first trace of each group of seismic traces.

[0128] The seismic data in Step 1 is in sgy format.

[0129] A seismic data feature extraction system based on the manifold learning algorithm includes a preprocessing module: used to preprocess seismic data to obtain a high-dimensional data set; a normalization module: used to normalize the high-dimensional data set to obtain high-dimensional seismic attribute data; a dimensionality reduction module: used to reduce the dimension of the high-dimensional seismic attribute data using the LLE algorithm and map it into a low-dimensional space to obtain low-dimensional optimal attribute data; a feature extraction module: used to extract seismic data features from the low-dimensional optimal attribute data using a convolutional neural network.

[0130] The present invention will be further described below in conjunction with the drawings and embodiments:

[0131] Generally speaking, there are relatively many seismic parameters that can be extracted from seismic data. Although the more parameters, the better for understanding seismic data, too many seismic data parameters will result in the high-dimensional characteristics of the data. In order to apply seismic attribute analysis well for subsequent prediction and interpretation, it is necessary to reduce the dimension of the parameters of seismic data.

[0132] The present invention applies the manifold learning dimensionality reduction algorithm to the dimensionality reduction of seismic data parameters based on cognitive science technology. Based on the understanding of existing manifold learning algorithms, the LLE algorithm among them is applied to the dimensionality reduction of seismic data parameters. By reducing the three-dimensional parameters of seismic data to two dimensions, the optimal k value for the dimensionality reduction of seismic data parameters is obtained. After dimensionality reduction, fewer attributes are used to replace more attributes, and a lot of redundant information can be removed. Through the comparison of the data before and after dimensionality reduction, it can be concluded that when the LLE algorithm is applied to the dimensionality reduction of seismic attribute parameters, if the k value is appropriately selected, a good dimensionality reduction effect can be achieved.

[0133] LLE algorithm is a new dimensionality reduction technique. This method can transform the data points of high-dimensional input into an overall low-dimensional coordinate system, enabling it to maintain the original topological structure after dimensionality reduction. LLE can learn low-dimensional manifolds of any dimension with local linear structures, ensuring that the proximity relationship between data points and their neighboring points remains unchanged under translations, rotations, and scaling transformations. Moreover, LLE has a globally optimal solution and does not require iteration. Its basic idea is to construct a local linear plane between any sample point and its neighboring points. This sample point can be linearly reconstructed from its neighboring points, and the reconstruction weights are such that the linear reconstruction error between the sample point and its neighborhood is minimized. In the low-dimensional space, the reconstruction relationship between each sample point and its neighborhood is maintained, and it is assumed that the low-dimensional embedding is locally linear. By minimizing the reconstruction error, the embedding result of the high-dimensional vector can be obtained. The minimum reconstruction error in LLE preserves the geometric properties of the local neighborhood of the data, that is, the minimum reconstruction error is invariant under data translations, rotations, and scaling.

[0134] It can be seen from Figure 1 that the LLE algorithm mainly consists of three steps. The first step is to select the k nearest neighbors of the samples, and this process uses a method similar to the KNN algorithm to obtain the proximity. The second step is to determine the linear relationship among the k adjacent samples in the adjacent region for each sample and obtain the linear relationship weighting coefficient W. The third step is to use the weighting factor to reconstruct the sample data in low dimensions. Specifically as follows:

[0135] Step 1: Calculate the Euclidean distances between each seismic data sample point and its k adjacent points x i1 , x i2 , …, x ik . k is a value given according to prior conditions, and the following objective function is constructed using the Euclidean distance:

[0136] δ E (x i , x j ) = [(x i - x j ) T (x i - x j )] 1 / 2 (1)

[0137] where x i is the high-dimensional seismic attribute data sample point, x j is the neighboring point, and δ E (x i , x j ) is the Euclidean distance between x i and x j ;

[0138] Step 2: Minimize the error function shown in Equation (2) to obtain the local reconstruction weight matrix W of the seismic data sampling points:

[0139]

[0140] The above equation should follow the following two constraints:

[0141] Constraint 1: If x ij is not a neighbor of x i at this time, w ij = 0;

[0142] Constraint 2: If x ij is a neighbor of x i at this time

[0143] From Constraint 1, it can be concluded that when considering the weights of sample points, only the weights of the k neighbors of each sample point x i need to be considered. Therefore, Equation (2) is further rewritten as:

[0144]

[0145] where minε(W) is the error function, N is the high-dimensional dimension, x i is the high-dimensional seismic attribute data sample point, x ij (j = 1, 2,..., k) are the neighbors of x i , k is the number of neighbors, w ij is the weight between x i and x ij . The constraint conditions that the above equation needs to satisfy are: Therefore, for each sample point and its neighbors, Equation (3) can be further rewritten:

[0146]

[0147] In the above equation, Q i is a k×k local covariance matrix, and:

[0148]

[0149] where x ij , x im are the jth and mth neighbors of x i respectively, w im is the weight between x i and x im ;

[0150] From this, the local linear reconstruction weight matrix can be solved:

[0151]

[0152] Step 3: In the low-dimensional space, map all seismic data sampling points to calculate the optimal low-dimensional embedding Y = (y 1 , y 2 , …, y m ). The conditions to be satisfied by the mapping are as follows:

[0153]

[0154] where ε(Y) is the loss function value, y i is the output vector of x i , y ij (j = 1, 2, …, k) are the k nearest neighbor points of y i , Y = (y 1 , y 2 , …, y m ) is the low-dimensional sample set, m is the number of non-zero eigenvalues, x i is the high-dimensional seismic attribute data sample point, N is the number of data points, and the standardized low-dimensional data should satisfy two conditions, namely:

[0155]

[0156]

[0157] where y i is the sample element after dimensionality reduction, I is the identity matrix, is the transpose of y i .

[0158] The LLE algorithm can learn local linear low-dimensional manifolds of any dimension, and it has only two undetermined coefficients k and d. At the same time, the optimal weights obtained by minimizing the objective function are symmetric, and the neighbor weights of each point remain unchanged under translation, rotation, and scaling transformations. The LLE algorithm has an analytical global optimal solution and does not require iteration. The calculation of the low-dimensional embedding is reduced to the calculation of the eigenvalues of a sparse matrix, so the computational complexity is relatively small. Applying the LLE algorithm to seismic attribute dimensionality reduction has the following advantages:

[0159] 1. Like ISOMAP, the LLE algorithm uses the number k of adjacent points as the output of the dimension, and the value of k plays an important role in the algorithm. When the value of k is too large, LLE cannot reflect the local characteristics, so that the LLE algorithm is more suitable for PCA; on the contrary, LLE cannot preserve the topological structure of the sampling points in the low-dimensional space;

[0160] 2. If the output order d is too high, it will interfere with the output data and the inherent characteristics of the sampling data cannot be accurately extracted;

[0161] 3. The difference is that the LLE method requires the parameter r which cannot be used by ISOMAP. The selection of r not only affects the convergence of the characteristic analysis of the matrix, but also is very helpful for subsequent data dimensionality reduction.

[0162] The optimized attribute data obtained by LLE dimensionality reduction can achieve reservoir prediction after feature extraction by a convolutional neural network.

[0163] The principle of the convolutional neural network is briefly introduced below. The kernel of the convolutional neural network is determined by the visual nerve mechanism. It mainly uses a series of operations such as convolution and deep pooling to extract image features and is currently mainly applied to the field of image recognition and processing. As a feedforward neural network, its neurons can respond to a part covered by surrounding units.

[0164] In the 1990s, the Super Vision team used a convolutional neural network to classify images in the Large Scale Visual Recognition Challenge (ILSVRC) and achieved remarkable results. The error rate was only 15.3%, while the error rate of the second place was 26.3%, far exceeding the error range, which made image classification attract people's attention. In 2012, Super Vision was the only team using a convolutional neural network to classify images. In 2013, more teams used convolutional neural networks for image classification. In ILSVRC 2014, almost all teams were using convolutional neural networks.

[0165] Currently, convolutional neural network models all adopt the concept of layering, that is, from surface signal features to inner layer data features, and each layer has different features. To control the fitting degree, neurons' outputs in the downstream layer are usually randomly deleted to increase the training time, improving the network's learning performance. In addition, local response normalization also standardizes the activities of neurons, reducing the error rate.

[0166] The convolutional neural network (CNN) can be regarded as an extension of the fully connected neural network. It consists of neurons such as an input layer, an output layer, and hidden layers, and these neuron layers are regularly connected together. Compared with the fully connected neural network, the main feature of the convolutional neural network is more the convolutional operation. Therefore, it has excellent performance in fields such as image retrieval, classification, and target feature recognition. In the forward operation, the convolutional neural network is a process of data superposition. The data representation form of each layer can be expressed as a three-dimensional tensor, as Figure 2 shown. The optimization of the convolutional neural network depends on the convolutional training of the neural network, calculating the error between the loss function's expected output and the actual output, and then using the error backpropagation algorithm to backpropagate from the last layer to update the parameters. The optimized operation parameters continue to run forward and loop until the loss function value is minimized, that is, the final target data reached after training.

[0167] The data input layer of the CNN mainly preprocesses the original image data. The input format of the input layer of the CNN preserves the structure of the picture itself. The main purpose of the convolutional calculation layer of the CNN is to extract features from the input image. Each image can be regarded as a matrix of pixel values. For example, for a 5×5 image, a convolutional kernel with a size of 3×3 (convolution kernel size) is used for convolution. The size of the feature map (convolutional feature) is controlled by three parameters: depth, stride, and zero padding, which need to be determined before convolution. After convolution, the convolutional neural network needs to be activated. Making a non-linear mapping of the output result of the convolutional layer is the role of the ReLU activation layer of the CNN. The activation function is used to introduce non-linear factors because the expressive ability of the linear model is insufficient. The pooling layer is sandwiched between consecutive convolutional layers and is used to compress the amount of data and parameters, reducing overfitting. In short, if the input is an image, the most important role of the pooling layer is to compress the image. The downsampling layer is also called the pooling layer. The fully connected layer plays the role of a "classifier" in the entire convolutional neural network. In other words, it integrates the features together (highly purifies the features) for easy handing over to the final classifier or regression. The fully connected layer will convert the two-dimensional feature map output by convolution into a one-dimensional vector.

[0168] Embodiment

[0169] First, use the standard Swiss Roll in manifold learning to verify the dimensionality reduction of the manifold learning algorithm. Figure 3 is the standard graph of the Swiss Roll. Figure 4 is the result graph of dimensionality reduction using the LLE algorithm. It can be seen from the graph that the graph after dimensionality reduction by the LLE algorithm has an obvious linear structure, and the topological structure of the data is also relatively well preserved.

[0170] Next, apply the LLE algorithm to the dimensionality reduction of attributes such as the instantaneous frequency, instantaneous amplitude, and envelope peak instantaneous frequency of seismic data. Figure 5 is the common shot gather profile. Figure 6 is the instantaneous frequency profile. Figure 7 is the instantaneous amplitude profile. Figure 8 is the envelope peak instantaneous frequency profile. First, preprocess the seismic data. Divide the 595 seismic traces in each seismic profile (both are matrices of 595×1250) into 7 groups, with 85 seismic traces in each group. Apply the LLE algorithm for dimensionality reduction to each group of seismic traces (only perform dimensionality reduction on the first trace of each group, and the remaining seismic traces are approximately considered the same). Apply the proposed method to different groups of seismic traces and their attribute data.

[0171] Figure 9 is the dimensionality reduction result graph of the first group of seismic data. From the dimensionality reduction result of the first group, as Figure 9As shown in (a), when k = 20, the dimensionality reduction result of the LLE algorithm is relatively ideal. The data topological structure is not damaged, the data also exhibits good linear characteristics, and its local characteristics are also well reflected. As the value of k increases, as Figure 9 shown in (b) and (d), when k = 40 and k = 80, the dimensionality reduction results of the LLE algorithm are not very ideal. The data topological structure is damaged, the linear characteristics of the data are not well reflected, and the local characteristics of the data after dimensionality reduction are not well reflected either. As Figure 9 shown in (c), when k = 60, the result after dimensionality reduction by the LEE algorithm is better than that when k = 40 and k = 80. It can be seen that the data after dimensionality reduction has good linear characteristics, but the data points are too dense and the local characteristics of the data after dimensionality reduction are not shown. Therefore, it is still not as ideal as the result of dimensionality reduction when k = 20. Therefore, in the processing results of the first group, the selected value of k is 20.

[0172] Figure 10 It is the dimensionality reduction result diagram of the second group of seismic data. It can be seen from the results of LLE dimensionality reduction that as the seismic trace changes, the dimensionality reduction results also change significantly. As Figure 10 shown in (a), when k = 20, the dimensionality reduction result is not very ideal. The number of data points after dimensionality reduction is significantly reduced, the topological structure is also slightly damaged, the linear characteristics are not very obvious, and the local characteristics of the data are not shown. As Figure 10 shown in (b), when k = 40, the dimensionality reduction result of the data is very ideal. The linear and local characteristics of the data are well shown, and the topological structure of the data is obvious and not damaged. As the value of k increases, as Figure 10 shown in (c) and (d), when k increases to 60 and 80, the topological structure of the data is damaged, the data points become dense, and the linear and local characteristics of the data are not well shown. The results after dimensionality reduction are not ideal. Therefore, the selected value of k in the processing results of the second group is 40.

[0173] The same dimensionality reduction processing is performed on the third to seventh groups, which will not be elaborated here.

[0174] After dimensionality reduction of the model data and actual seismic data by the LLE algorithm and comparative analysis of the dimensionality reduction results of the embodiments, it can be seen that, first, the results of dimensionality reduction of different high-dimensional seismic attribute data by the LLE algorithm are different. After comparing the dimensionality reduction data results of seven groups, it is considered that the optimal seismic trace group after dimensionality reduction is the third group. Second, in terms of the optimization results, LLE can use fewer attributes to replace the original attributes, and at the same time can retain most of the information of the original high-dimensional data set like PCA, achieving lossless dimensionality reduction to a certain extent, and its optimization effect is obviously better. The data obtained by dimensionality reduction using the LLE algorithm is seismic trace attribute data. Initially, three seismic attributes were set, namely: instantaneous frequency, instantaneous amplitude, and envelope peak instantaneous frequency. After comparing the results of dimensionality reduction, it is concluded that finally two attributes are obtained after dimensionality reduction: instantaneous frequency and envelope peak instantaneous frequency. Finally, the dimension d after dimensionality reduction by the LLE algorithm is 2, and the value of the dimensionality reduction parameter k is selected as 20. The LLE algorithm can not only reduce the dimension of the data structure but also help to discover the inherent dimension of the data set structure. After dimensionality reduction of the actual seismic data, a fixed dimension of 2 is obtained.

[0175] Apply the proposed method to the actual post-stack seismic data. After dimensionality reduction by LLE, then use a convolutional neural network to perform feature analysis on the dimensionality-reduced seismic attributes. The original image of the actual post-stack seismic data is as Figure 11 shown. Figure 12 is the image feature representation obtained by the network model after training multiple training images. Image feature extraction is performed through these features. Figure 13 is the histogram parameter for network training. In this network, the training effect is relatively concentrated, and the training accuracy reaches more than 75%. Through the training of this network, feature extraction is performed on the low-dimensional optimal attribute data after dimensionality reduction, and the feature extraction image as shown in Figure 14 is obtained. The texture features of the feature extraction image have been extracted by the network model. Compared with the original image, the texture features of the feature image are more obvious and there is less useless information, which is significantly helpful for identifying the features of seismic images. The experimental results show that the texture features of the attribute data can well reflect the texture features of the formation, providing strong support for subsequent seismic interpretation, reservoir prediction, etc. Analyzing seismic profiles by combining various attribute data feature extraction techniques is one of the research and development directions in future seismic exploration.

[0176] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent replacements to the specific implementation manners of the invention, but these changes, modifications or equivalent replacements are all within the protection scope of the pending claims of the invention.

Claims

1. A method for extracting seismic data features based on a manifold learning algorithm, characterized in that, it includes the following steps: Step 1: Preprocess the seismic data to obtain a high-dimensional data set; Step 2: Normalize the high-dimensional data set to obtain high-dimensional seismic attribute data; Step 3: Use the LLE algorithm to reduce the dimension of the high-dimensional seismic attribute data and map it into a low-dimensional space to obtain low-dimensional optimal attribute data; specifically: find the neighbor points of the high-dimensional seismic attribute data sample points, calculate the local linear reconstruction weight matrix of the high-dimensional seismic attribute data sample points from the neighbor points, and use the local linear reconstruction weight matrix to map all high-dimensional seismic attribute data sample points onto the internal global coordinate low-dimensional vectors to output low-dimensional optimal attribute data; the specific method for finding the neighbor points of the high-dimensional seismic attribute data sample points is: use the KNN algorithm to calculate the neighbor points of each high-dimensional seismic attribute data sample point, and define the neighbor points that meet the Euclidean distance relative to the high-dimensional seismic attribute data sample point to be found as the neighbor points of the high-dimensional seismic attribute data sample point to be found. The Euclidean distance objective function is as follows: Among them is a sample point of high-dimensional seismic attribute data is a neighboring point is and is the Euclidean distance objective function of is the transpose of the difference between the high-dimensional seismic attribute data sample point and the neighboring point Step 4: Use a convolutional neural network to extract seismic data features from the low-dimensional optimal attribute data.

2. The method for extracting seismic data features based on a manifold learning algorithm according to claim 1, characterized in that, the specific method for calculating the local linear reconstruction weight matrix of the high-dimensional seismic attribute data sample points from the neighbor points is: construct an error function and then minimize the error function to solve for the local linear reconstruction weight matrix.

3. The method for extracting seismic data features based on a manifold learning algorithm according to claim 2, characterized in that, the error function is as follows: Among them is the error function N is the high-dimensional dimension is the high-dimensional seismic attribute data sample point , is the neighbor point of k is the number of neighbor points is the weight between and The process of minimizing the error function follows two constraint conditions: The first constraint is that if is not a neighboring point of ; The second constraint is: If is 's neighbor point, then ; The local linear reconstruction weight matrix is as follows: wherein is a k × k local covariance matrix: Among them , , , are respectively of j , m , p , q nearest neighbor points, is the transpose of the difference between the high-dimensional seismic attribute data sample point and the nearest neighbor points.

4. The method for extracting seismic data features based on a manifold learning algorithm according to claim 1, characterized in that, the conditions to be satisfied by the mapping are as follows: where is the loss function value, is the low-dimensional feature vector, m is the number of non-zero eigenvalues, is 's output vector, is the high-dimensional seismic attribute data sample point, , is 's k nearest neighbor points, and N is the number of high-dimensional seismic attribute data sample points.

5. The method for extracting seismic data features based on a manifold learning algorithm according to claim 1, characterized in that, the specific preprocessing in Step 1 is: find the seismic attribute profiles of the pre-stack seismic data or the post-stack seismic data and group the seismic attribute profiles according to seismic traces. The seismic attribute profiles include instantaneous amplitude attribute profiles, instantaneous frequency attribute profiles, envelope peak instantaneous frequency attribute profiles, and dip stack peak amplitude attribute profiles; The formula for finding the seismic attribute profiles of the pre-stack seismic data or the post-stack seismic data is as follows: The calculation formula for the instantaneous amplitude is as follows: Among them, is the analytic part of the signal calculated by wavelet transform, is the modulus of the instantaneous amplitude, is the imaginary unit, is a seismic data, is the wavelet transform of the seismic data, is the wavelet function is the real part of the Fourier transform of is the scale factor, is the translation factor, Regarding the wavelet transform is defined as: In the formula, , ; , is the complex conjugate of The envelope peak instantaneous frequency refers to the envelope peak instantaneous frequency of a constant-phase wavelet. The envelope peak instantaneous frequency of a constant-phase wavelet is the average frequency obtained by weighting the wavelet amplitude spectrum by the Fourier frequency, as shown in the following formula: Among them, is the envelope peak instantaneous frequency of the constant phase wavelet, is the frequency, is the amplitude spectrum of the source wavelet; The calculation of the instantaneous frequency uses the derivative of the instantaneous phase as follows: Among them, is the seismic signal, is the seismic signal of the Hilbert transform, is the instantaneous frequency; The inclined stack peak amplitude profile is obtained by performing a local linear Radon transform on a seismic profile and stacking the seismic traces near the reference trace along the selected slope. When the selected stacking slope is close to or equal to the slope of the event, the stacking value along this line in the time-distance domain is the largest. The average value of the maximum stacking values is placed at the corresponding position in the time-distance domain to construct a super trace gather to increase the signal-to-noise ratio. This profile is called the inclined stack peak amplitude profile.

6. The method for extracting seismic data features based on the manifold learning algorithm according to claim 5, characterized in that, Step three is only performed on the first trace of each group of seismic traces.

7. The method for extracting seismic data features based on the manifold learning algorithm according to claim 1, characterized in that, The seismic data in step one is in sgy format.

8. A seismic data feature extraction system based on the manifold learning algorithm, characterized in that, It includes a preprocessing module: used to preprocess seismic data to obtain a high-dimensional data set; A normalization module: used to normalize the high-dimensional data set to obtain high-dimensional seismic attribute data; A dimensionality reduction module: used to reduce the dimension of the high-dimensional seismic attribute data using the LLE algorithm and map it into a low-dimensional space to obtain low-dimensional optimal attribute data; specifically: find the neighbor points of the high-dimensional seismic attribute data sample points, calculate the local linear reconstruction weight matrix of the high-dimensional seismic attribute data sample points from the neighbor points, and use the local linear reconstruction weight matrix to map all high-dimensional seismic attribute data sample points to the internal global coordinate low-dimensional vectors to output low-dimensional optimal attribute data; The specific method for finding the neighbor points of the high-dimensional seismic attribute data sample points is: use the KNN algorithm to calculate the neighbor points of each high-dimensional seismic attribute data sample point, and define the neighbor points that meet the Euclidean distance relative to the desired high-dimensional seismic attribute data sample point as the neighbor points of the desired high-dimensional seismic attribute data sample point. The Euclidean distance objective function is as follows: Among them is a sample point of high-dimensional seismic attribute data is a neighboring point is and is the Euclidean distance objective function is the transpose of the difference between the high-dimensional seismic attribute data sample point and the neighboring point A feature extraction module: used to extract seismic data features from the low-dimensional optimal attribute data using a convolutional neural network.

Images