Reservoir fracture index prediction method based on deep residual shrinkage network

Abstract

The invention provides a reservoir fracture index prediction method based on a depth residual shrinkage network. The prediction method comprises the following steps: S1, preferably selecting a fracture sensitive conventional logging curve according to a shear wave anisotropy value; s2, a crack index is calculated and used for representing the development degree of a crack beside a well; s3, extracting various fracture sensitive seismic attributes of the post-stack seismic data body; s4, a recursive feature elimination method is used for optimizing fracture sensitive seismic attributes; s5, performing adjustment and training of the deep residual shrinkage network model by using a data set constructed by seismic attributes and fracture indexes; and S6, using the trained deep residual shrinkage network model to predict the fracture index of the target stratum. And the trained depth residual shrinkage network model is used to obtain the fracture index of the target layer, so that a more comprehensive and accurate reservoir fracture distribution range prediction result can be efficiently obtained.

Description

Technical Field

[0001] The invention relates to the field of reservoir geophysical exploration, and in particular to a reservoir fracture index prediction method based on a deep residual shrinkage network. Background Art

[0002] Fractures are important reservoirs and migration channels for oil and gas. They significantly improve reservoir permeability while controlling the distribution of oil and gas enrichment areas. Accurately evaluating the degree of fracture development is a key task for oil reservoirs to achieve high and stable production. However, fractured reservoirs have strong heterogeneity, making quantitative and refined identification and characterization of reservoir fractures a major challenge for oil and gas exploration. The main methods for identifying reservoir fractures include imaging logging, conventional logging, and three-dimensional seismic. Imaging logging is characterized by high resolution and can intuitively describe the distribution of fractures around the wellbore, but it is not suitable for promotion to the entire area due to its high cost. Conventional logging is often used in fracture identification work because of its relatively cheap acquisition fee, but its disadvantage is that the detection range is limited. Although the three-dimensional seismic method has low vertical resolution and cannot identify small-scale fractures and microfractures, it has a wide coverage range and a large vertical exploration depth. When the reservoir has strong anisotropy, if only part of the geophysical data is considered, the reliability of the reservoir fracture prediction results is low. Therefore, it is necessary to explore a new method that integrates well and seismic data to more comprehensively and accurately predict the distribution range of reservoir fractures.

[0003] Machine Learning (ML) with a certain degree of automation is the most widely used intelligent algorithm in the field of reservoir prediction. In recent years, deep learning (DL), as a subclass of machine learning, has achieved great success in dealing with problems such as speech recognition, image classification, and target detection, showing its powerful ability to deal with pattern recognition problems. Deep learning algorithms can mine deep data features from massive data by building deep network models, providing new ideas and methods for solving the problem of reservoir prediction in the field of oil exploration. When applying deep learning algorithms to reservoir fracture prediction, there is no need to establish a logging interpretation model in advance. The network can automatically find the implicit relationship between fractures of different development levels and seismic attribute data, thereby efficiently obtaining more comprehensive and accurate prediction results of the distribution range of reservoir fractures. Summary of the invention

[0004] In view of the above problems, the present invention is proposed to provide a reservoir fracture index prediction method based on a deep residual shrinkage network, which overcomes the above problems or at least partially solves the above problems.

[0005] According to one aspect of the present invention, a reservoir fracture index prediction method based on a deep residual shrinkage network is provided, the prediction method comprising:

[0006] Step S1: Optimizing fracture-sensitive conventional logging curves according to shear wave anisotropy values;

[0007] Step S2: Calculate the fracture index to characterize the degree of fracture development near the well;

[0008] Step S3: extracting multiple fracture-sensitive seismic attributes from the post-stack seismic data volume;

[0009] Step S4: using a recursive feature elimination method to optimize fracture-sensitive seismic attributes;

[0010] Step S5: using the data set constructed by seismic attributes and fracture index to perform training of the deep residual shrinkage network model;

[0011] Step S6: Use the trained deep residual shrinkage network model to predict the crack index of the target layer.

[0012] Optionally, the step S1: optimizing the conventional logging curve of fracture sensitivity according to the shear wave anisotropy value specifically includes:

[0013] Shear wave anisotropy reflects the degree of change in the velocity difference between fast and slow shear waves caused by cracks;

[0014] The shear wave anisotropy is obtained based on the time difference between fast and slow shear waves:

[0015]

[0016] Where, ANTA is the shear wave anisotropy (%); Δt s and Δt f are the time differences between fast and slow shear waves, μs / ft;

[0017] The strength of shear wave anisotropy is usually proportional to the development of fractures, and its magnitude reflects the fracture density and cracking degree of the rock formation near the well.

[0018] Calculate the Pearson correlation coefficient between conventional logging data and shear wave anisotropy:

[0019]

[0020] Among them, cov(X,Y) is the covariance of variables X and Y, σ x σ y is the standard deviation of variables X and Y;

[0021] The Pearson correlation system is used to count and analyze the correlation between the conventional logging response curve and the shear wave anisotropy, and the sensitivity analysis of various conventional logging parameters and fractures is indirectly realized.

[0022] Optionally, the use of Pearson correlation system counting to analyze the correlation between the conventional logging response curve and the shear wave anisotropy specifically includes:

[0023] When the absolute value of the Pearson correlation coefficient is greater than 0.7, the well logging parameter is considered to be highly sensitive; when it is between 0.2 and 0.7, the well logging response curve is considered to have weak sensitivity; when it is less than 0.2, the two are considered to have no sensitivity;

[0024] After filtering out conventional logging data with correlation coefficients lower than 0.2, the fracture-sensitive conventional logging response curves with high correlation and relative independence were selected.

[0025] Optionally, the step S2: calculating the fracture index to characterize the degree of development of the fractures near the wellbore specifically includes:

[0026] The optimal fracture sensitive logging parameters are normalized, and the formula is as follows:

[0027]

[0028] Where W is the measured reading value of the selected conventional logging curve, W min is the minimum value of the conventional logging curve in the target layer, W max is the maximum value of the conventional logging curve in the target layer, and WI is the coefficient of the conventional logging curve indicating the degree of fracture development;

[0029] According to the sensitivity of each logging curve to the fracture, multiple logging curves are combined by arithmetic weighted synthesis to obtain the fracture index:

[0030]

[0031] Where FIA ​​is the calculated fracture index, n is the number of optimal fracture sensitive logging parameters, ai is the weighting coefficient of the i-th fracture sensitive logging parameter, WI i is the coefficient of the i-th fracture sensitive logging parameter indicating the degree of fracture development.

[0032] Optionally, the step S3: extracting multiple fracture-sensitive seismic attributes from the post-stack seismic data volume specifically includes:

[0033] Extract fracture-sensitive seismic volume attributes using post-stack 3D seismic data volumes;

[0034] According to the layer data, slice along the layer in the volume attributes to extract the corresponding layer attributes.

[0035] Optionally, the fracture-sensitive seismic volume attributes include: dip, coherence, curvature, similarity, enhanced coherence, and variance.

[0036] Optionally, the step S4: using a recursive feature elimination method to optimize fracture-sensitive seismic attributes specifically includes:

[0037] The seismic attribute optimization problem is converted into a feature selection problem in machine learning. The recursive feature elimination algorithm is used to evaluate the influence of N different input seismic attributes on the fracture index, and then S combinations of fracture-sensitive seismic attributes are selected after eliminating redundant seismic attributes.

[0038] Suppose the earthquake attribute set after standardization is: X i ={x1,x2,…,x n} N , the label set composed of crack index is: Y = {y1,y2,…,y n}, where n represents the length of the data set and N represents the dimension, i.e. the number of seismic attributes;

[0039] Since N is large and dimensionality reduction is required, a recursive feature elimination algorithm based on a random forest regression model is used to select a set of important seismic attribute combinations X s , whose number S is less than N;

[0040] When training the model, the X and Y data sets are divided into training sets and cross-validation sets, that is, X = {X T ,X V} and Y={Y T ,Y V}, where subscripts T and V represent the training set and validation set, respectively;

[0041] By X T and Y T Iteratively train the regression model, using X V and Y V Test the learner and calculate the predicted value of the crack index The error between the true value Y and the predicted value is calculated to determine the importance of each feature. After each round of training, the earthquake attribute X with the least importance is eliminated. f , and then based on the remaining earthquake attributes X i-1 =X i -X f Construct a new feature set for the next round of training, and optimize the model by recursively deleting earthquake attributes and building a new model based on the remaining earthquake attributes until the number of earthquake attributes that need to be retained is reached. The corresponding earthquake attribute combination X s This is the optimal attribute combination.

[0042] Optionally, the step S5: using the data set constructed by seismic attributes and fracture indexes to perform training on the deep residual shrinkage network model specifically includes:

[0043] The collected seismic attribute data usually contain noise or redundant information, and the deep residual shrinkage network DRSN is selected for reservoir fracture identification;

[0044] To train the model, the seismic traces near the wells in the 3D seismic attribute volume are extracted as the training set, and the fracture index of the corresponding well points is used as the label set;

[0045] The number of channels in the input layer of the training model should be set to the number of seismic attributes in the preferred fracture-sensitive seismic attribute combination;

[0046] The structure of the simplified deep residual shrinkage network has more network layers than the deep residual network.

[0047] Optionally, the structure of the simplified deep residual shrinkage network includes the following network layers compared to the deep residual network:

[0048] The deep residual shrinkage network includes a residual module and a sub-network;

[0049] There are 4 hidden layers between the first convolutional layer and the last fully connected layer. Every two hidden layers form a residual contraction group. There are 2 residual contraction groups in total. Each residual contraction group includes a residual block and a contraction subnetwork.

[0050] Two convolutional layers are used in the residual block, each of which contains 8 convolutional kernels of size 3×3;

[0051] An output feature is obtained through a nonlinear activation function, which converts important seismic attributes into values ​​with larger absolute values ​​and seismic attributes corresponding to redundant information into values ​​with smaller absolute values;

[0052] The residual shrinkage sub-network is used to learn the boundaries between the values ​​of the convolutional layer outputs, preparing for the subsequent soft threshold processing;

[0053] The subnetwork includes two branches, one of which includes a normalization layer BN, a ReLU activation function, a fully connected layer, and a sigmoid function in sequence to obtain a set of positive thresholds, and the obtained thresholds are used to perform soft threshold processing on the seismic attributes of each channel to achieve the whole process from automatic threshold setting to sample denoising; the other branch is connected to the global average pooling layer GAP to obtain the global average value of the feature;

[0054] The output end of the network includes a global average pooling layer and a fully connected layer, and the crack index prediction result is obtained through the fully connected layer.

[0055] Optionally, the step S6: using the trained deep residual shrinkage network model to predict the target layer crack index specifically includes:

[0056] After the training of the deep residual shrinkage network model is completed, the target layer optimal seismic attribute combination data obtained by slicing the three-dimensional seismic volume attributes along the layer is used as the input of the network model;

[0057] Obtain the fracture index prediction results for the entire target layer.

[0058] The present invention provides a reservoir fracture index prediction method based on a deep residual shrinkage network, the prediction method comprising: step S1: optimizing fracture-sensitive conventional logging curves according to shear wave anisotropy values; step S2: calculating fracture indexes to characterize the degree of fracture development near the well; step S3: extracting multiple fracture-sensitive seismic attributes from the post-stack seismic data volume; step S4: optimizing fracture-sensitive seismic attributes using a recursive feature elimination method; step S5: training a deep residual shrinkage network model using a data set constructed using seismic attributes and fracture indexes; step S6: predicting the fracture index of a target layer using the trained deep residual shrinkage network model. The fracture index of the target layer is obtained using the trained deep residual shrinkage network model, and a more comprehensive and accurate prediction result of the reservoir fracture distribution range is efficiently obtained.

[0059] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0061] Figure 1 A flow chart of a reservoir fracture index prediction method based on a deep residual shrinkage network provided in an embodiment of the present invention;

[0062] Figure 2 The four seismic attributes included in the optimal seismic attribute preferred combination of the embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram of the structure of a deep residual shrinkage network model according to an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of the prediction results of the target layer crack index in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0066] The terms "comprises" and "having" and any variations thereof in the description embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.

[0067] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0068] like Figure 1 As shown, Figure 1 The flowchart of the reservoir comprehensive fracture index prediction method based on the deep residual shrinkage network according to the embodiment of the present invention comprises the following steps:

[0069] S1. Optimize the fracture-sensitive conventional logging curve according to the shear wave anisotropy value.

[0070] S1 specifically includes:

[0071] As an important indicator for qualitatively identifying cracks, shear wave anisotropy can reflect the degree of change in the velocity difference between fast and slow shear waves caused by cracks. Its magnitude can be calculated from the time difference between fast and slow shear waves:

[0072]

[0073] Where ANTA is the shear wave anisotropy (%); Δt s and Δt f are the time differences between fast and slow shear waves (μs / ft).

[0074] The strength of shear wave anisotropy is usually proportional to the degree of fracture development, and its size reflects the fracture density and cracking degree of the rock formation near the well. Calculate the Pearson correlation coefficient between conventional logging data and shear wave anisotropy:

[0075]

[0076] Where cov(X,Y) is the covariance of variables X and Y, σ x σ yis the standard deviation of variables X and Y. The correlation between conventional logging response curve and shear wave anisotropy is analyzed by Pearson correlation system counting, and the sensitivity analysis of each conventional logging parameter and fracture is indirectly realized: when the absolute value of Pearson correlation coefficient is greater than 0.7, the sensitivity of logging parameter is considered to be very high; when it is between 0.2 and 0.7, the logging response curve is considered to have weak sensitivity; when it is less than 0.2, the two are considered to have no sensitivity. After filtering out conventional logging data with correlation coefficients less than 0.2, the conventional logging response curve with high correlation and relative independence of fracture sensitivity is selected.

[0077] S2. Calculate the fracture index to characterize the degree of fracture development near the well.

[0078] S2 specifically includes:

[0079] First, the optimal fracture sensitive logging parameters are normalized, and the formula is as follows:

[0080]

[0081] Where W is the measured reading value of the selected conventional logging curve, W min is the minimum value of the conventional logging curve in the target layer, W max is the maximum value of the conventional logging curve in the target layer, and WI is the coefficient of the conventional logging curve indicating the degree of fracture development.

[0082] In order to avoid misjudgment caused by using a single logging parameter to characterize reservoir fractures, multiple logging curves are combined through arithmetic weighted synthesis according to the sensitivity of each logging curve to fractures to obtain the fracture index:

[0083]

[0084] Where FIA ​​is the calculated fracture index, n is the number of optimal fracture sensitive logging parameters, ai is the weighting coefficient of the i-th fracture sensitive logging parameter, WI i is the coefficient of the i-th fracture sensitive logging parameter indicating the degree of fracture development. In this embodiment, the fracture index is calculated using three logging curves: acoustic time difference, micropotential resistivity and microgradient resistivity. The following table lists the calculated fracture index of the target layer in the work area.

[0085] hashtag Crack Index W1 0.1381 W2 0.1520 W3 0.3234 W4 0.2314 W5 0.3271 W6 0.1913 W7 0.1472 W8 0.2875

[0086] S3. Extract various fracture-sensitive seismic attributes from the post-stack seismic data volume.

[0087] S3 specifically includes:

[0088] In order to ensure that the seismic attribute information can better characterize the fracture information, the post-stack 3D seismic data volume is used to extract the fracture-sensitive seismic volume attributes, which mainly include dip, coherence, curvature, similarity, enhanced coherence, and variance. Then, according to its stratigraphic data, the corresponding layer attributes are extracted by slicing along the layer in the volume attributes.

[0089] S4. Optimize fracture-sensitive seismic attributes using a recursive feature elimination method.

[0090] S4 specifically includes:

[0091] The seismic attribute optimization problem is converted into a feature selection problem in machine learning. The recursive feature elimination algorithm is used to evaluate the impact of N different seismic attributes on the crack index, and then S combinations of fracture-sensitive seismic attributes are selected after eliminating redundant seismic attributes. The seismic attribute set after standardization is assumed to be: X i ={x1,x2,…,x n} N , the label set composed of crack index is: Y = {y1,y2,…,y n}, where n represents the length of the dataset and N represents the dimension, i.e. the number of seismic attributes. Since N can be very large and dimensionality reduction is required, a recursive feature elimination algorithm based on a random forest regression model is used to select a set of important seismic attribute combinations X s , whose number S is less than N.

[0092] When training the model, the X and Y data sets are divided into training sets and cross-validation sets, that is, X = {X T ,X V} and Y={Y T ,Y V}, where subscripts T and V represent the training set and validation set respectively. T and Y T Iteratively train the regression model, using X V and Y V Test the learner and calculate the predicted value of the crack index The error between the true value Y and the predicted value is calculated to determine the importance of each feature. After each round of training, the earthquake attribute X with the least importance is eliminated. f , and then based on the remaining earthquake attributes X i-1 =X i -X f Construct a new feature set for the next round of training, and optimize the model by recursively deleting earthquake attributes and building a new model based on the remaining earthquake attributes until the number of earthquake attributes that need to be retained is reached. The corresponding earthquake attribute combination X s This is the optimal attribute combination.

[0093] The seismic attribute optimization method based on the recursive feature elimination method of the random forest regression model is used to optimize the attributes. The 18 seismic attributes are optimized to obtain 4 fracture development sensitive seismic attributes, namely coherence, variance, maximum curvature and maximum positive curvature, as shown in Figure 2. Figure 2 shown.

[0094] S5. Use the dataset constructed by seismic attributes and fracture index to train the deep residual shrinkage network model.

[0095] Step S5 specifically includes:

[0096] Since the datasets constructed by seismic attributes and fracture index are relatively small, and the structure of the deep residual shrinkage network is more complex than that of the deep residual network, it is necessary to appropriately simplify the number of network layers. The deep residual shrinkage network used is mainly composed of residual modules and subnetworks. The network contains 4 hidden layers between the first convolutional layer and the last fully connected layer. Every two hidden layers constitute a residual shrinkage group. There are 2 residual shrinkage groups in total, and each residual shrinkage group includes a residual block and a shrinkage subnetwork. Two convolutional layers are used in the residual block, each of which contains 8 convolution kernels of size 3×3. Then, an output feature is obtained through a nonlinear activation function. This process can convert important seismic attributes into values ​​with larger absolute values ​​and convert seismic attributes corresponding to redundant information into values ​​with smaller absolute values. The boundary between the values ​​output by the convolutional layer is learned through the residual shrinkage subnetwork to prepare for the subsequent soft threshold processing. The residual shrinkage subnetwork is mainly composed of two branches, one of which is composed of a normalization layer (BN), a ReLU activation function, a fully connected layer, and a sigmoid function in sequence to obtain a set of positive thresholds. The obtained thresholds are used to perform soft threshold processing on the seismic attributes of each channel, realizing the whole process from automatic threshold setting to sample denoising; the other branch is connected to the global average pooling layer (GAP) to obtain the global average of the features. The output end of the network contains a global average pooling layer and a fully connected layer, and the fracture index prediction result is obtained through the fully connected layer. The network structure used is as follows Figure 3 shown.

[0097] S6. Use the trained deep residual shrinkage network model to predict the crack index of the target layer.

[0098] In this embodiment, step S6 specifically includes:

[0099] After the deep residual network model training is completed, the target layer preferred seismic attribute combination data obtained along the layer in the three-dimensional seismic volume attribute slice is used as the input of the network model to obtain the fracture index prediction result on the entire target layer, and realize the prediction of the reservoir fracture distribution range. The prediction results obtained in this embodiment are as follows: Figure 4 shown.

[0100] In one embodiment of the present invention, Figure 2 The four seismic attributes included in the optimal seismic attribute preferred combination of the embodiment of the present invention are: Figure 2 (a) is relevant, Figure 2 (b) is the variance, Figure 2 (c) is the maximum curvature, Figure 2 (d) is the maximum positive curvature. Figure 3 The structure of the deep residual shrinkage network model used in the embodiment of the present invention, the deep residual shrinkage network used is mainly composed of a residual module and a subnetwork, the network includes 4 hidden layers between the first convolutional layer and the last fully connected layer, and every two hidden layers constitute a residual shrinkage group, there are 2 residual shrinkage groups in total, and each residual shrinkage group includes a residual block and a shrinkage subnetwork. Two convolutional layers are used in the residual block, each of which contains 8 convolution kernels of size 3×3, and is accompanied by a nonlinear activation function. The residual shrinkage subnetwork is mainly composed of two branches, one of which is composed of a normalization layer (BN), a ReLU activation function, a fully connected layer, and a sigmoid function in sequence; the other branch is connected to a global average pooling layer (GAP) to obtain the global average of the features. The output end of the network includes a global average pooling layer and a fully connected layer, and the prediction result is obtained through the fully connected layer. Figure 4 This is the target layer fracture index prediction result of the embodiment of the present invention.

[0101] The present invention first optimizes the conventional logging curve of fracture sensitivity according to the shear wave anisotropy value, and then uses the optimized conventional logging curve of fracture sensitivity to calculate the fracture index for characterizing the degree of fracture development near the wellbore; then, multiple fracture-sensitive seismic attributes are extracted from the post-stack seismic body, and the fracture-sensitive seismic attributes are optimized by using a recursive feature elimination method; then, the data set established by the optimized seismic attribute combination and the fracture index is used for training in a deep residual shrinkage network model; finally, the trained deep residual shrinkage network model is used to predict the fracture index of the target layer.

[0102] Beneficial effects: Based on the deep residual shrinkage network, a method for predicting reservoir fracture index is provided. By optimizing and calculating the fracture index of multiple conventional logging curves, optimizing multiple seismic attributes to obtain a combination of fracture-sensitive seismic attributes, and using the trained deep residual shrinkage network model to obtain the fracture index of the target layer.

[0103] When applying deep learning algorithms to reservoir fracture prediction, there is no need to establish a well logging interpretation model in advance. The network can automatically find the implicit relationship between fractures of different development degrees and seismic attribute data, and thus efficiently obtain more comprehensive and accurate prediction results of the reservoir fracture distribution range.

[0104] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A reservoir fracture index prediction method based on a deep residual shrinkage network, characterized in that: The prediction method comprises: Step S1: Optimizing fracture-sensitive conventional logging curves according to shear wave anisotropy values; Step S2: Calculate the fracture index to characterize the degree of fracture development near the well; Step S3: extracting multiple fracture-sensitive seismic attributes from the post-stack seismic data volume; Step S4: using a recursive feature elimination method to optimize fracture-sensitive seismic attributes; Step S5: using the data set constructed by seismic attributes and fracture index to perform training of the deep residual shrinkage network model; Step S6: Use the trained deep residual shrinkage network model to predict the crack index of the target layer.

2. The reservoir fracture index prediction method based on the deep residual shrinkage network according to claim 1 is characterized in that: The step S1: optimizing the conventional logging curve of fracture sensitivity according to the shear wave anisotropy value specifically includes: Shear wave anisotropy reflects the degree of change in the velocity difference between fast and slow shear waves caused by cracks; The shear wave anisotropy is obtained based on the time difference between fast and slow shear waves: Where, ANTA is the shear wave anisotropy (%); Δt s and Δt f are the time differences between fast and slow shear waves, μs / ft; The strength of shear wave anisotropy is usually proportional to the development of fractures, and its magnitude reflects the fracture density and cracking degree of the rock formation near the well. Calculate the Pearson correlation coefficient between conventional logging data and shear wave anisotropy: Among them, cov(X,Y) is the covariance of variables X and Y, σ x σ y is the standard deviation of variables X and Y; The Pearson correlation system is used to count and analyze the correlation between the conventional logging response curve and the shear wave anisotropy, and the sensitivity analysis of various conventional logging parameters and fractures is indirectly realized.

3. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 1 is characterized in that: The correlation between the conventional logging response curve and the shear wave anisotropy analyzed by using the Pearson correlation system count specifically includes: When the absolute value of the Pearson correlation coefficient is greater than 0.7, the well logging parameter is considered to be highly sensitive; when it is between 0.2 and 0.7, the well logging response curve is considered to have weak sensitivity; when it is less than 0.2, the two are considered to have no sensitivity; After filtering out conventional logging data with correlation coefficients lower than 0.2, the fracture-sensitive conventional logging response curves with high correlation and relative independence were selected.

4. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 1 is characterized in that: The step S2: calculating the fracture index to characterize the degree of development of the fractures near the well specifically includes: The optimal fracture sensitive logging parameters are normalized, and the formula is as follows: Where W is the measured reading value of the selected conventional logging curve, W min is the minimum value of the conventional logging curve in the target layer, W max is the maximum value of the conventional logging curve in the target layer, and WI is the coefficient of the conventional logging curve indicating the degree of fracture development; In order to avoid misjudgment caused by using a single logging parameter to characterize reservoir fractures, multiple logging curves are combined through arithmetic weighted synthesis according to the sensitivity of each logging curve to fractures to obtain the fracture index: Where FIA ​​is the calculated fracture index, n is the number of optimal fracture sensitive logging parameters, ai is the weighting coefficient of the i-th fracture sensitive logging parameter, WI i is the coefficient of the i-th fracture sensitive logging parameter indicating the degree of fracture development.

5. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 1 is characterized in that: The step S3: extracting multiple fracture-sensitive seismic attributes from the post-stack seismic data volume specifically includes: Extract fracture-sensitive seismic volume attributes using post-stack 3D seismic data volumes; According to the layer data, slice along the layer in the volume attributes to extract the corresponding layer attributes.

6. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 5, characterized in that: The fracture sensitive seismic body attributes include: dip, coherence, curvature, similarity, enhanced coherence, and variance.

7. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 1, characterized in that: The step S4: using the recursive feature elimination method to optimize fracture sensitive seismic attributes specifically includes: The seismic attribute optimization problem is converted into a feature selection problem in machine learning. The recursive feature elimination algorithm is used to evaluate the influence of N different input seismic attributes on the fracture index, and then S combinations of fracture-sensitive seismic attributes are selected after eliminating redundant seismic attributes. Suppose the earthquake attribute set after standardization is: X i ={x1,x2,…,x n } N , the label set composed of crack index is: Y = {y1,y2,…,y n }, where n represents the length of the data set and N represents the dimension, i.e. the number of seismic attributes; Since N is large and dimensionality reduction is required, a recursive feature elimination algorithm based on a random forest regression model is used to select a set of important seismic attribute combinations X s , whose number S is less than N; When training the model, the X and Y data sets are divided into training sets and cross-validation sets, that is, X = {X T ,X V } and Y={Y T ,Y V }, where subscripts T and V represent the training set and validation set, respectively; By X T and Y T Iteratively train the regression model, using X V and Y V Test the learner and calculate the predicted value of the crack index The error between the true value Y and the predicted value is calculated to determine the importance of each feature. After each round of training, the earthquake attribute X with the least importance is eliminated. f , and then based on the remaining earthquake attributes X i-1 =X i -X f Construct a new feature set for the next round of training, and optimize the model by recursively deleting earthquake attributes and building a new model based on the remaining earthquake attributes until the number of earthquake attributes that need to be retained is reached. The corresponding earthquake attribute combination X s This is the optimal attribute combination.

8. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 1, characterized in that: The step S5: using the data set constructed by seismic attributes and crack index to perform training on the deep residual shrinkage network model specifically includes: The collected seismic attribute data usually contain noise or redundant information, and the deep residual shrinkage network DRSN is selected for reservoir fracture identification; To train the model, the seismic traces near the wells in the 3D seismic attribute volume are extracted as the training set, and the fracture index of the corresponding well points is used as the label set; The number of channels in the input layer of the training model should be set to the number of seismic attributes in the preferred fracture-sensitive seismic attribute combination; The structure of the simplified deep residual shrinkage network has more network layers than the deep residual network.

9. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 8, characterized in that: The structure of the simplified deep residual shrinkage network has a larger number of network layers than the deep residual network, specifically including: The deep residual shrinkage network includes a residual module and a sub-network; There are 4 hidden layers between the first convolutional layer and the last fully connected layer. Every two hidden layers form a residual contraction group. There are 2 residual contraction groups in total. Each residual contraction group includes a residual block and a contraction subnetwork. Two convolutional layers are used in the residual block, each of which contains 8 convolutional kernels of size 3×3; An output feature is obtained through a nonlinear activation function, which converts important seismic attributes into values ​​with larger absolute values ​​and seismic attributes corresponding to redundant information into values ​​with smaller absolute values; The residual shrinkage sub-network is used to learn the boundaries between the values ​​of the convolutional layer outputs, preparing for the subsequent soft threshold processing; The subnetwork includes two branches, one of which includes a normalization layer BN, a ReLU activation function, a fully connected layer, and a sigmoid function in sequence to obtain a set of positive thresholds, and the obtained thresholds are used to perform soft threshold processing on the seismic attributes of each channel to achieve the whole process from automatic threshold setting to sample denoising; the other branch is connected to the global average pooling layer GAP to obtain the global average value of the feature; The output end of the network includes a global average pooling layer and a fully connected layer, and the crack index prediction result is obtained through the fully connected layer.

10. The reservoir fracture index prediction method based on deep residual shrinkage network according to claim 1, characterized in that: The step S6: using the trained deep residual shrinkage network model to predict the target layer crack index specifically includes: After the training of the deep residual shrinkage network model is completed, the target layer optimal seismic attribute combination data obtained by slicing the three-dimensional seismic volume attributes along the layer is used as the input of the network model; Obtain the fracture index prediction results for the entire target layer.

Images