Volcano institution identification method based on well-seismic combination and neural network

Abstract

The invention relates to a volcanic institution identification method based on well-seismic combination and a neural network, and the method comprises the steps: 1, obtaining a volcanic institution profile two-dimensional picture data set of post-stack seismic data, and dividing the volcanic institution profile two-dimensional picture data set into training set data and test set data; obtaining a volcanic rock logging data set; step 2, based on the training set data, training a neural network mathematical model of a preset network architecture; step 3, inputting the test set data based on the trained neural network mathematical model, and identifying a volcano institution in a seismic section; 4, combining the identified volcanic institution profile with the volcanic logging data set to establish a fusion data set; and 5, based on the fused data set, establishing a CNN volcano institution classification network.

Description

Technical Field

[0001] The invention relates to the technical field of geological exploration in oil and gas exploitation, and in particular to a volcanic structure identification method based on well-seismic combination and neural network. Background Art

[0002] Existing technologies use the fusion of amplitude and coherence attributes to identify volcanic channels based on fusion attributes, and believe that low-frequency, chaotic, nearly vertical channel-like reflection features are volcanic channels. Then, the volcanic rock body is identified by the morphology of the volcano. The type of volcanic structure is identified by combining imaging logging, conventional logging data and core data. This method can currently only identify basic basic neutral and acidic volcano types, but the effect of more detailed classification and identification is general. It is currently unable to accurately identify the effusive phase, explosive phase and extrusive phase. Summary of the invention

[0003] The present invention provides a volcanic structure identification method based on the combination of well seismic data and neural network. Through the volcanic channels, volcanic structures and the distinction of volcanic types identified by experts, a sample database is established, a neural network model is trained, and an algorithm for intelligently identifying and distinguishing volcanic structures is constructed. The algorithm can identify volcanic types, which mainly include three types and seven categories, namely basic, neutral and acidic. The basic type is divided into single-channel volcanic structures in the extrusive phase, multi-channel volcanic structures in the effusive phase, single-channel volcanic structures in the effusive phase, and single-channel volcanic structures in the eruptive phase; the neutral type is divided into composite single-channel volcanic structures and single-channel volcanic structures in the effusive phase; and the acid type is divided into single-channel volcanic structures in the effusive phase. The algorithm improves the accuracy of identifying and distinguishing volcanic structures, improves efficiency, saves manpower, and has guiding significance for volcanic rock oil and gas exploration.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] In a first aspect, the present application provides a volcanic structure identification method based on well-seismic combination and neural network, comprising:

[0006] Step 1: obtaining a volcanic structure profile two-dimensional image dataset of post-stack seismic data, wherein the volcanic structure profile two-dimensional image dataset is divided into a training set data and a test set data; and obtaining a volcanic rock well logging dataset;

[0007] Step 2: Based on the training set data, a neural network mathematical model of a preset network architecture is trained;

[0008] Step 3: Based on the trained neural network mathematical model, the test set data is input to identify the volcanic structure in the seismic profile;

[0009] Step 4: Combine the identified volcanic structure profile with the volcanic rock logging dataset to create a fused dataset;

[0010] Step 5: Based on the fused data set, a CNN volcano structure classification network is established.

[0011] In one implementation, in step 1, obtaining a two-dimensional image dataset of a volcanic structure profile of post-stack seismic data includes:

[0012] Collect a preset number of 2D images of seismic profiles, one of each set as training data, and one profile depicting the volcano as label data;

[0013] Grayscale the seismic profile of the two-dimensional image used as training data;

[0014] The volcano identified by the expert is used as a sample label by covering the pixel area of ​​the image it occupies through a mask.

[0015] In one implementation, the neural network mathematical model of the preset network architecture is a U-net neural network model; the U-net neural network model processes input training data in the following ways: downsampling, upsampling, feature extraction and fusion.

[0016] In one implementation, in step 4, a characteristic curve of preset parameters is selected from a volcanic rock logging data set as a sample, and the type of volcano encountered in the well is selected as a sample label.

[0017] In one implementation, the preset parameters include: natural gamma (GR), spontaneous potential (SP), deep lateral resistivity (RLLD), acoustic time difference (AC), density (DEN), and compensated neutron (CN) in the logging data.

[0018] In one implementation, the step 4 specifically includes:

[0019] Based on the volcanic rock logging data set and the neural network mathematical model with the preset network architecture, a prediction vector for predicting the volcano type is obtained;

[0020] Based on the test set data and the neural network mathematical model with the preset network architecture, the corresponding prediction vector of the predicted volcano type is obtained;

[0021] Based on the fused average of the prediction vectors, the volcano type is determined.

[0022] The technical solution of the present invention extracts the feature combination according to the pixel feature parameters of the seismic profile image, establishes a neural network mathematical model, and calculates the weight between input and output. Finally, the test seismic profile is imported into the model to effectively identify the volcanic channel. After identifying the volcanic structure, a seismic well logging fusion data set is established, and a CNN network model is established to classify the volcanic structure. This method has high efficiency and accuracy in identifying and classifying seismic profile volcanic structures, saves manpower, and has certain significance in the field of oil and gas exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a processing schematic diagram of a neural network mathematical model of a volcano mechanism in one embodiment of the present invention;

[0024] Figure 2 It is a schematic diagram of processing test set data in establishing a fusion data set according to an embodiment of the present invention;

[0025] Figure 3 It is a schematic diagram of processing a volcanic rock logging data set in establishing a fusion data set in one embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiment of the present invention. Original Based on the described embodiments of the present invention, the present invention is described in detail. All other embodiments obtained by ordinary technicians in the field are within the scope of protection of the present invention.

[0027] In view of the defects and problems of the prior art, the present application provides a volcanic structure identification method based on well-seismic combination and neural network, including:

[0028] Step 1: obtaining a volcanic structure profile two-dimensional image dataset of post-stack seismic data, wherein the volcanic structure profile two-dimensional image dataset is divided into a training set data and a test set data; and obtaining a volcanic rock well logging dataset;

[0029] Step 2: Based on the training set data, a neural network mathematical model of a preset network architecture is trained;

[0030] Step 3: Based on the trained neural network mathematical model, the test set data is input to identify the volcanic structure in the seismic profile;

[0031] Step 4: Combine the identified volcanic structure profile with the volcanic rock logging dataset to create a fused dataset;

[0032] Step 5: Based on the fused data set, a CNN volcano structure classification network is established.

[0033] The above method is described below in a more detailed embodiment in conjunction with the accompanying drawings of the present application.

[0034] In a more detailed embodiment of the present application, a volcanic structure identification method based on the combination of well-seismic data and neural networks can identify 7 types of volcanoes of 3 types.

[0035] The steps of this method include:

[0036] S1, establish a two-dimensional image dataset of post-stack seismic volcanic structure profiles and a volcanic rock logging dataset;

[0037] The specific steps include:

[0038] S1.1. Collect 200 sets of seismic profiles with 256*256 pixels, one clear profile as training data, and one profile depicting a volcano as label data;

[0039] S1.2, grayscale the 100 sets of seismic profiles;

[0040] S1.3. The volcano identified by the expert is used as a sample label by covering the pixel area of ​​the image it occupies with a mask.

[0041] S2, establish a neural network mathematical model of the volcanic structure;

[0042] The specific steps include:

[0043] S2.1. The input of the U-type neural network is a 256*256 two-dimensional grayscale image.

[0044] S2.2. Add a layer of 0 value to the outermost layer of the 256*256 grayscale matrix to change it to a 258*258 grayscale matrix. Set 64 3*3 convolution kernels in the first layer to obtain a 64*256*256 feature matrix. Add another layer of 0 value to the outermost layer to change it to a 64*258*258 matrix. Then, use 64 3*3 convolution kernels in the second layer to obtain a 64*256*256 feature matrix. Finally, use the third layer 2*2 maxpool downsampling to convert it to a 64*128*128 feature matrix.

[0045] S2.3. Add a layer of 0 values ​​to the outermost layer of the 64*128*128 feature matrix to change it to a 64*130*130 matrix. Set 128 3*3 convolution kernels in the fourth layer to obtain a 128*258*258 feature matrix. Add another layer of 0 values ​​to the outermost layer to change it to a 128*258*258 matrix. Then, use 128 3*3 convolution kernels in the fifth layer to obtain a 128*258*258 feature matrix. Finally, use the 2*2 maxpool downsampling in the sixth layer to convert it to a 128*128*1282 feature matrix.

[0046] S2.4. Add a layer of 0 values ​​to the outermost layer of the 128*128*128 feature matrix to turn it into a 128*130*130 matrix. seventh Set 256 3*3 convolution kernels to get a 256*64*64 feature matrix, and then add a layer of 0 value to the outermost layer to become The matrix is ​​256*66*66, and then it is converted into a 256*64*64 feature matrix through 256 3*3 convolution kernels in the eighth layer, and finally it is converted into a 256*32*32 feature matrix through 2*2 maxpool downsampling in the ninth layer.

[0047] S2.5. Add a layer of 0 values ​​to the outermost layer of the 256*32*32 feature matrix to change it to a 256*34*34 matrix. Set 512 3*3 convolution kernels in the tenth layer to obtain a 512*32*32 feature matrix. Add another layer of 0 values ​​to the outermost layer to change it to a 512*34*34 matrix. Then, use 512 3*3 convolution kernels in the eleventh layer to obtain a 512*32*32 feature matrix. Finally, use the twelfth layer of 2*2 deconvolution upsampling to convert it to a 512*64*64 feature matrix.

[0048] S2.6. Connect the 512*64*64 feature matrix and the 256*64*64 feature matrix of the eighth layer to form a fused feature matrix of 768*64*64. Add a layer of 0 values ​​to the outermost layer to form a matrix of 768*66*66. Set 256 3*3 convolution kernels in the thirteenth layer to obtain a feature matrix of 256*64*64. Add another layer of 0 values ​​to the outermost layer to form a matrix of 256*66*66. Then, use 256 3*3 convolution kernels in the fourteenth layer to obtain a feature matrix of 256*64*64. Finally, use the 2*2 deconvolution upsampling in the fifteenth layer to convert it into a feature matrix of 256*128*128.

[0049] S2.6. Connect the 256*128*128 feature matrix and the 128*128*128 feature matrix of the fifth layer to form a fused feature matrix of 384*128*128. Add a layer of 0 values ​​to the outermost layer to form a matrix of 384*130*130. Set 128 3*3 convolution kernels in the sixteenth layer to obtain a feature matrix of 128*128*128. Add another layer of 0 values ​​to the outermost layer to form a matrix of 128*130*130. Then, use 128 3*3 convolution kernels in the seventeenth layer to obtain a feature matrix of 128*128*128. Finally, use the 2*2 deconvolution upsampling in the eighteenth layer to convert it into a feature matrix of 128*256*256.

[0050] S2.7. Connect the 128*256*256 feature matrix and the 64*256*256 feature matrix of the second layer to become a 192*256*256 fusion feature matrix. Add a layer of 0 values ​​to the outermost layer to become a 192*258*258 matrix. Set 64 3*3 convolution kernels in the nineteenth layer to obtain a 64*256*256 feature matrix. Add another layer of 0 values ​​to the outermost layer to become a 64*258*258 matrix. Then, use 64 3*3 convolution kernels in the twentieth layer to obtain a 64*256*256 feature matrix. Finally, use the twenty-first fully connected layer to convert it into a 1*256*256 volcano prediction matrix.

[0051] S2.8, 1*128*128 are the artificially established mask label data of the volcanic channel seismic profile. After the model is built, the training set is brought into the model to fit the neural network weight parameters and establish the network model.

[0052] S3, passing the seismic profile test set data into the established neural network model to identify the volcanic structure in the seismic profile;

[0053] S3.1. Establish a test seismic profile data set;

[0054] S3.2. Convert the test data set into a 256*256 grayscale image by grayscale processing and resizing.

[0055] S3.3. Input the established neural network model and output the identified volcanic structure.

[0056] S4, combining the identified volcanic structure profile with the well logging data set profile to establish a fused data set;

[0057] The specific steps include:

[0058] S4.1. Establish a well logging data set, select GR, SP, LLD, AC, DEN and CN as characteristic curves as samples, and the volcano type encountered by the well as the sample label. The sample label data type is a one-dimensional list data of 1*7, such as (1, 0, 0，0，0，0，0)， The vector represents the type of volcano in 7 categories. The logging data is preprocessed into a two-dimensional 1024*7 Matrix list data.

[0059] S4.2. Establish a preprocessed volcanic grayscale seismic profile data set. The sample label data type is a 1*7 one-dimensional list data such as (0, 0, 1, 0, 0, 0, 0). The vector represents the type of the 7 types of volcanoes. The volcanic grayscale seismic profile data is a 256*256 grayscale image.

[0060] S4.3. Add a layer of 0 values ​​to the outermost layer of the 1024*7 two-dimensional matrix list to turn it into a 1026*9 grayscale matrix. Set 64 3*3 convolution kernels in the first layer to obtain a 64*1024*7 feature matrix. Add another layer of 0 values ​​to the outermost layer to turn it into a 64*1026*9 matrix. Then, use 64 3*3 convolution kernels in the second layer to obtain a 64*1024*7 feature matrix. Finally, use the third layer 1*2 maxpool downsampling to convert it into a 64*512*7 feature matrix. The outermost layer of the 64*512*7 two-dimensional matrix list is added with a layer of 0 values ​​to turn it into a 64*512*9 grayscale matrix. The fourth layer is set with 128 3*3 convolution kernels to obtain a 128*512*7 feature matrix. Then, the outermost layer is added with a layer of 0 values ​​to turn it into a 128*512*9 matrix. Then, the fifth layer is passed through 64 3*3 convolution kernels to obtain a 128*512*7 feature matrix. Finally, it is converted into a 128*256*7 feature matrix through the sixth layer of 1*2 maxpool downsampling. The outermost layer of the 128*256*7 two-dimensional matrix list is added with a layer of 0 value to turn it into a 128*256*9 grayscale matrix. The seventh layer is set with 256 3*3 convolution kernels to obtain a 256*256*7 feature matrix. Then, the outermost layer is added with a layer of 0 value to turn it into a 256*256*9 matrix. Then, 64 3*3 convolution kernels are used in the eighth layer to obtain a 256*256*7 feature matrix. Finally, the ninth layer is downsampled with 1*2 maxpool to convert it into a 256*128*7 feature matrix. The outermost layer of the 256*128*7 feature matrix is ​​added with a layer of 0 values ​​to become a 256*130*9 matrix. The tenth layer is set with 256 3*3 convolution kernels to obtain a 256*128*7 feature matrix. Another layer of 0 values ​​is added to the outermost layer to become a 256*130*9 matrix. Then, the eleventh layer is passed through 256 3*3 convolution kernels to obtain a 256*128*7 feature matrix. Finally, the twelfth layer is converted into a 256*256*7 feature matrix through 1*2 deconvolution upsampling. The outermost layer of the 256*256*7 feature matrix is ​​added with a layer of 0 values ​​to become a 256*258*9 matrix. The thirteenth layer is set with 128 3*3 convolution kernels to obtain a 128*256*7 feature matrix. Then, the outermost layer is added with a layer of 0 values ​​to become a 128*258*9 matrix. Then, through the fourteenth layer, 128 3*3 convolution kernels are used to obtain a 128*256*7 feature matrix. Finally, through the fifteenth layer, 1*2 deconvolution upsampling is performed to convert it into a 128*512*7 feature matrix. The outermost layer of the 128*512*7 feature matrix is ​​added with a layer of 0 values ​​to become a 128*514*9 matrix. The sixteenth layer is set with 64 3*3 convolution kernels to obtain a 64*512*7 feature matrix. Another layer of 0 values ​​is added to the outermost layer to become a 64*514*9 matrix. Then, the seventeenth layer is passed through 64 3*3 convolution kernels to obtain a 64*512*7 feature matrix. Finally, the eighteenth layer is converted into a 64*1024*7 feature matrix through 1*2 deconvolution upsampling.Finally, it is converted into a 1*1*7 volcano prediction matrix through the nineteenth fully connected layer.

[0061] S4.4, add a layer of 0 values ​​to the outermost layer of the 256*256 grayscale matrix to change it to a 258*258 grayscale matrix, set 64 3*3 convolution kernels in the first layer to obtain a 64*256*256 feature matrix, then add a layer of 0 values ​​to the outermost layer to change it to a 64*258*258 matrix, then pass the second layer of 64 3*3 convolution kernels to obtain a 64*256*256 feature matrix, and finally pass the third layer of 2*2 maxpool downsampling to convert it to a 64*128*128 feature matrix. Add a layer of 0 values ​​to the outermost layer of the 64*128*128 feature matrix to change it to a 64*130*130 matrix, and set 128 3*3 convolution kernels in the fourth layer Get a 128*258*258 feature matrix, and then add the outermost layer A layer of 0 values ​​becomes a 128*258*258 matrix, thenThe 128 3*3 convolution kernels in the fifth layer are used to obtain a 128*258*258 feature matrix, which is then converted to a 128*128*1282 feature matrix through 2*2 maxpool downsampling in the sixth layer. The 128*128*128 feature matrix is ​​converted to a 128*130*130 matrix by adding a layer of 0 values ​​to the outermost layer. The 256 3*3 convolution kernels in the seventh layer are used to obtain a 256*64*64 feature matrix, which is then converted to a 256*66*66 matrix by adding another layer of 0 values ​​to the outermost layer. The 256 3*3 convolution kernels in the eighth layer are used to obtain a 256*64*64 feature matrix, which is then converted to a 256*32*32 feature matrix through 2*2 maxpool downsampling in the ninth layer. The outermost layer of the 256*32*32 feature matrix is ​​added with a layer of 0 values ​​to become a 256*34*34 matrix. The tenth layer is set with 512 3*3 convolution kernels to obtain a 512*32*32 feature matrix. Another layer of 0 values ​​is added to the outermost layer to become a 512*34*34 matrix. Then, the eleventh layer is passed through 512 3*3 convolution kernels to obtain a 512*32*32 feature matrix. Finally, the twelfth layer is converted into a 512*64*64 feature matrix through 2*2 deconvolution upsampling. The 512*64*64 feature matrix is ​​converted into a 512*66*66 matrix by adding a layer of 0 values ​​to the outermost layer. The thirteenth layer has 256 3*3 convolution kernels to obtain a 256*64*64 feature matrix. Then, a layer of 0 values ​​is added to the outermost layer to obtain a 256*66*66 matrix. Then, the fourteenth layer has 256 3*3 convolution kernels to obtain a 256*64*64 feature matrix. Finally, the fifteenth layer is converted into a 256*128*128 feature matrix by upsampling with 2*2 deconvolution. The outermost layer of the 256*128*128 feature matrix is ​​added with a layer of 0 values ​​to become a 384*130*130 matrix. The sixteenth layer is set with 128 3*3 convolution kernels to obtain a 128*128*128 feature matrix. Another layer of 0 values ​​is added to the outermost layer to become a 128*130*130 matrix. Then, the seventeenth layer is passed through 128 3*3 convolution kernels to obtain a 128*128*128 feature matrix. Finally, the eighteenth layer is converted into a 128*256*256 feature matrix through 2*2 deconvolution upsampling. The outermost layer of the 128*256*256 feature matrix is ​​added with a layer of 0 values ​​to become a 192*258*258 matrix. The nineteenth layer has 64 3*3 convolution kernels to obtain a 64*256*256 feature matrix. Another layer of 0 values ​​is added to the outermost layer to become a 64*258*258 matrix. Then, the twentieth layer has 64 3*3 convolution kernels to obtain a 64*256*256 feature matrix, and finally, it is converted into a 1*1*7 vector matrix through the twenty-first fully connected layer.

[0062] S4.5. Integrate and fuse the vectors predicted by S4.3 and S4.4 and take the average value. The maximum value is the volcano type predicted by the well logging seismic fusion model.

[0063] S5, build a CNN volcano structure classification network.

[0064] Specifically, define the loss function:

[0065]

[0066] The above formula reflects the residual between the predicted value and the true value. The smaller the residual, the better the model training effect. c represents the true value of the sample pixel (0 or 1, 1 represents that the pixel is an expert-labeled volcano pixel), P c represents the predicted value of the sample pixel (between 0 and 1), c is the pixel number, and M is the total number of pixels. The weight is the value of the convolution kernel. The weights are calculated through partial differential equations to establish a network.

[0067] It can be seen that the above technical solution extracts the feature combination according to the pixel feature parameters of the seismic profile image, establishes a neural network mathematical model, and calculates the weight between input and output. Finally, the test seismic profile is imported into the model to effectively identify the volcanic channel. After identifying the volcanic structure, a seismic well logging fusion data set is established, and a CNN network model is established to classify the volcanic structure. This method is efficient and accurate in identifying and classifying seismic profile volcanic structures, saving manpower.

[0068] In the several embodiments provided by the present invention, it should be understood that the disclosed method can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0069] The above are only preferred embodiments of the present invention and are not intended to limit 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 volcanic structure identification method based on well-seismic combination and neural network, characterized in that: include: Step 1: obtaining a volcanic structure profile two-dimensional image dataset of post-stack seismic data, wherein the volcanic structure profile two-dimensional image dataset is divided into a training set data and a test set data; and obtaining a volcanic rock well logging dataset; Step 2: Based on the training set data, a neural network mathematical model of a preset network architecture is trained; Step 3: Based on the trained neural network mathematical model, the test set data is input to identify the volcanic structure in the seismic profile; Step 4: Combine the identified volcanic structure profile with the volcanic rock logging dataset to create a fused dataset; Step 5: Based on the fused data set, a CNN volcano structure classification network is established.

2. The volcanic structure identification method based on well-seismic combination and neural network according to claim 1 is characterized in that: In step 1, a two-dimensional image dataset of a volcanic structure profile of post-stack seismic data is obtained, including: Collect a preset number of 2D images of seismic profiles, one of each set as training data, and one profile depicting the volcano as label data; Grayscale the seismic profile of the two-dimensional image used as training data; The volcano identified by the expert is used as a sample label by covering the pixel area of ​​the image it occupies through a mask.

3. The volcanic structure identification method based on well-seismic combination and neural network according to claim 1 is characterized in that: The neural network mathematical model of the preset network architecture is a U-net neural network model; the U-net neural network model processes input training data in the following ways: downsampling, upsampling, feature extraction and fusion.

4. The volcanic structure identification method based on well-seismic combination and neural network according to claim 3 is characterized in that: In the step 4, characteristic curves with preset parameters are selected from the volcanic rock logging data set as samples, and the volcano types encountered in the well drilling are used as sample labels.

5. The volcanic structure identification method based on well-seismic combination and neural network according to claim 4 is characterized in that: The preset parameters include: natural gamma (GR), spontaneous potential (SP), deep lateral resistivity (RLLD), acoustic time difference (AC), density (DEN), and compensated neutron (CN) in the logging data.

6. The volcanic structure identification method based on well-seismic combination and neural network according to claim 4 is characterized in that: The step 4 specifically includes: Based on the volcanic rock logging data set and the neural network mathematical model with the preset network architecture, a prediction vector for predicting the volcano type is obtained; Based on the test set data and the neural network mathematical model with the preset network architecture, the corresponding prediction vector of the predicted volcano type is obtained; Based on the fused average of the prediction vectors, the volcano type is determined.

Images