Modeling method for processing multi-source geological data and constructing three-dimensional geological model based on pre-trained U-Net3D network

Abstract

The invention discloses a modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network, and the method comprises the steps: carrying out the sampling of drilling data and geophysical profile data to a uniform grid size through a data up-sampling method on the basis of the comprehensive processing of the drilling data and geophysical profile data in a modeling region; assigning a known grid to the to-be-modeled region; and then, carrying out down-sampling feature extraction and up-sampling feature completion on the modeling area containing part of known data by utilizing the adjusted U-Net3D network, and finally, constructing a three-dimensional geologic model by part of known multi-source data. According to the method, the organization of the multi-source geological data can be effectively realized, and the three-dimensional geological model is constructed by part of known multi-source data; the method is easy to implement and remarkable in effect, and the application requirement is met.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional geological modeling, and in particular relates to a modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network. Background Art

[0002] Understanding the direction of underground strata and rock masses, the distribution of ore bodies, and the extension of structures will help geologists and other construction personnel to effectively design ground construction plans and determine the direction of underground construction. The establishment of an accurate three-dimensional geological model can allow geologists to more intuitively understand the distribution of underground strata, ore bodies, and structures.

[0003] The existing methods for constructing three-dimensional geological models mainly include explicit modeling and implicit modeling. Explicit modeling includes: automatic modeling based on generalized triangular prism elements and free curve and surface modeling technology modeling; the existing implicit modeling methods mainly include radial basis function (RBF), Hermite RBF (HRBF), moving least squares (MLS), anisotropic RBF (ARBF), generalized RBF (GRBF) and other implicit modeling methods. Among them, explicit modeling is to directly construct a three-dimensional geological model based on borehole or profile data through manual interaction; implicit modeling is to extract the boundary information in the borehole or profile data based on implicit functions, solve the structural surface of the stratum and rock mass, and construct a three-dimensional geological model. However, there are different problems when the existing modeling methods are used to construct three-dimensional geological models. First, the explicit modeling method requires a lot of manual operation and a lot of experience, which makes it very time-consuming and labor-intensive to process large amounts of data. Second, when faced with multi-source heterogeneous deep exploration data in underground engineering areas, the implicit three-dimensional modeling method faces a series of problems, such as difficulty in collaborative interpretation of structural boundaries, time-consuming and labor-intensive geometric modeling, difficulty in model updating, and lack of uncertainty analysis and accuracy assessment. Summary of the invention

[0004] The purpose of the present invention is to solve the problems raised in the background technology, and propose a method for processing multi-source data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network, so as to realize the organization and comprehensive analysis of drilling data and geophysical profile data, and finally realize rapid analysis and automatic modeling of multi-source data.

[0005] In order to achieve the purpose of the present invention, the present invention discloses a modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network, comprising the following steps:

[0006] Step 1: extracting drilling data and geophysical profile data; upsampling multi-source data by calculating the minimum cell size;

[0007] Step 2, assigning known multi-source data to corresponding areas of the modeling area grid;

[0008] Step 3: Build U-Net3D deep learning network;

[0009] Step 4: Use the U-Net3D pre-trained deep learning network to invert the modeling area containing known data to construct a three-dimensional geological model using multi-source data.

[0010] Furthermore, in step 1, due to the differences in storage format and representation range between borehole data and geophysical data, it is necessary to upsample the multi-source data. The specific steps are as follows:

[0011] Step 1-1, determining the data types of drilling data and geophysical data;

[0012] Step 1-2, determine the representation range of different data, and determine the minimum grid size according to the resolution of different data;

[0013] Steps 1-3: Determine the uniform depth and perform data upsampling.

[0014] Furthermore, in step 2, the known multi-source data are assigned to the corresponding areas of the modeling area grid; based on the calculated unit grid size, the modeling area is divided into N grids of uniform size, and the known multi-source data are assigned to the corresponding grids; the unknown area grid only contains coordinate information (X, Y, Z), and does not contain attribute information and lithology information.

[0015] Furthermore, in step 3, the U-Net3D network includes 4 downsamplings, 4 upsamplings and one 3D horizontal convolution.

[0016] Furthermore, in step 4, the U-Net3D pre-trained deep learning network is used to invert the modeling area containing known data to realize the construction of a three-dimensional geological model using multi-source data.

[0017] Furthermore, in step 1-1, the header information of the drilling data includes lithology name, X, Y, depth from, and depth to; the header information of geophysical data varies depending on the storage format; the geophysical data formats include SEG-Y, LAS, NetCDF, GRD, HDF5, XYZ, and Matlab.mat.

[0018] Furthermore, in step 1-2, the coverage of the data is determined. For geophysical data, its coverage is determined and recorded as [d profile min ,d profile max ]; for drilling data, determine its coverage, denoted as [d drill min ,d drill max]; Select a target grid size that will be used to evenly divide the geophysical and drillhole data. The target grid size is based on the smallest cell size in the geophysical and drillhole datasets.

[0019] Furthermore, in steps 1-3, the depth range [d min ,d max ] need to cover the full range of geophysical data and borehole data, i.e.

[0020] d min =min(d profile min ,d drill min )

[0021] d max =maxx(d profile max , d drill max )

[0022] The entire depth range [d min ,d max ][d min ,d max ] Divide the target grid size Δd and calculate the grid size S grids

[0023]

[0024] Data upsampling, redistribute all data according to the calculated grid size, and upsample the original data of a known length into multiple grids of the same size.

[0025] Furthermore, step 3 specifically includes the following steps:

[0026] Step 3-1, the length, width and height of the area to be modeled are X_Length, Y_Length and Z_Length; during the training process, the modeling area containing some known grids is input into the U-Net3D network and undergoes four downsampling operations. The specific downsampling parameters are set as follows:

[0027] The first downsampling, the parameter settings of the first downsampling include a 4*4 three-dimensional convolution layer, where the stride is 2, the padding is 1, and leakyReLu is used as the activation function; after the first downsampling, the length, width and height of the area to be modeled become 1 / 2 of the initial modeling area, and the number of channels becomes 32;

[0028] The second downsampling, the parameter settings of the second downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function; after the second downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 64;

[0029] The third downsampling, the parameter settings of the third downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function; after the third downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 128;

[0030] The fourth downsampling, the parameter settings of the fourth downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function; after the fourth downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 256;

[0031] Step 3-2: Further feature extraction operation is performed on the modeling area after four downsampling. The horizontal feature extraction operation includes a 4*4 three-dimensional convolution layer, where the stride is 2, the padding is 1, and leakyReLu is used as the activation function; after horizontal convolution, the length, width and height of the modeling area are reduced to 1 / 4 of the initial modeling area, and the number of channels is increased to 256;

[0032] Step 3-3: After four downsampling and horizontal convolution operations, the eigenvalues ​​of the extracted three-dimensional model need to be upsampled. The upsampling is also performed in four steps. The specific upsampling parameters are set as follows:

[0033] The first upsampling, the parameter settings of the first upsampling include a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function; after the first upsampling, the length, width and height of the area to be modeled are increased to 1 / 2 of the initial modeling area, and the number of channels is reduced to 128; in addition, the data after the first upsampling and the third downsampling are jump-connected to ensure the accuracy of the model;

[0034] The second upsampling, the parameter settings of the second upsampling include a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function; after the second upsampling, the length, width and height of the area to be modeled still remain 1 / 2 of the initial modeling area, and the number of channels is reduced to 64; the data after the second upsampling and the second downsampling are jump-connected to ensure the accuracy of the model;

[0035] The third upsampling, the parameter settings of the third upsampling include a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function; after the third upsampling, the length, width and height of the area to be modeled still remain 1 / 2 of the initial modeling area, and the number of channels is reduced to 32; the third upsampling is skipped with the data after the first downsampling to ensure the accuracy of the model;

[0036] The fourth upsampling, the parameter setting of the fourth upsampling includes a 4*4 three-dimensional convolution layer, in which the stride is 2, the padding is 1, and ReLu is used as the activation function; after the fourth upsampling, the length, width and height of the area to be modeled are consistent with those of the initial modeling area, and the prediction result of each cell can be output.

[0037] Further, in step 4, based on the pre-trained model of the U-Net3D inverted three-dimensional geological model, the pre-trained model is loaded to invert the area to be modeled, including the following steps:

[0038] Step 4-1, manually generate a large number of three-dimensional geological models of random occurrences, random tectonic events and random lithology for training;

[0039] Step 4-2, divide the generated 3D geological models of different events into training set, test set and validation set according to the classification method of 8:1:1, and input the training set into the U-Net3D deep learning network mentioned in step 3 for training; use precision, recall rate and F1 score to evaluate the modeling effect of the test set and validation set;

[0040] Step 4-3, save the pre-trained model with precision, recall and F1 score greater than 90% for the test set and validation set;

[0041] Step 4-4: Load the pre-trained model to perform inversion modeling on the area to be modeled that contains known multi-source data.

[0042] Compared with the prior art, the significant progress of the present invention is that it can effectively organize multi-source geological data and construct a three-dimensional geological model, which is reflected in two aspects: first, it can use the data upsampling method to unify the borehole data and geophysical profile data into the same grid size and assign them to the modeling area; second, it can use the adjusted U-Net3D network to perform down-sampling feature extraction and up-sampling feature completion on the modeling area containing some known data, and finally realize the construction of a three-dimensional geological model from some known multi-source data. The present invention is simple to implement, has significant effects, and meets the requirements of the application.

[0043] In order to more clearly illustrate the functional characteristics and structural parameters of the present invention, further description is given below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 It is a schematic diagram of the U-Net3D deep learning network structure used for multi-source data inversion;

[0046] Figure 2 This is a flow chart for constructing a three-dimensional geological model based on multi-source data based on U-Net3D. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0048] The present invention uses a data upsampling method to sample the borehole data and geophysical profile data to a uniform grid size based on the comprehensive processing of the borehole data and geophysical profile data in the modeling area, and assigns the known grid to the area to be modeled. Afterwards, the adjusted U-Net3D network is used to perform down-sampling feature extraction and up-sampling feature completion on the modeling area containing part of the known data, and finally realizes the construction of a three-dimensional geological model from part of the known multi-source data. The present invention is simple to implement, has significant effects, and meets the application requirements.

[0049] like Figure 2 As shown, a modeling method for constructing a three-dimensional geological model based on a pre-trained U-Net3D network using multi-source data can effectively extract and process drilling data and geophysical profile data, and realize the construction of a three-dimensional geological model based on multi-source data, which specifically includes the following steps:

[0050] Step 1: Extract drilling data and geophysical profile data; upsample multi-source data by calculating the minimum cell size.

[0051] Since the storage format and representation range of drilling data and geophysical data are different, it is necessary to upsample the multi-source data. The specific steps are as follows:

[0052] Step 1-1. Determine the data type of drilling data and geophysical data.

[0053] The header information of the drilling data includes: lithology name, X, Y, depth from, depth to.

[0054] Geophysical data has different header information depending on the storage format. Common geophysical data formats include: SEG-Y, LAS, NetCDF, GRD, HDF5, XYZ, Matlab.mat, etc.

[0055] Step 1-2: Determine the representation range of different data and determine the minimum grid size based on the resolution of different data.

[0056] 1. Determine the coverage of the data

[0057] For geophysical data, determine its coverage, denoted as [d profile min ,d profile max ].

[0058] For drilling data, determine its coverage, denoted as [d drill min ,d drill max ].

[0059] 2. Select a target grid size Δd that will be used to evenly divide the geophysical and drillhole data. This can be based on the smallest cell size of the two datasets.

[0060] Step 1-3: Determine the uniform depth and perform data upsampling

[0061] 1. Determine a uniform depth range

[0062] Unified depth range [d min ,d max ] need to cover the full range of geophysical data and borehole data, namely:

[0063] d min =min(d profile min , d drill min )

[0064] d max =max(d profile max ,d drill max )

[0065] 2. The entire depth range [d min ,d max ][d min ,d max ] Divide the target grid size Δd and calculate the grid size S grids :

[0066]

[0067] 3. Data upsampling:

[0068] Redistribute all data according to the calculated grid size, and upsample the original data of a known length into multiple grids of the same size.

[0069] Step 2: Assign known multi-source data to the corresponding areas of the modeling area grid.

[0070] Based on the calculated unit grid size, the modeling area is divided into N grids of the same size, and the known multi-source data are assigned to the corresponding grids. The unknown area grid only contains coordinate information (X, Y, Z), and does not contain attribute information and lithology information.

[0071] Step 3: Build U-Net3D deep learning network;

[0072] like Figure 1 As shown, the U-Net3D network designed by the present invention includes 4 downsamplings, 4 upsamplings and one three-dimensional horizontal convolution.

[0073] Step 3-1, the length, width and height of the area to be modeled are X_Length, Y_Length and Z_Length. During the training process, the modeling area containing some known grids is input into the U-Net3D network and undergoes four downsampling operations. The specific downsampling parameters are set as follows:

[0074] First downsampling: The parameter settings for the first downsampling include a 4*4 three-dimensional convolution layer, where the stride is 2, the padding is 1, and leakyReLu is used as the activation function. After the first downsampling, the length, width and height of the area to be modeled become 1 / 2 of the initial modeling area, and the number of channels becomes 32.

[0075] Second downsampling: The parameter settings for the second downsampling include a 3*3 three-dimensional convolution layer with a stride of 1, a padding of 1, and leakyReLu as the activation function. After the second downsampling, the length, width, and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 64.

[0076] The third downsampling: The parameter settings for the third downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function. After the third downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 128.

[0077] Fourth downsampling: The parameter settings for the fourth downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function. After the fourth downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 256.

[0078] Step 3-2: Further feature extraction operation is performed on the modeling area after four downsampling. The horizontal feature extraction operation includes a 4*4 three-dimensional convolution layer, where the stride is 2, the padding is 1, and leakyReLu is used as the activation function. After the horizontal convolution, the length, width and height of the modeling area are reduced to 1 / 4 of the initial modeling area, and the number of channels is increased to 256.

[0079] Step 3-3, after four downsampling and horizontal convolution operations, the eigenvalues ​​of the extracted three-dimensional model need to be upsampled. The upsampling is also performed in four steps. The specific upsampling parameters are set as follows:

[0080] First upsampling: The parameter settings for the first upsampling include a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function. After the first upsampling, the length, width, and height of the area to be modeled are increased to 1 / 2 of the initial modeling area, and the number of channels is reduced to 128. In addition, the data after the first upsampling and the third downsampling are jump-connected to ensure the accuracy of the model.

[0081] Second upsampling: The parameter setting of the second upsampling includes a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function. After the second upsampling, the length, width and height of the area to be modeled remain 1 / 2 of the initial modeling area, and the number of channels is reduced to 64. The data after the second upsampling and the second downsampling are skipped to ensure the accuracy of the model.

[0082] The third upsampling: The parameter setting of the third upsampling includes a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function. After the third upsampling, the length, width and height of the area to be modeled remain 1 / 2 of the initial modeling area, and the number of channels is reduced to 32. The third upsampling is skipped with the data after the first downsampling to ensure the accuracy of the model.

[0083] Fourth upsampling: The parameter settings for the fourth upsampling include a 4*4 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function. After the fourth upsampling, the length, width and height of the area to be modeled are consistent with those of the initial modeling area, and the prediction results of each cell can be output.

[0084] Step 4: Use the U-Net3D pre-trained deep learning network to invert the modeling area containing known data to construct a three-dimensional geological model using multi-source data.

[0085] A deep learning network that has not been pre-trained cannot predict and classify unknown blank events. Therefore, the present invention is designed to first make a pre-trained model based on the U-Net3D inversion three-dimensional geological model, and then invert the area to be modeled by loading the pre-trained model.

[0086] Step 4-1, manually generate a large number of three-dimensional geological models with random occurrences, random tectonic events and random lithology for training.

[0087] Step 4-2, divide the generated three-dimensional geological models of different events into training set, test set and validation set according to the classification method of 8:1:1, and input the training set into the U-Net3D deep learning network mentioned in step 3 for training; use precision, recall rate and F1 score to evaluate the modeling effect of the test set and validation set.

[0088] Step 4-3. Save the pre-trained model with precision, recall and F1 score greater than 90% for the test set and validation set.

[0089] Step 4-4, load the pre-trained model to perform inversion modeling on the area to be modeled that contains known multi-source data.

[0090] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0091] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network, characterized in that: The following steps are involved: Step 1: extracting drilling data and geophysical profile data; upsampling multi-source data by calculating the minimum cell size; Step 2, assigning known multi-source data to corresponding areas of the modeling area grid; Step 3: Build U-Net3D deep learning network; Step 4: Use the U-Net3D pre-trained deep learning network to invert the modeling area containing known data to construct a three-dimensional geological model using multi-source data.

2. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 1, characterized in that: In step 1, due to the differences in the storage format and representation range of borehole data and geophysical data, it is necessary to upsample the multi-source data. The specific steps are as follows: Step 1-1, determining the data types of drilling data and geophysical data; Step 1-2, determine the representation range of different data, and determine the minimum grid size according to the resolution of different data; Steps 1-3: Determine the uniform depth and perform data upsampling.

3. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 1, characterized in that: In step 2, the known multi-source data are assigned to the corresponding areas of the modeling area grid; based on the calculated unit grid size, the modeling area is divided into N grids of uniform size, and the known multi-source data are assigned to the corresponding grids; the unknown area grid only contains coordinate information (X, Y, Z), and does not contain attribute information and lithology information.

4. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 1, characterized in that: In step 3, the U-Net3D network includes 4 downsamplings, 4 upsamplings, and one 3D horizontal convolution.

5. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 1, characterized in that: In step 4, the U-Net3D pre-trained deep learning network is used to invert the modeling area containing known data to construct a three-dimensional geological model using multi-source data.

6. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 2, characterized in that: In step 1-1, the header information of the drilling data includes lithology name, X, Y, depth from, and depth to. The header information of geophysical data varies depending on the storage format. The geophysical data formats include SEG-Y, LAS, NetCDF, GRD, HDF5, XYZ, and Matlab.mat.

7. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 2, characterized in that: In step 1-2, the coverage of the data is determined. For geophysical data, its coverage is determined and recorded as [d profile min ,d profile max ]; for drilling data, determine its coverage, denoted as [d drill min ,d drill max ]; Select a target grid size that will be used to evenly divide the geophysical and drillhole data. The target grid size is based on the smallest cell size in the geophysical and drillhole datasets.

8. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 2, characterized in that: In steps 1-3, the uniform depth range [d min ,d max ] need to cover the full range of geophysical data and borehole data, i.e. the min =min(d profile min ,d drill min ) d max =max(d proflie max ,d drill max ) The entire depth range [d min ,d max [d min ,d max ] Divide the target grid size Δd and calculate the grid size S grids Data upsampling, redistribute all data according to the calculated grid size, and upsample the original data of a known length into multiple grids of the same size.

9. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 4, characterized in that: In step 3, the following steps are specifically included: Step 3-1, the length, width and height of the area to be modeled are X_Length, Y_Length and Z_Length; during the training process, the modeling area containing some known grids is input into the U-Net3D network and undergoes four downsampling operations. The specific downsampling parameters are set as follows: The first downsampling, the parameter settings of the first downsampling include a 4*4 three-dimensional convolution layer, where the stride is 2, the padding is 1, and leakyReLu is used as the activation function; after the first downsampling, the length, width and height of the area to be modeled become 1 / 2 of the initial modeling area, and the number of channels becomes 32; The second downsampling, the parameter settings of the second downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function; after the second downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 64; The third downsampling, the parameter settings of the third downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function; after the third downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 128; The fourth downsampling, the parameter settings of the fourth downsampling include a 3*3 three-dimensional convolution layer, where the stride is 1, the padding is 1, and leakyReLu is used as the activation function; after the fourth downsampling, the length, width and height of the area to be modeled are still 1 / 2 of the initial modeling area, and the number of channels is increased to 256; Step 3-2: Further feature extraction operation is performed on the modeling area after four downsampling. The horizontal feature extraction operation includes a 4*4 three-dimensional convolution layer, where the stride is 2, the padding is 1, and leakyReLu is used as the activation function; after horizontal convolution, the length, width and height of the modeling area are reduced to 1 / 4 of the initial modeling area, and the number of channels is increased to 256; Step 3-3: After four downsampling and horizontal convolution operations, the eigenvalues ​​of the extracted three-dimensional model need to be upsampled. The upsampling is also performed in four steps. The specific upsampling parameters are set as follows: The first upsampling, the parameter settings of the first upsampling include a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function; after the first upsampling, the length, width and height of the area to be modeled are increased to 1 / 2 of the initial modeling area, and the number of channels is reduced to 128; in addition, the data after the first upsampling and the third downsampling are jump-connected to ensure the accuracy of the model; The second upsampling, the parameter settings of the second upsampling include a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function; after the second upsampling, the length, width and height of the area to be modeled still remain 1 / 2 of the initial modeling area, and the number of channels is reduced to 64; the data after the second upsampling and the second downsampling are jump-connected to ensure the accuracy of the model; The third upsampling, the parameter settings of the third upsampling include a 3*3 three-dimensional convolution layer, where the stride is 2, the padding is 1, and ReLu is used as the activation function; after the third upsampling, the length, width and height of the area to be modeled still remain 1 / 2 of the initial modeling area, and the number of channels is reduced to 32; the third upsampling is skipped with the data after the first downsampling to ensure the accuracy of the model; The fourth upsampling, the parameter setting of the fourth upsampling includes a 4*4 three-dimensional convolution layer, in which the stride is 2, the padding is 1, and ReLu is used as the activation function; after the fourth upsampling, the length, width and height of the area to be modeled are consistent with those of the initial modeling area, and the prediction result of each cell can be output.

10. A modeling method for processing multi-source geological data and constructing a three-dimensional geological model based on a pre-trained U-Net3D network according to claim 5, characterized in that: In step 4, based on the pre-trained model of the U-Net3D inverted three-dimensional geological model, the pre-trained model is loaded to invert the area to be modeled, including the following steps: Step 4-1, manually generate a large number of three-dimensional geological models of random occurrences, random tectonic events and random lithology for training; Step 4-2, divide the generated 3D geological models of different events into training set, test set and validation set according to the classification method of 8:1:1, and input the training set into the U-Net3D deep learning network mentioned in step 3 for training; use precision, recall rate and F1 score to evaluate the modeling effect of the test set and validation set; Step 4-3, save the pre-trained model with precision, recall and F1 score greater than 90% for the test set and validation set; Step 4-4: Load the pre-trained model to perform inversion modeling on the area to be modeled that contains known multi-source data.

Images