A three-dimensional geological modeling method and system based on deep learning

Abstract

The application discloses a kind of three-dimensional geological modeling method and system based on deep learning, the method is first to the modeling area is drilled and the unit grid of known geological attribute is as reference unit;Again, target unit is combined with any N reference units to build data set containing order of magnitude sample, wherein, the three-dimensional coordinates and geological attribute of each sample data are combined with the three-dimensional coordinates of target unit to form sample feature matrix, and the geological attribute of target unit is as label;Then, the sample feature matrix is extracted using auto-encoder to obtain the depth feature map of sample;Finally, the geological attribute prediction model of discrete unit based on deep convolutional neural network is constructed, the depth feature map of sample and label are used for model training, for traversing the geological attribute of unknown unit grid, realize three-dimensional geological modeling, solve the technical obstacles existing in the limited known geological attribute data, improve modeling precision.

Description

Technical Field

[0001] This invention pertains to three-dimensional geological modeling technology, specifically relating to a three-dimensional geological modeling method and system based on deep learning. Background Technology

[0002] Three-dimensional geological modeling technology can invert complete geological structural information from limited exploration information. It plays an important foundational role in mineral resource exploration, hydrogeological surveys, mineral resource mining, and underground engineering excavation.

[0003] Traditional geological modeling methods primarily rely on explicit modeling. Explicit modeling is a method based on defining geological structural boundaries using a series of contour lines. For example, in mining engineering, contour lines of the ore body are manually delineated from borehole sampling data, and a three-dimensional (3D) ore body model is periodically reconstructed. However, explicit modeling methods suffer from drawbacks such as the inability to dynamically update the model, over-reliance on experience, difficulty in real-time model updates, and inability to reflect special geological structures. To address the problems of explicit modeling and improve the efficiency and accuracy of geological modeling, implicit modeling methods have been proposed. Implicit modeling uses implicit methods, such as formulas or functions, to represent a geometric surface. However, traditional implicit modeling methods require a high degree of model control.

[0004] With the development of artificial intelligence technology, machine learning theory has been introduced into the research of implicit modeling. Goncalveset et al. defined the implicit modeling problem as multi-class classification in machine learning and used maximum likelihood to train the model, which became the main idea of ​​many subsequent studies. De La Varga and Grosse combined uncertain prior information with the geological motivation likelihood function in the Bayesian inference framework to reduce the uncertainty of the generated model set. Deep learning provides a more advanced method than machine learning for implicit modeling research. Bi and Wu treated implicit structural modeling as an image rendering task, using channel concern points and multi-scale fusion convolutional neural networks (CNNs) to systematically aggregate features at different spatial resolutions, and used a hybrid loss function to enhance tectonic boundaries and remove blurred geometric features to produce realistic geological models. Hillier et al. proposed two geometric deep learning methods, using graph neural networks (GNNs) and implicit neural representations (INRs), utilizing unstructured grids as graphs to perform coupled implicit and discrete geological unit modeling on them. Compared with traditional implicit modeling methods, implicit modeling methods that incorporate machine learning and deep learning reduce the control requirements of the model. By discretizing the geological region and classifying the geological information of the discrete units, an implicit interpolation function for the region is constructed.

[0005] Therefore, implicit modeling methods based on artificial intelligence are a development trend in this field. However, existing deep learning methods still have shortcomings in terms of dataset organization, input / output settings, network architecture design, 3D modeling accuracy, and intelligence level, making it difficult to meet the actual needs of 3D geological implicit modeling. First, deep learning requires a large amount of data for training, and how to obtain or construct a sufficiently large dataset is one of the main technical obstacles to the application of this technology. Second, current deep learning methods do not fully mine features in the geological modeling process, resulting in significant errors in modeling accuracy. Summary of the Invention

[0006] To overcome the limitations of existing technologies in acquiring large-scale datasets and improving modeling accuracy, this invention provides a deep learning-based 3D geological modeling method and system. Specifically, the technical solution of this invention discretizes the region to be modeled into grid cells. Utilizing the relationships between borehole grid cells, a large-scale deep learning training set can be formed using limited borehole data by combining cells to create a feature matrix, thus meeting the requirements of deep learning. Furthermore, an unsupervised learning method using an autoencoder is used to extract the deep feature matrix, eliminating potential cell order features in the aforementioned sample feature matrix. That is, while N cells are actually randomly selected, a natural order is created when they are combined, but this order feature does not actually exist. Therefore, this invention further introduces an autoencoder to eliminate this order feature, further ensuring the reliability of data features and improving modeling accuracy.

[0007] Therefore, the present invention provides the following technical solution:

[0008] On the one hand, the present invention provides a three-dimensional geological modeling method based on deep learning, comprising the following steps:

[0009] Step 1: Conduct borehole exploration in the modeling area to obtain borehole exploration data and discretize the modeling area into a cell network, and then use the cell grid with known geological properties through which the borehole channel passes as the reference cell.

[0010] Step 2: Select the target cell and arbitrarily choose reference cells to combine and construct a dataset containing orders of magnitude of samples. Each sample data is for the target cell, and the three-dimensional coordinates x of N reference cells are used. i ,y i ,z i and geological attributes L i The three-dimensional coordinates x of the target unit label ,y label ,z label The sample feature matrix is ​​formed by combining the features, and the geological attributes L of the target unit are also included. label For tags;

[0011] Step 3: Use an autoencoder to extract deep features from the sample feature matrix to obtain the sample's deep feature map;

[0012] Step 4: Construct a geological attribute prediction model for discrete units based on a deep convolutional neural network, and train the model using the depth feature map and labels of the samples.

[0013] Among them, the trained geological attribute prediction model predicts the geological attributes of discrete unit grids with unknown geological attributes.

[0014] Preferably, the geological properties of the unit grids through which the passage from the surface to the ground corresponds to a single borehole location are all determined based on borehole exploration data, and the passage location is regarded as a vertical superposition of several unit grids with known geological properties.

[0015] During the dataset construction process, the unit grid is arbitrarily selected as the target unit at the vertical position of the borehole channel, and the three-dimensional coordinates and geological attributes of the reference units at the same borehole channel and / or different borehole channels are fused to construct the sample feature matrix.

[0016] Preferably, N is 2048 or any multiple of 2 greater than 2028. If N is 2048, then the sample feature matrix is ​​a 2048×7 feature matrix, expressed as:

[0017]

[0018] In the formula, x1, y1, z1, L1 are the three-dimensional coordinates and geological properties of the first reference unit, and x2, y2, z2, L2 are the three-dimensional coordinates and geological properties of the second reference unit. 2048 ,y 2048 ,z 2048 ,L 2048 The three-dimensional coordinates and geological properties of the 2048th reference unit.

[0019] Preferably, the autoencoder consists of three convolutional layers, three max pooling layers, and three deconvolutional layers. The network structure is set as follows: convolutional layer 1, max pooling layer 1, convolutional layer 2, max pooling layer 2, convolutional layer 3, and max pooling layer 3 to obtain a depth feature map, which is then decoded by the three deconvolutional layers in sequence.

[0020] The input to convolutional layer 1 is the sample feature matrix;

[0021] The training of the autoencoder aims to reduce the difference between the vector corresponding to the input sample feature matrix and the output vector decoded by three deconvolutional layers, and adjusts the network parameters of the autoencoder accordingly.

[0022] Preferably, the deep convolutional neural network consists of n residual blocks connected sequentially and one capsule network block, and preferably n is greater than 8;

[0023] The input to the first residual block is the depth feature map of the sample, the input to other residual blocks is the output of the previous residual block, and the output of the last residual block is used as the input to the capsule network. The capsule network outputs the geological attribute prediction result of the target unit.

[0024] Preferably, the geological attribute is lithology.

[0025] Secondly, the present invention provides a deep learning-based 3D modeling method, comprising the following steps:

[0026] S1: Discretize the modeling object into a cell network and use the cell mesh with known properties as the reference cell;

[0027] S2: Select the target cell and arbitrarily choose reference cells to combine and construct a dataset containing orders of magnitude of samples. Each sample data is for the target cell, and the three-dimensional coordinates x of N reference cells are used. i ,y i ,z i and attributes l i The three-dimensional coordinates x of the target unit label ,y label ,z label The sample feature matrix is ​​formed by combining the attributes of the target unit. label For tags;

[0028] S3: Use an autoencoder to extract deep features from the sample feature matrix to obtain the sample's deep feature map;

[0029] S4: Construct an attribute prediction model for discrete units based on a deep convolutional neural network, and train the model using the deep feature maps and labels of the samples.

[0030] Among them, the trained attribute prediction model predicts the attributes of discrete cell grids with unknown attributes.

[0031] In three aspects, a system based on the above modeling method includes at least the following:

[0032] Mesh generation units are used to acquire borehole exploration data of the modeling area and to discretize the modeling area into a unit network, thereby using the unit grids with known geological properties along the borehole passage as reference units; or to discretize the modeling object into a unit network and use the unit grids with known properties as reference units.

[0033] Dataset construction units are used to select target units and arbitrarily choose reference units to combine and construct a dataset containing orders of magnitude of samples. Each sample data is for the target unit, and the three-dimensional coordinates x of N reference units are used to construct the dataset. i ,y i ,z i and geological attributes L i The three-dimensional coordinates x of the target unit label ,y label ,z label The sample feature matrix is ​​formed by combining the features, and the geological attributes L of the target unit are also included. label For labels; or for selecting target units and arbitrarily choosing reference units to construct a dataset containing orders of magnitude of samples, where each sample is for the target unit, and the three-dimensional coordinates x of N reference units are used. i ,y i ,z i and attributes l i The three-dimensional coordinates x of the target unit label ,y label ,z label The sample feature matrix is ​​formed by combining the attributes of the target unit. label For tags;

[0034] The deep feature extraction unit is used to extract deep features from the sample feature matrix using an autoencoder to obtain the deep feature map of the sample.

[0035] The model training unit is used to construct a geological attribute prediction model or attribute prediction model based on discrete units of a deep convolutional neural network, and to train the model using the deep feature map and labels of the samples.

[0036] Prediction unit, used by a trained geological attribute prediction model or attribute prediction model to predict the geological attributes or properties of a discrete cell grid of unknown attributes.

[0037] In four aspects, the computer device provided by the present invention includes at least:

[0038] One or more processors;

[0039] A memory that stores one or more computer programs;

[0040] The processor calls the computer program to implement:

[0041] The steps of a deep learning-based 3D geological modeling method or the steps of a deep learning-based 3D modeling method.

[0042] In five aspects, the present invention provides a computer storage device that stores a computer program, which is called by a processor to implement: the steps of a deep learning-based three-dimensional geological modeling method or the steps of a deep learning-based three-dimensional modeling method.

[0043] Beneficial effects

[0044] Compared with existing methods, the advantages of the present invention are:

[0045] 1. Based on discretizing the region to be modeled into grid cells, the present invention proposes a self-organizing method for borehole data: taking a random borehole cell A as an example, the three-dimensional coordinates x, y, z of this cell, and the three-dimensional coordinates x of any other cell... i ,y i ,z i and geological properties L obtained through drilling i Seven values ​​form one row of the feature matrix of unit A. Keeping A unchanged, N other borehole units are randomly selected to form an N×7 sample feature matrix corresponding to unit A. Correspondingly, the geological attributes of unit A are called the labels corresponding to the aforementioned sample feature matrix in supervised learning. By constructing the sample feature matrix, the problem of borehole sparsity is solved. A dataset of any size can be obtained through a one-to-many approach, realizing the data foundation for deep learning. This effectively solves the technical obstacle of insufficient training dataset due to limited data in deep learning. Then, this invention proposes an encoder-decoder network model suitable for sample feature matrices as input to mine deep feature matrices. This eliminates the correlation error introduced by the unit order of the sample feature matrix, further ensuring the accuracy of feature data and improving modeling accuracy.

[0046] 2. The preferred technical solution of this invention combines residual networks and capsule networks. Using a deep feature matrix as input, the residual network enhances the network depth, and the capsule network retains feature information, ultimately achieving high-precision lithology prediction, i.e., lithology prediction based on convolutional neural networks. Through the lithology prediction method, each cell is predicted, ultimately realizing the lithology prediction of each cell in the three-dimensional model, i.e., the construction of a three-dimensional geological model. Attached Figure Description

[0047] Figure 1 These are exploration schematic diagrams of a 3D geological modeling area. Figure a is a schematic diagram of the 3D geological structure, and Figure b is a schematic diagram of borehole exploration.

[0048] Figure 2 This is a schematic diagram of the discretization preprocessing of borehole exploration data;

[0049] Figure 3 This is a schematic diagram of the construction of the sample feature matrix;

[0050] Figure 4 This is a schematic diagram of the network structure of an autoencoder;

[0051] Figure 5 This is a schematic diagram of the structure of a deep convolutional neural network;

[0052] Figure 6 This is a schematic diagram illustrating the prediction of unknown attribute cells in a three-dimensional mesh.

[0053] Figure 7 This is a flowchart illustrating the three-dimensional geological modeling method provided by the technical solution of this invention. Detailed Implementation

[0054] This invention provides a deep learning-based 3D geological modeling method and system. Specifically, it uses the channel structure of borehole exploration combined with a discrete grid to deduce a novel sample set construction method. The proposed sample feature matrix construction method enables data-level sample construction and extracts depth features through an autoencoder, eliminating potential unit order features in the aforementioned sample feature matrix, thus meeting the requirements of deep learning. Furthermore, this technical approach can be extended to other fields, specifically applicable to objects that construct complete regional models from sparse data, where the attributes of the modeled object are related to the grid cell positions. The invention will be further described below with reference to embodiments.

[0055] Example 1

[0056] This invention illustrates a three-dimensional geological modeling method based on deep learning, using lithology as an example. The method includes the following steps:

[0057] Step 1: Discretize the area to be modeled in 3D and obtain borehole exploration data. Then, use the grid with known geological properties as the reference unit. Geological properties are determined by borehole exploration data, such as lithology (soil, quartz, lead-zinc ore, etc.).

[0058] like Figure 1 As shown, borehole exploration is conducted in the area to be modeled in three dimensions. The borehole exploration data obtained through multiple borehole explorations constitutes the data foundation for the three-dimensional geological modeling of this area. For example, in this embodiment, core data is obtained by vertically sampling the strata. Figure 2 As shown, based on the accuracy requirements of geological modeling, the area to be modeled (the area to be modeled in 3D) is discretized into smaller grid cells. Based on the rock core extracted from the borehole, the grid cells through which the borehole passes are assigned corresponding lithology, thus creating cells with unknown geological properties and cells with known geological properties (reference cells).

[0059] Step 2: Construct the training and test sets.

[0060] Specifically, this invention targets the target unit and uses the three-dimensional coordinates x of N reference units. i ,y i ,z i and geological attributes L i The three-dimensional coordinates x of the target unit label ,y label ,z label and geological properties L label Combine them to form a feature matrix.

[0061] In this embodiment, if N is set to 2048, a feature matrix of size 2048×7 is formed for any target cell. Figure 3 As shown in Figure a. Subsequently, the feature matrices formed using all geologically known units as target units are aggregated to create the training and test sets for supervised learning.

[0062] The entire area to be modeled was divided into a cell grid. For a single borehole, the location of the borehole from the surface to the subsurface consists of several cells stacked vertically, and the lithology of these cells is known (determined through borehole exploration). Therefore, all the boreholes form a grid similar to... Figure 3 Figure a shows multiple stacked columns composed of cells. Taking any cell A from these drilled cells as the target cell, its three-dimensional coordinates are (x...). label ,y label ,z label The lithology of this cell is denoted as L. label Then, 2048 cells are randomly selected from other drilling units, with three-dimensional coordinates (x... i ,y i ,z i ), i = 1, 2, 3, ..., I, ..., 2048, the lithology of the corresponding cell is denoted as L. i ,i=1,2,3…,I,…,2048; Then combine the three-dimensional coordinates and lithology of cell A with those of other cells one by one, that is, construct a data vector for the first cell out of 2048 cells: x label ,y label ,z label The second unit constructs a data vector: x1, y1, z1, L1. label ,y label ,z label The data vector is constructed from the i-th unit: x2, y2, z2, L2. label ,y label ,z label ,x i ,y i ,z i ,Li The 2048th unit constructs the data vector: x label ,y label ,z label ,x 2048 ,y 2048 ,z 2048 ,L 2048 These arrays are stacked to form a 2048×7 feature matrix, which serves as the input to the deep learning network. The corresponding training label is the lithology L of cell A. label This results in a single input-label pair required for supervised learning. By repeating the above steps with randomly selected cells, theoretically, datasets of any size can be obtained. For example... Figure 3 As shown, the sample feature matrix in this embodiment is constructed as follows:

[0063]

[0064] Similarly, the matrix of an element with unknown lithology can be represented as:

[0065]

[0066] In the formula, x pred ,y pred ,z pred The three-dimensional coordinates of grid cells considered as having unknown geological properties correspond to the label L to be predicted. pred This refers to the geological properties of the grid cell, specifically the lithology in this embodiment.

[0067] The technical solution of this invention utilizes the above-mentioned technical means to form a corresponding unit feature matrix by taking advantage of the interrelationship between borehole units, thereby realizing the organization of features and construction of datasets for limited borehole data and solving the technical obstacle of large amounts of data required for deep learning.

[0068] Step 3: Use an autoencoder to extract deep features from the feature matrix, that is, build an Autoencoder network and train it to extract deep features.

[0069] The technical solution of this invention, after preprocessing borehole sampling data, utilizes an autoencoder network to extract more geological information, ensuring the accuracy of lithology prediction. The autoencoder effectively performs supervised learning compression of the original input through encoding and decoding mechanisms, making it an excellent tool for automatically extracting data features. Therefore, the technical solution of this invention constructs such... Figure 4 The Autoencoder network shown consists of three convolutional layers, three max-pooling layers, and three deconvolutional layers, in the following order: Convolutional layer 1, Max-pooling layer 1, Convolutional layer 2, Max-pooling layer 2, Convolutional layer 3. Max-pooling layer 3 produces a depth feature map, which is then decoded by the three deconvolutional layers. Figure 4 The encoder, depth feature map, and decoder shown in this embodiment of the invention do not limit the parameters (size) of each network layer. They are set according to the input / output and network accuracy requirements, and the number of network layers can also be adjusted according to accuracy requirements. The input to the first convolutional layer is the feature matrix constructed in the previous step, i.e., the 2048×7 feature matrix in this embodiment.

[0070] The output of a convolutional layer can be represented as:

[0071]

[0072] Where Relu(x) = max(0,x) is the activation function, x ij It is the i-th input of the j-th neuron, k ij It is the convolution kernel between the j-th neuron and the i-th input, b j It is the bias of neuron j, M j This refers to the selection of input features. In the Autoencoder, the encoder can be represented as:

[0073] y = g φ (x) (2)

[0074] Where x represents the input sample and y represents the encoder output, the decoder can be represented as:

[0075] x = f θ (g φ (x)) (3)

[0076] Where (θ, φ) are parameters to be learned, f θ The objective function of the autoencoder represents the decoding process:

[0077] x≈f θ (g φ (x)) (4)

[0078] This invention uses evaluation metrics to compare the difference between two vectors, such as cross-entropy or a simple MSE (mean squared error) loss function. This invention does not limit the type of evaluation metric chosen, as shown in the following loss function L. AE (θ,φ) can be represented as:

[0079]

[0080] Where n represents the number of samples used for training in a single iteration, x (i) Let i be the i-th sample.

[0081] Step 4: Lithology prediction based on deep convolutional neural networks.

[0082] This invention combines residual blocks and capsule networks to construct a lithology prediction model for discrete units based on a deep convolutional neural network. To accurately predict the lithology of each element, such as... Figure 5 As shown, a deep convolutional neural network consists of n residual blocks and one capsule network block. In lithology prediction based on deep convolutional neural networks, the residual blocks are responsible for mining the information contained in the data, and the data size changes every two residual blocks. The capsule network block is a modern deep learning model that can overcome the information loss problem in fully connected networks and make full use of limited data for deep learning. From the detailed descriptions of residual and capsule network blocks, it can be seen that their similarity lies in retaining as much information as possible when mining features to achieve the final goal. Therefore, this invention preferably combines residual blocks and capsule network blocks to form a deep convolutional neural network model to predict the lithology of each element in a three-dimensional discrete geological model; in other feasible embodiments, other neural networks can be selected for lithology prediction, provided that the application accuracy requirements are met.

[0083] In this approach, the first residual block takes the depth feature map and labels obtained from the autoencoder in the previous step as input, while the other residual blocks take the output of the previous residual block as input. The core idea of ​​residual blocks is to fuse the extracted features with the input features, ensuring the integrity of the feature information. This can be simply described as:

[0084] y1'=g φ2 (g φ1 (x1))+x1 (6)

[0085] In the formula, g φ1 g φ2 This represents the convolutional layer of the residual block, where x1 is the input of the residual block and y1' is the output of the residual block.

[0086] The output of the last residual block serves as the input to the capsule network. Capsule network blocks within the capsule network are used to retain as much key information as possible during the deep learning process. Each capsule network block is fully connected between primary and higher-level capsules. That is, in the capsule network, primary and higher-level capsules establish a hierarchical feature representation relationship through a dynamic routing algorithm. Primary capsules are responsible for extracting local, low-level visual features (such as edges and textures), and their output vectors represent the probability of existence and pose parameters of specific features. Higher-level capsules adaptively learn the weight relationships between the outputs of each primary capsule through dynamic routing, aggregating and modeling higher-level global or semantic features (such as object parts or the whole). The dynamic routing process iteratively optimizes the information transfer strategy from primary to higher-level capsules, enabling features to be combined in a reasonable "part-whole" structure, ultimately forming an expression with spatial consistency and hierarchical abstraction capabilities. The capsule network and its blocks are existing structures, and this invention will not describe them in detail.

[0087] Similarly, capsule network blocks have been proposed to retain as much key information as possible during deep learning. Unlike fully connected layers in CNNs, capsule network blocks consist of fully connected primary and higher-level capsules. A capsule is a vector that can contain all the information from the input data, including its position and orientation. The core of capsule network blocks is a dynamic routing algorithm. This algorithm uses a compression function, formalized as:

[0088]

[0089] Among them, v j It is the output of advanced capsule j, s j This is the input for advanced capsule j, along with input s. j It is also the prediction vector of all capsules in the previous layer. The weighted sum. It is the output u of the upper capsule. i With weighted matrix W ij The result of the multiplication is the prediction vector:

[0090]

[0091] Among them, c ij Let b be the coupling coefficient, representing the consistency between capsule i and capsule j. ij and b ik b is the log-prior probability between the two coupled capsules. ij Update using the following formula

[0092]

[0093] b ij The initial value is 0, and the weighting matrix W ij Initially generated randomly. ij The dynamic update continuously optimizes the coupling coefficient c. ij .

[0094] It should be understood that by using deep feature maps as input to a deep convolutional neural network, the lithology prediction results of the target unit are obtained. During the network training process, the network is trained using the labels of the target units, and the network parameters are updated by constructing a loss function.

[0095] In this embodiment, all parameters are evaluated using the following loss function L. k Update:

[0096] L k =T k max(0,m + -||v k ||) 2 +λ(1-T k max(0,||v) k ||-m - ) 2 (12)

[0097] Among them, T k =1,m + =0.9,m - =0.1, T is the label indicator function (1 when category c exists), and m is the preset boundary threshold (forcing positive sample probabilities close to 1 and negative sample probabilities close to 0). In classification tasks, if the sample is hematite (category 5), setting T5 = 1, the loss will penalize the case where ||v5|| < 0.9. For other non-category 5 capsules (such as pyrite of category 3), the loss will be constrained by ||v3|| < 0.1. λ can reduce the loss weight of missing categories and prevent the initial learning from shrinking the length of the activity vectors of all high-level capsules. k This is the L2 norm output by the capsule corresponding to lithology category k.

[0098] In summary, the 2048×7 matrix corresponding to any borehole cell is input into the Autoencoder. The Autoencoder does not require labels; its network is trained by comparing the output with the input. After the Autoencoder is trained, its output is not used; instead, the results of the intermediate layers are used as a depth feature map. This depth feature map is then input into the capsule network and trained along with the lithology of the corresponding cell as the classification label, thus obtaining a trained lithology prediction model.

[0099] It should be understood that, such as Figure 6As shown, this invention constructs a method for building a three-dimensional discrete geological model through preprocessing of borehole sampling data, feature extraction using an Autoencoder, and lithology prediction based on ResCapsNet. Each element of the three-dimensional discrete geological model undergoes traversal processing: First, each element is preprocessed to obtain an initial feature matrix. Then, the initial feature matrix is ​​used as input to the Autoencoder, which outputs a depth feature map. Finally, using the depth feature map as input, the lithology prediction model ResCapsNet is used to predict the lithology of the elements. These steps are repeated until the lithology of all elements is predicted, thus completing the construction of the three-dimensional geological model.

[0100] Example 2

[0101] This embodiment of the invention differs from Embodiment 1, which is limited to the geological field. This embodiment is applicable to objects that construct complete regional models using sparse data. This embodiment provides a deep learning-based 3D modeling method, including the following steps:

[0102] S1: Discretize the modeling object into a cell network and use the cell mesh with known properties as the reference cell;

[0103] S2: Select the target cell and arbitrarily choose reference cells to combine and construct a dataset containing orders of magnitude of samples. Each sample data is for the target cell, and the three-dimensional coordinates x of N reference cells are used. i ,y i ,z i and attributes l i The three-dimensional coordinates x of the target element label ,y label ,z label The sample feature matrix is ​​formed by combining the attributes of the target unit. label For tags;

[0104] S3: Use an autoencoder to extract deep features from the sample feature matrix to obtain the sample's deep feature map;

[0105] S4: Construct an attribute prediction model for discrete units based on a deep convolutional neural network, and train the model using the deep feature maps and labels of the samples; wherein, the trained attribute prediction model predicts the attributes of discrete unit grids with unknown attributes.

[0106] Example 3

[0107] The present invention also provides a three-dimensional modeling system based on the modeling object described in Embodiment 1 or Embodiment 2, which includes: a mesh generation unit, a dataset construction unit, a depth feature extraction unit, a model training unit, and a prediction unit.

[0108] Among them, the grid division unit is used to obtain borehole exploration data of the modeling area and to discretize the modeling area into a unit network, and then use the unit grid with known geological properties through which the borehole channel passes as the reference unit; or it is used to discretize the modeling object into a unit network and use the unit grid with known properties as the reference unit.

[0109] Dataset construction units are used to select target units and arbitrarily choose reference units to combine and construct a dataset containing orders of magnitude of samples. Each sample data is for the target unit, and the three-dimensional coordinates x of N reference units are used to construct the dataset. i ,y i ,z i and geological attributes L i The three-dimensional coordinates x of the target element label ,y label ,z label The sample feature matrix is ​​formed by combining the features, and the geological attributes L of the target unit are also included. label For labels; or for selecting target units and arbitrarily choosing reference units to construct a dataset containing orders of magnitude of samples, where each sample is for the target unit, and the three-dimensional coordinates x of N reference units are used. i ,y i ,z i and attributes l i The three-dimensional coordinates x of the target element label ,y label ,z label The sample feature matrix is ​​formed by combining the attributes of the target unit. label For tags;

[0110] The deep feature extraction unit is used to extract deep features from the sample feature matrix using an autoencoder to obtain the deep feature map of the sample.

[0111] The model training unit is used to construct a geological attribute prediction model or attribute prediction model based on discrete units of a deep convolutional neural network, and to train the model using the deep feature map and labels of the samples.

[0112] Prediction unit, used by a trained geological attribute prediction model or attribute prediction model to predict the geological attributes or properties of a discrete cell grid of unknown attributes.

[0113] It should be understood that the specific implementation process of each module is described in the above method. This invention will not repeat the details here. The above division of functional modules is only for illustrative purposes. In some embodiments, some functional modules can be combined and some functional modules can be separated. Each functional module can be implemented in software, hardware, or a combination of software and hardware. The software and hardware devices include, but are not limited to, general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0114] Example 4

[0115] The present invention also provides a computer device, comprising at least:

[0116] One or more processors;

[0117] A memory that stores one or more computer programs;

[0118] The processor calls the computer program to implement:

[0119] The steps of a deep learning-based 3D geological modeling method or the steps of a deep learning-based 3D modeling method.

[0120] The specific process of implementing a deep learning-based 3D geological modeling method is as follows:

[0121] Step 1: Discretize the area to be modeled in 3D and obtain borehole exploration data, and then use the grid with known geological properties as the reference unit.

[0122] Step 2: Construct the training and test sets.

[0123] Step 3: Use an autoencoder to extract deep features from the feature matrix, that is, build an Autoencoder network and train it to extract deep features.

[0124] Step 4: Lithology prediction based on deep convolutional neural networks.

[0125] The specific process of implementing a deep learning-based 3D modeling method is as follows:

[0126] S1: Discretize the modeling object into a cell network and use the cell mesh with known properties as the reference cell;

[0127] S2: Select the target cell and arbitrarily choose reference cells to combine and construct a dataset containing orders of magnitude of samples. Each sample data is for the target cell, and the three-dimensional coordinates x of N reference cells are used. i ,y i ,z i and attributes l iThe three-dimensional coordinates x of the target unit label ,y label ,z label The sample feature matrix is ​​formed by combining the attributes of the target unit. label For tags;

[0128] S3: Use an autoencoder to extract deep features from the sample feature matrix to obtain the sample's deep feature map;

[0129] S4: Construct an attribute prediction model for discrete units based on a deep convolutional neural network, and train the model using the deep feature maps and labels of the samples; wherein, the trained attribute prediction model predicts the attributes of discrete unit grids with unknown attributes.

[0130] Please refer to the explanation of the method above for the specific implementation process of each step.

[0131] It should be understood that, in the embodiments of the present invention, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0132] Example 5

[0133] The present invention also provides a computer storage device storing a computer program, which is called by a processor to implement: the steps of a deep learning-based three-dimensional geological modeling method or the steps of a deep learning-based three-dimensional modeling method.

[0134] The specific process of implementing a deep learning-based 3D geological modeling method is as follows:

[0135] Step 1: Discretize the area to be modeled in 3D and obtain borehole exploration data, and then use the grid with known geological properties as the reference unit.

[0136] Step 2: Construct the training and test sets.

[0137] Step 3: Use an autoencoder to extract deep features from the feature matrix, that is, build an Autoencoder network and train it to extract deep features.

[0138] Step 4: Lithology prediction based on deep convolutional neural networks.

[0139] The specific process of implementing a deep learning-based 3D modeling method is as follows:

[0140] S1: Discretize the modeling object into a cell network and use the cell mesh with known properties as the reference cell;

[0141] S2: Select the target cell and arbitrarily choose reference cells to combine and construct a dataset containing orders of magnitude of samples. Each sample data is for the target cell, and the three-dimensional coordinates x of N reference cells are used. i ,y i ,z i and attributes l i The three-dimensional coordinates x of the target unit label ,y label ,z label The sample feature matrix is ​​formed by combining the attributes of the target unit. label For tags;

[0142] S3: Use an autoencoder to extract deep features from the sample feature matrix to obtain the sample's deep feature map;

[0143] S4: Construct an attribute prediction model for discrete units based on a deep convolutional neural network, and train the model using the deep feature maps and labels of the samples; wherein, the trained attribute prediction model predicts the attributes of discrete unit grids with unknown attributes.

[0144] Please refer to the explanation of the method above for the specific implementation process of each step.

[0145] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the hardware and software device described in any of the foregoing embodiments, such as the hard drive or memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the controller. Further, the readable storage medium can include both internal storage units and external storage devices of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0146] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0147] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application refers to flowchart illustrations and / or instructions executed by a processor of a method, apparatus (system), and computer program product according to embodiments of this application to create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more blocks of a block diagram.

[0148] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1. A three-dimensional geological modeling method based on deep learning, characterized in that: Includes the following steps: Step 1: Conduct borehole exploration in the modeling area to obtain borehole exploration data and discretize the modeling area into a cell network, and then use the cell grid with known geological properties through which the borehole channel passes as the reference cell. Step 2: Select the target cell and arbitrarily choose reference cells to combine and construct a dataset containing orders of magnitude of samples. Each sample data is specific to the target cell. N Three-dimensional coordinates of each reference unit x i , y i , z i and geological properties L i Three-dimensional coordinates of the target element x label , y label , z label The sample feature matrix is ​​formed by combining the features, and the geological attributes of the target unit are included. L label For tags; where, N The value of is 2048 or any multiple of 2 greater than 2028. N If the value is 2048, then the sample feature matrix is ​​a 2048×7 feature matrix, expressed as: ； In the formula, x 1 , y 1 , z 1 , L 1 x1 represents the three-dimensional coordinates and geological properties of the first reference unit, and x2, y2, z2, L2 represent the three-dimensional coordinates and geological properties of the second reference unit. 2048 , y 2048 , z 2048 ,L 2048 The three-dimensional coordinates and geological properties of the 2048th reference unit; Step 3: The autoencoder is used to extract deep features from the sample feature matrix to obtain the sample's deep feature map. The autoencoder consists of three convolutional layers, three max-pooling layers, and three deconvolutional layers. The network structure is set sequentially as follows: Convolutional layer 1, max-pooling layer 1, convolutional layer 2, max-pooling layer 2, convolutional layer 3. The deep feature map is obtained from max-pooling layer 3 and then decoded sequentially through the three deconvolutional layers. The input to convolutional layer 1 is the sample feature matrix. The training of the autoencoder aims to reduce the difference between the vector corresponding to the input sample feature matrix and the output vector decoded by the three deconvolutional layers, thereby adjusting the network parameters of the autoencoder. Step 4: Construct a geological attribute prediction model for discrete units based on a deep convolutional neural network. Train the model using the depth feature maps and labels of the samples. The deep convolutional neural network is composed of… n The system consists of a residual block and a capsule network block connected in sequence. The input of the first residual block is the depth feature map of the sample, the input of the other residual blocks is the output of the previous residual block, and the output of the last residual block is used as the input of the capsule network. The capsule network outputs the geological attribute prediction result of the target unit. Among them, the trained geological attribute prediction model predicts the geological attributes of discrete unit grids with unknown geological attributes.

2. The method according to claim 1, characterized in that: The geological properties of the unit grids through which the passage from the surface to the ground corresponds to a single borehole location are determined based on borehole exploration data. The passage location is considered as a vertical superposition of several unit grids with known geological properties. During the dataset construction process, the unit grid is arbitrarily selected as the target unit at the vertical position of the borehole channel, and the three-dimensional coordinates and geological attributes of the reference units at the same borehole channel and / or different borehole channels are fused to construct the sample feature matrix.

3. The method according to claim 1, characterized in that: The geological attribute mentioned is lithology.

4. A system based on the method of any one of claims 1-3, characterized in that: At least including: The grid division unit is used to obtain borehole exploration data of the modeling area and to discretize the modeling area into a unit network, and then use the unit grid with known geological properties through which the borehole channel passes as the reference unit. Alternatively, it can be used to discretize the modeling object into a cell network and use the cell mesh with known properties as the reference cell; Dataset building units are used to select target units and arbitrarily choose reference units to combine and construct a dataset containing orders of magnitude of samples. Each sample data is specific to the target unit. N Three-dimensional coordinates of each reference unit x i , y i , z i and geological properties L i Three-dimensional coordinates of the target element x label , y label , z label The sample feature matrix is ​​formed by combining the features, and the geological attributes of the target unit are included. L label For labels; or for selecting target units and arbitrarily choosing reference units to construct a dataset containing orders of magnitude of samples, where each sample is for the target unit. N Three-dimensional coordinates of each reference unit x i , y i , z i and attributes l i Three-dimensional coordinates of the target element x label , y label , z label The sample feature matrix is ​​formed by combining the attributes of the target unit. l label For tags; The deep feature extraction unit is used to extract deep features from the sample feature matrix using an autoencoder to obtain the deep feature map of the sample. The model training unit is used to construct a geological attribute prediction model or attribute prediction model based on discrete units of a deep convolutional neural network, and to train the model using the deep feature map and labels of the samples. Prediction unit, used by a trained geological attribute prediction model or attribute prediction model to predict the geological attributes or properties of a discrete cell grid of unknown attributes.

5. A computer device, characterized in that: At least includes: One or more processors; A memory that stores one or more computer programs; The processor calls the computer program to implement: The steps of a deep learning-based three-dimensional geological modeling method as described in any one of claims 1-3.

6. A computer storage device, characterized in that: A computer program is stored, which is invoked by a processor to implement the steps of the deep learning-based three-dimensional geological modeling method according to any one of claims 1-3.

Images