WO2021179558A1 - 一种构建数字岩心的方法及系统 - Google Patents

一种构建数字岩心的方法及系统 Download PDF

Info

Publication number
WO2021179558A1
WO2021179558A1 PCT/CN2020/114481 CN2020114481W WO2021179558A1 WO 2021179558 A1 WO2021179558 A1 WO 2021179558A1 CN 2020114481 W CN2020114481 W CN 2020114481W WO 2021179558 A1 WO2021179558 A1 WO 2021179558A1
Authority
WO
WIPO (PCT)
Prior art keywords
discriminant
network
digital core
image
loss function
Prior art date
Application number
PCT/CN2020/114481
Other languages
English (en)
French (fr)
Inventor
杨永飞
刘夫贵
姚军
汪远博
宋怀森
徐伯钊
张凯
张磊
孙海
宋文辉
Original Assignee
中国石油大学(华东)
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中国石油大学(华东) filed Critical 中国石油大学(华东)
Priority to US17/295,493 priority Critical patent/US11934488B2/en
Publication of WO2021179558A1 publication Critical patent/WO2021179558A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2148—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the process organisation or structure, e.g. boosting cascade
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/088—Non-supervised learning, e.g. competitive learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/143—Segmentation; Edge detection involving probabilistic approaches, e.g. Markov random field [MRF] modelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/95—Pattern authentication; Markers therefor; Forgery detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10056—Microscopic image
    • G06T2207/10061—Microscopic image from scanning electron microscope
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20076—Probabilistic image processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20112—Image segmentation details
    • G06T2207/20152—Watershed segmentation

Definitions

  • the invention relates to the technical field of digital cores, in particular to a method and system for constructing digital cores.
  • digital core technology can reproduce complex pore spaces and establish a model that effectively characterizes the complex pore structure of porous media, allowing core samples to be visualized and quantified. Furthermore, the digital core technology can perform flow simulation in the reconstructed digital core and pore network model.
  • Digital core technology It has become one of the necessary technologies and methods for unconventional oil and gas development. However, due to shale, carbonate rock and some deep formations, there are great difficulties in obtaining cores. There are problems such as difficulty in obtaining cores and high cost of obtaining cores. The core samples obtained are also very precious. Therefore, with the support of existing experimental instruments and theories, the existing methods of constructing digital cores have the problems of high cost and time-consuming to obtain high-resolution sample images.
  • the purpose of the present invention is to provide a method and system for constructing a digital core, so as to solve the problems of high cost and time-consuming acquisition of high-resolution sample images by the existing method for constructing a digital core.
  • the present invention provides the following solutions:
  • a method of constructing digital cores includes:
  • the digital core training image is a digital core sample image of a rock whose core is known;
  • the digital core construction model is a generated confrontation network trained through the sample set and the random sample noise, and the digital core construction The model is used to construct the target digital core image;
  • the target random noise is input into the digital core construction model to obtain the target digital core image.
  • the acquiring a digital core training image specifically includes:
  • the watershed segmentation method is used to segment the smooth digital core image to obtain a digital core training image.
  • the training and generating a confrontation network using the sample set and random sample noise to obtain a digital core construction model specifically includes:
  • the generative countermeasure network includes a generative network and a discriminant network
  • the pseudo sample set includes a plurality of first pseudo-digital core images
  • the discriminant network model is a trained discriminant network
  • the input of the discriminant network model is the first pseudo-digital core image
  • the output Is the true and false probability value of the first pseudo-digital core image
  • the generation network model is a trained generation network
  • the output of the generation network model is the The target digital core image
  • the discriminant network model and the generation network model constitute a digital core construction model.
  • the discriminating network includes: discriminating the input layer, discriminating the intermediate layer, and discriminating the output layer;
  • the generating network includes: generating an input layer, generating an intermediate layer, and generating an output layer; the generating input layer is a fully connected layer, and the generating intermediate layer and the generating output layer are both micro-step convolutional layers;
  • the activation function includes: the activation function of the judgment middle layer, the activation function of the judgment output layer, the activation function of the middle layer and the activation function of the output layer; the activation function of the judgment middle layer is the LeakyReLu activation function, and the judgment
  • the activation function of the output layer is a Sigmoid activation function
  • the activation function of the generating intermediate layer is a ReLu activation function
  • the activation function of the generating output layer is a Tanh activation function
  • the loss function includes: a discriminant loss function of a discriminating network and a generation loss function of a generating network.
  • the training of the discriminant network using the pseudo sample set and the sample set to obtain a discriminant network model specifically includes:
  • the step of taking the random sample noise as input and using the discriminant network model to train the generation network to obtain the generation network model specifically includes:
  • Loss_S1 lg(D(G(z, ⁇ ), ⁇ )); in the above formula, Loss_S1 represents the first loss function , D( ⁇ ) represents the discriminant network model, G( ⁇ ) represents the generation network, z represents the random sample noise, ⁇ represents the discriminant network parameter, and ⁇ represents the generation network parameter;
  • the true and false probability value of the network model is a preset true and false probability value, and the corresponding generation network parameter when the preset number of iterations is reached or when the true and false probability value is the preset true and false probability value is determined as the optimal generation network parameter,
  • a generation network model is obtained according to the optimal generation network parameters; the generation network parameters are the weights and biases of each layer of the generation network.
  • a system for building digital cores including:
  • the digital core training image module is used to obtain a digital core training image;
  • the digital core training image is a digital core sample image of a rock whose core is known;
  • the sample set module is used to divide the digital core training image into a plurality of sub-samples, and store all the sub-samples as a sample set;
  • the digital core model module is used to train and generate a confrontation network using the sample set and random sample noise to obtain a digital core construction model;
  • the digital core construction model is a generated confrontation trained by the sample set and the random sample noise Network, the digital core construction model is used to construct a target digital core image;
  • the acquisition module is used to acquire target random noise
  • the target digital core image module is used to input the target random noise into the digital core construction model to obtain the target digital core image.
  • the digital core training image module specifically includes:
  • Scanning unit used to scan the known rock in the core using image scanning technology to obtain the gray-scale image of the known rock in the core;
  • a smoothing processing unit for extracting a representative unit body of a known rock gray image center position of the core, and smoothing the representative unit body to obtain a smooth digital core image
  • the segmentation unit is used to segment the smooth digital core image by adopting a watershed segmentation method to obtain a digital core training image.
  • the digital core model module specifically includes:
  • the acquiring unit is used to acquire the activation function and the loss function of the generative confrontation network;
  • the generative confrontation network includes a generation network and a discriminant network;
  • the pseudo sample set unit is used to input random sample noise into the generating network to obtain a pseudo sample set;
  • the pseudo sample set includes a plurality of first pseudo-digital core images;
  • the discriminant network model unit is used to train the discriminant network using the pseudo sample set and the sample set to obtain a discriminant network model;
  • the discriminant network model is a trained discriminant network, and the input of the discriminant network model is the first A pseudo-digital core image, output as the true and false probability value of the first pseudo-digital core image;
  • the digital core construction model unit is used to take the random sample noise as input and use the discriminant network model to train the generation network to obtain a generation network model; the generation network model is a trained generation network, and The output of the generated network model is the target digital core image; the discriminant network model and the generated network model constitute a digital core construction model.
  • the discriminating network model unit specifically includes:
  • the first discriminant loss function subunit is configured to extract N sub-samples from the sample set and input them into the discriminant network to calculate the first discriminant loss function;
  • the first discriminant gradient subunit is used to calculate the first discriminant gradient of each layer of the discriminant network by using the first discriminant loss function
  • the second discriminant loss function subunit is used to extract N of the first pseudo-digital core images from the pseudo sample set and input them into the discriminant network to calculate a second discriminant loss function
  • the second discriminant gradient subunit is used to calculate the second discriminant gradient of each layer of the discriminant network by using the second discriminant loss function
  • a discriminant loss function subunit configured to add the first discriminant loss function and the second discriminant loss function to obtain the discriminant loss function
  • the discriminant network model subunit is used to optimize the discriminant loss function by using the first discriminant gradient, the second discriminant gradient and the mini-batch gradient descent algorithm to obtain optimal discriminant network parameters, and to discriminate network parameters according to the optimal Obtain the discriminant network model; the discriminant network parameters are the weights and biases of each layer of the discriminant network.
  • the present invention discloses the following technical effects:
  • the invention provides a method and system for constructing a digital core.
  • the method includes: acquiring a digital core training image; the digital core training image is a digital core sample image of a rock whose core is known; dividing the digital core training image into multiple sub-samples, and storing all the sub-samples as a sample set; using the sample set and Random sample noise training generates a confrontation network to obtain a digital core construction model; the digital core construction model is a generated confrontation network trained through sample sets and random sample noise, and the digital core construction model is used to construct the target digital core image; obtain the target random noise; The random noise of the target is input into the digital core to construct the model, and the target digital core image is obtained.
  • This method uses the Generative Adversarial Network (GAN) to build the target digital core image, and stores the trained digital core construction model used to construct the target digital core image.
  • GAN Generative Adversarial Network
  • the stored digital core construction model can be used directly to quickly build the target
  • the digital core image greatly reduces the calculation cost, and reduces the cost and time-consuming of obtaining high-resolution sample images.
  • Fig. 1 is a flowchart of a method for constructing a digital core provided by an embodiment of the present invention
  • Figure 2 is a two-dimensional grayscale image of a rock provided by an embodiment of the present invention
  • Figure 2(a) is a two-dimensional grayscale image of the rock
  • Figure 2(b) is a three-dimensional grayscale image of the rock
  • Figure 3 is a representation unit volume diagram provided by an embodiment of the present invention
  • Figure 3(a) is a representation unit volume diagram extracted from 400*400*400 pixels
  • Figure 3(b) is a three-dimensional display diagram of the extracted representation unit volume ;
  • Fig. 4 is a two-dimensional display diagram provided by an embodiment of the present invention
  • Fig. 4(a) is a two-dimensional display diagram of the extracted characterization unit
  • Fig. 4(b) is a two-dimensional display of the smoothed characterization unit picture
  • FIG. 5 is a two-dimensional display diagram of a segmented representation unit provided by an embodiment of the present invention.
  • Fig. 6 is a target digital core image provided by an embodiment of the present invention
  • Fig. 6(a) is a three-dimensional display diagram of a constructed 400*400*400 pixel target digital core image
  • Fig. 6(b) is a constructed 400*400* A two-dimensional display image of a 400-pixel target digital core image
  • Fig. 7 is a structural diagram of a system for constructing a digital core provided by an embodiment of the present invention.
  • the purpose of the present invention is to provide a method and system for constructing a digital core, so as to solve the problems of high cost and time-consuming acquisition of high-resolution sample images by the existing method for constructing a digital core.
  • FIG. 1 is a flowchart of the method for constructing a digital core provided by an embodiment of the present invention. Referring to Figure 1, the method of constructing a digital core includes:
  • Step 101 Obtain a digital core training image; the digital core training image is a digital core sample image of a rock whose core is known.
  • Step 101 specifically includes:
  • the image scanning technology is used to scan the known rock in the core, and the gray-scale image of the known rock in the core is obtained.
  • Image scanning technology includes the use of CT scanning equipment and Focused Ion Beam Scanning Electron Microscope (FIB-SEM) and other equipment for image scanning.
  • the scanning resolution of this embodiment is p micrometers ( ⁇ m)
  • the rock grayscale image includes: two-dimensional rock gray For the three-dimensional grayscale image and the rock, see Figure 2(a) and Figure 2(b).
  • the representative unit body of the known rock gray image center position of the core is extracted, and the representative unit body is smoothed to obtain a smooth digital core image.
  • This step is specifically: extracting a 400*400*400 pixel representation unit volume in the middle of the rock 3D gray image to improve the calculation speed of the subsequent simulation.
  • the extracted 400*400*400 pixel representation unit volume is shown in Figure 3 (a).
  • Figure 3(b) for the three-dimensional display of the extracted representative unit, and refer to Figure 4(a) for the two-dimensional display of the extracted representative unit.
  • the non-local average method is used to smooth the characterization unit body to improve the color contrast between the rock pores and the contact edge of the matrix, so that the rock pores and skeleton can be more clearly distinguished in the next step.
  • Figure 4(b) for the two-dimensional display of the smoothed characterization unit.
  • Figure 2 (a), Figure 2 (b), Figure 3 (a), Figure 3 (b), Figure 4 (a) and Figure 4 (b) are grayscale images.
  • the watershed segmentation method is used to segment the smooth digital core image to obtain the digital core training image.
  • the threshold-based watershed segmentation method is used to segment the pores and the matrix according to the gray value of the matrix.
  • the two-dimensional display diagram of the characterization unit after segmentation is shown in Figure 5.
  • the black is the pore and the white is the rock matrix.
  • 1mm in Fig. 2 to Fig. 5 represents the unit size of the image. Save the digital core training image as 3Dtif format.
  • Step 102 Segment the digital core training image into multiple sub-samples, and store all the sub-samples as a sample set.
  • Step 102 specifically includes: In order to ensure that each digital core training image contains several complete particles, set the step size to 16 pixels and the sub-sample size to 64*64*64 pixels, and to train the digital core saved in a .tif format file. The image is segmented once per step to generate 10648 sub-samples; 10648 sub-samples are stored as a sample set and saved in the .hdf5 file format, which is convenient for reading the sample set during training in step 103.
  • Step 103 Use the sample set and random sample noise to train a generated confrontation network to obtain a digital core construction model; the digital core construction model is a generated confrontation network trained through the sample set and random sample noise, and the digital core construction model is used to construct the target digital core image.
  • Step 103 specifically includes:
  • the generative confrontation network includes the generation network and the discriminant network.
  • the discriminant network includes: discriminating the input layer, discriminating the intermediate layer and discriminating the output layer.
  • the generating network includes: generating an input layer, generating an intermediate layer, and generating an output layer; generating an input layer as a fully connected layer, generating an intermediate layer as a three-layer micro-step convolutional layer, and generating an output layer as a micro-step convolutional layer.
  • the input of the generating network is random sample noise z, and the output is image data. It is judged that the output value of the output layer is in the (0,1) interval, as a two-classifier.
  • the input of the discriminant network is an image, and the output is the possibility that the input image is a real image, that is, the probability value of true and false, and it is also the output value of the discriminating output layer.
  • the activation function includes: discriminating the activation function of the intermediate layer, discriminating the activation function of the output layer, generating the activation function of the intermediate layer, and generating the activation function of the output layer.
  • the activation function of the middle layer is judged to be the LeakyReLu activation function
  • the activation function of the output layer is judged to be the Sigmoid activation function
  • the activation function of the generated middle layer is the ReLu activation function
  • the activation function of the generated output layer is the Tanh activation function.
  • the loss function includes: the discriminant loss function of the discriminant network and the generation loss function of the generative network.
  • the discriminant loss function of the discriminant network is:
  • Loss_D lg(D(x, ⁇ ))+lg(1-D(G(z, ⁇ ), ⁇ )) (1)
  • Loss_D represents the discriminative loss function
  • D( ⁇ ) represents the discriminant network model
  • x represents the sub-sample
  • ⁇ represents the discriminant network parameters
  • G( ⁇ ) represents the generation network
  • z represents the random sample noise
  • ⁇ represents the generation network parameters.
  • the discriminant network parameter is the discriminant gradient of each layer of the discriminating network
  • the generated network parameter is the gradient of each layer of the generated network.
  • the generative loss function of the generative network is:
  • Loss_G represents the generative loss function
  • G( ⁇ ) represents the generative network
  • the random sample noise is input to the generation network to obtain a pseudo sample set; the pseudo sample set includes a plurality of first pseudo-digital core images. Specifically: first initialize the generation network and fix the parameters of the generation network; then input random sample noise into the initialized generation network to generate multiple first pseudo-digital core images; finally store multiple first pseudo-digital core images as pseudo Sample set.
  • the random sample noise in this embodiment conforms to the (0, 1) standard normal distribution.
  • the discriminant network model is a trained discriminant network
  • the input of the discriminant network model is the first pseudo-digital core image
  • the output is the true or false of the first pseudo-digital core image Probability value. This step is to improve the ability of the discrimination network to identify true and false images.
  • the sub-samples in the sample set are true images
  • the first pseudo-digital core image in the pseudo-sample set is a fake image.
  • the first discriminant loss function is used to calculate the first discriminant gradient of each layer of the discriminant network.
  • Calculating the discriminative gradient of each layer of the discriminant network includes: first calculating the net input of each layer And activation value Use the net input and activation value of the last layer to calculate the error term of the last layer; then use the error term of the last layer (ie the first-to-last layer) to calculate the error term of the second-to-last layer, back propagation calculation from back to front From the error term of the previous layer to the first layer, the error term of each layer is obtained.
  • the error term is used to calculate the weight and bias of each layer of the discriminant network to obtain the discriminant network parameters; finally, the first discriminant loss function is calculated for the discriminant network parameters Partial derivative, get the first discriminant gradient of each layer.
  • lr represents the initial learning rate; Indicates the weight of the first layer of the judgment network; Represents the error term of the first layer of the discriminant network, that is, the influence of the neurons in the first layer on the final loss; Indicates the activation value of the l-1th layer of the discrimination network, that is, the result of the net input after the activation function is applied; Represents the bias of the first layer of the discrimination network; T represents the transposition; l represents the number of layers of the discrimination network, which includes the discrimination input layer, the discrimination intermediate layer and the discrimination output layer.
  • Loss_2 lg(1-D(G(z, ⁇ ))) (7)
  • the second discriminant loss function is used to calculate the second discriminant gradient of each layer of the discriminant network.
  • the specific calculation method is the same as the above step "Using the first discriminant loss function to calculate the first discriminant gradient of each layer of the discriminant network".
  • the first discriminant loss function and the second discriminant loss function are added to obtain the discriminant loss function.
  • the parameters of the corresponding discriminant network are determined as the optimal discriminant network parameters, the optimal discriminant network parameters are saved, and the corresponding discriminant network when the discriminant loss function is maximized is determined as the discriminant network model.
  • the discriminant network model is the trained generative network
  • the output of the generative network model is the target digital core image
  • the discriminant network model and the generative network model Compose a digital core construction model.
  • the discriminant network model to train the generative network to obtain the generative network model, which specifically includes:
  • the random sample noise is input to the generating network, and the first pseudo sample is generated.
  • a random sample noise z of 512*1*1*1 pixels is specifically input to the generating network, and a first pseudo sample of 1*64*64*64 pixels is generated.
  • Loss_S1 lg(D(G(z, ⁇ ), ⁇ )) (8)
  • D( ⁇ ) represents the discriminant network model
  • G( ⁇ ) represents the generation network
  • z represents the random sample noise
  • ⁇ represents the generation network parameters
  • ⁇ represents the discrimination network parameters
  • the first loss function is used to calculate the generated gradient of each layer of the generated network.
  • Calculating the generation gradient of each layer of the generative network includes: first calculating the net input of each layer And activation value Use the net input and activation value of the last layer to calculate the error term of the last layer; then use the error term of the last layer (ie the first-to-last layer) to calculate the error term of the second-to-last layer, back propagation calculation from back to front From the error term of the previous layer to the first layer, the error term of each layer is obtained, and the weight and bias of each layer of the generated network are calculated by using the error term to obtain the generated network parameters; finally, the generated loss function (or first loss function) is calculated For the partial derivative of the generated network parameters, the generated gradient of each layer is obtained. Calculate and generate network parameters according to formula (9), calculate the weight of each layer according to formula (10), and calculate the bias of each layer according to formula (11):
  • lr represents the initial learning rate; Indicates the weight of the i-th layer of the generated network; Represents the error term of the i-th layer of the generated network, that is, the influence of the neurons in the i-th layer on the final loss; Indicates the activation value of the i-1th layer of the generated network, that is, the result of the net input after the activation function is applied; Represents the bias of the i-th layer of the generated network; T represents the transpose; i represents the number of layers of the generated network.
  • the number of layers of the generated network includes the generation of the input layer, the generation of the intermediate layer, and the generation of the output layer.
  • the network parameters are the weights and biases of each layer of the generated network.
  • the preset true and false probability value is close to or equal to 0.5; when the preset number of iterations is reached or the true and false probability value is the preset true and false probability value, the corresponding generation network parameters are determined For the optimal generation network parameters, the optimal generation network parameters are saved, and the generation network corresponding to the optimal generation network parameters is determined as the generation network model. When the number of iterations is reached, but the true and false probability value is not equal to the preset true and false probability value, the number of iterations is increased; when the true and false probability value is equal to the preset true and false probability value but the number of iterations is not reached, the iteration is stopped early.
  • the optimal generated network parameters are saved in the file format of .pth. Since a training process is performed by selecting N sample data, the loss function should be averaged when calculating the loss function. Therefore, all the loss functions in the above steps are average loss functions.
  • Step 104 Obtain target random noise.
  • Step 105 Input the target random noise into the digital core construction model to obtain the target digital core image.
  • Step 105 specifically includes: firstly setting the size of the target digital core image to be constructed, that is, setting the size of ngf; then inputting the target random noise into the generation network model of the digital core construction model to obtain the target digital core image.
  • Fig. 6 is a 400*400*400 pixel target digital core image constructed according to this embodiment. The three-dimensional display of the target digital core image is shown in Fig. 6(a), and the two-dimensional display is shown in Fig. 6(b).
  • FIG. 7 is a structural diagram of the system for constructing a digital core provided by an embodiment of the present invention.
  • the system for constructing a digital core includes:
  • the digital core training image module 201 is used to obtain digital core training images; the digital core training images are digital core sample images of rocks whose cores are known.
  • the digital core training image module 201 specifically includes:
  • the scanning unit is used to scan the known rock in the core using image scanning technology to obtain the gray-scale image of the known rock in the core.
  • Image scanning technology includes the use of CT scanning equipment and Focused Ion Beam Scanning Electron Microscope (FIB-SEM) and other equipment for image scanning.
  • FIB-SEM Focused Ion Beam Scanning Electron Microscope
  • the scanning resolution of this embodiment is p micrometers ( ⁇ m)
  • the rock grayscale image includes: two-dimensional rock gray Degree images and 3D grayscale images of rocks.
  • the smoothing processing unit is used to extract the representative unit body of the known rock gray image center position of the core, and perform smoothing processing on the representative unit body to obtain a smooth digital core image.
  • the segmentation unit is used to segment the smooth digital core image by adopting the watershed segmentation method to obtain the digital core training image.
  • the sample set module 202 is used to divide the digital core training image into multiple sub-samples, and store all the sub-samples as a sample set.
  • the digital core model module 203 is used to train and generate a confrontation network using sample sets and random sample noise to obtain a digital core construction model; the digital core construction model is a generated confrontation network trained through the sample set and random sample noise, and the digital core construction model is used To construct the target digital core image.
  • the digital core model module 203 specifically includes:
  • the acquiring unit is used to acquire the activation function and the loss function of the generated confrontation network; the generated confrontation network includes a generation network and a discriminant network.
  • the discriminant network includes: discriminating the input layer, discriminating the intermediate layer and discriminating the output layer.
  • the generating network includes: generating an input layer, generating an intermediate layer, and generating an output layer; generating an input layer as a fully connected layer, generating an intermediate layer as a three-layer micro-step convolutional layer, and generating an output layer as a micro-step convolutional layer.
  • the input of the generating network is random sample noise z, and the output is image data. It is judged that the output value of the output layer is in the (0,1) interval, as a two-classifier.
  • the input of the discriminant network is an image
  • the output is the possibility that the input image is a real image, that is, the probability value of true and false, and it is also the output value of the discriminating output layer.
  • the activation function includes: discriminating the activation function of the intermediate layer, discriminating the activation function of the output layer, generating the activation function of the intermediate layer, and generating the activation function of the output layer.
  • the activation function of the intermediate layer is judged as the LeakyReLu activation function
  • the activation function of the output layer is judged as the Sigmoid activation function
  • the activation function of the intermediate layer is generated as the ReLu activation function
  • the activation function of the output layer is generated as the Tanh activation function.
  • the loss function includes: the discriminant loss function of the discriminant network and the generation loss function of the generative network.
  • the pseudo sample set unit is used to input random sample noise into the generating network to obtain a pseudo sample set; the pseudo sample set includes a plurality of first pseudo-digital core images.
  • the pseudo sample set unit is specifically used to: first initialize the generation network and fix the parameters of the generation network; then input random sample noise into the initialized generation network to generate multiple first pseudo-digital core images; and finally set multiple first pseudo-digital core images
  • the core image is stored as a pseudo sample set.
  • the random sample noise in this embodiment conforms to the (0, 1) standard normal distribution.
  • the discriminant network model unit is used to train the discriminant network using the pseudo sample set and sample set to obtain the discriminant network model;
  • the discriminant network model is a trained discriminant network, the input of the discriminant network model is the first pseudo-digital core image, and the output is the first The true and false probability value of the pseudo-digital core image.
  • the discriminant network model unit specifically includes:
  • the first discriminant loss function subunit is used to extract N sub-samples from the sample set and input into the discriminant network to calculate the first discriminant loss function.
  • N 128, and N is generally a power of 2, which can improve calculation efficiency.
  • the first discriminant gradient subunit is used to calculate the first discriminant gradient of each layer of the discriminant network by using the first discriminant loss function.
  • the second discriminant loss function subunit is used to extract N first pseudo-digital core images from the pseudo sample set and input it into the discriminant network to calculate the second discriminant loss function.
  • the second discriminant gradient subunit is used to calculate the second discriminant gradient of each layer of the discriminant network by using the second discriminant loss function.
  • the discriminant loss function subunit is used to add the first discriminant loss function and the second discriminant loss function to obtain the discriminant loss function.
  • the discriminant network model subunit is used to optimize the discriminant loss function by using the first discriminant gradient, the second discriminant gradient and the mini-batch gradient descent algorithm to obtain the optimal discriminant network parameters, and obtain the discriminant network model according to the optimal discriminant network parameters; discriminate network parameters To determine the weight and bias of each layer of the network.
  • the digital core construction model unit is used to take random sample noise as input, and use the discriminant network model to train the generation network to obtain the generation network model; the generation network model is the trained generation network, and the output of the generation network model is the target digital core image ;
  • the discriminant network model and the generated network model constitute a digital core construction model.
  • the digital core construction model unit specifically includes:
  • the first pseudo sample subunit is used to input random sample noise into the generating network to generate the first pseudo sample.
  • the first loss function subunit is used to input the first pseudo sample into the discriminant network model, and calculate the first loss function Loss_S1 according to formula (8):
  • Loss_S1 lg(D(G(z, ⁇ ), ⁇ )) (8)
  • D( ⁇ ) represents the discriminative network model
  • G( ⁇ ) represents the generation network
  • z represents the random sample noise
  • ⁇ represents the discriminant network parameters
  • ⁇ represents the generation network parameters.
  • the gradient generation subunit is used to calculate the gradient of each layer of the generated network using the first loss function.
  • the network parameters are the weights and biases of each layer of the generated network.
  • the obtaining module 204 is used to obtain target random noise.
  • the target digital core image module 205 is used to input target random noise into the digital core construction model to obtain the target digital core image.
  • the method and system for constructing digital cores can establish digital core images of real porous media based on CT scanning technology, focused ion beam scanning electron microscope (FIB-SEM) and other image scanning technologies, which are compared with other numerical-based
  • the reconstruction method is more authentic and representative.
  • Digital core technology converts real porous media into a data volume that can be identified by a computer. On this basis, storage spaces of micro-fractures and other forms can be added, making it possible to simulate many microscopic seepage mechanisms.
  • the GAN model will save the hyperparameters, that is, the discriminant network parameter ⁇ and the generated network parameter ⁇ , and then when constructing the digital core image, only the size of the target random noise needs to be modified to quickly and massively generate the target digital core image.
  • the training samples are derived from the three-dimensional data volume (three-dimensional image) of real digital cores established by image scanning technologies such as CT scanning technology and focused ion beam scanning electron microscope (FIB-SEM). Due to CT scanning technology, focused ion beam Scanning electron microscope (FIB-SEM) and other image scanning technologies can establish digital cores of real porous media, which preserves the true nature and high precision of porous media. Therefore, the digital core images established based on the GAN model (digital core construction model) are true High performance and high precision.
  • image scanning technologies such as CT scanning technology and focused ion beam scanning electron microscope (FIB-SEM)
  • FIB-SEM focused ion beam Scanning electron microscope
  • the digital core images established based on the GAN model are true High performance and high precision.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Multimedia (AREA)
  • Geometry (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

一种构建数字岩心的方法及系统，涉及数字岩心技术领域。该方法利用图像扫描技术获取能够反映真实岩心信息的三维数字岩心图像，并对其进行处理得到数字岩心训练图像来训练生成对抗网络；将训练好的生成对抗网络存储起来，得到数字岩心构建模型，可以直接利用存储的数字岩心构建模型快速构建目标数字岩心图像，这不仅大大减小了计算成本，降低了获取高分辨率样本图像成本和耗时，同时，所构建的目标数字岩心图像也能反映真实的岩心信息。

Description

一种构建数字岩心的方法及系统
本申请要求于2020年03月13日提交中国专利局、申请号为202010175137.6、发明名称为“一种构建数字岩心的方法及系统”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及数字岩心技术领域，特别是涉及一种构建数字岩心的方法及系统。
背景技术
基于物理实验方法或者数值重建手段，数字岩心技术可以再现复杂的孔隙空间，建立有效表征多孔介质复杂孔隙结构特性的模型，使得岩心样品可视化和定量化。进而数字岩心技术可以在重构的数字岩心和孔隙网络模型中进行流动模拟，通过数字岩心技术可以获得油气田开发所必须的数据，为高效开发非常规油气资源提供理论指导和技术支持，数字岩心技术已成为目前非常规油气开发必备的技术和方法之一。但是由于页岩、碳酸盐岩以及一些深层地层在获取岩心时存在着很大的困难，存在着岩心不易获取和获取岩心成本高等问题，获取的岩心样本也非常珍贵。所以在已有实验仪器和理论的支持下，现有的构建数字岩心的方法存在获取高分辨率样本图像成本高和耗时长的问题。
发明内容
本发明的目的是提供一种构建数字岩心的方法及系统，以解决现有构建数字岩心的方法获取高分辨率样本图像成本高和耗时长的问题。
为实现所述目的，本发明提供了如下方案：
一种构建数字岩心的方法，包括：
获取数字岩心训练图像；所述数字岩心训练图像为岩心已知的岩石的数字岩心样本图像；
将所述数字岩心训练图像分割为多个子样本，将所有所述子样本存储为样本集；
利用所述样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型；所述数字岩心构建模型为通过所述样本集和所述随机样本噪声训练好的生成对抗网络，所述数字岩心构建模型用于构建目标数字岩心图像；
获取目标随机噪声；
将所述目标随机噪声输入所述数字岩心构建模型，得到目标数字岩心图像。
可选的，所述获取数字岩心训练图像，具体包括：
利用图像扫描技术对岩心已知的岩石进行扫描，得到岩心已知的岩石灰度图像；
提取所述岩心已知的岩石灰度图像中心位置的表征单元体，并对所述表征单元体进行平滑处理，得到平滑数字岩心图像；
采用分水岭分割方法对所述平滑数字岩心图像进行分割，得到数字岩心训练图像。
可选的，所述利用所述样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型，具体包括：
获取生成对抗网络的激活函数和损失函数；所述生成对抗网络包括生成网络和判别网络；
将随机样本噪声输入至所述生成网络，得到伪样本集；所述伪样本集包括多个第一伪数字岩心图像；
利用所述伪样本集和所述样本集训练所述判别网络，得到判别网络模型；所述判别网络模型为训练好的判别网络，所述判别网络模型的输入为第一伪数字岩心图像，输出为所述第一伪数字岩心图像的真假概率值；
将所述随机样本噪声作为输入，利用所述判别网络模型对所述生成网络进行训练，得到生成网络模型；所述生成网络模型为训练好的生成网络，所述生成网络模型的输出为所述目标数字岩心图像；所述判别网络模型和所述生成网络模型构成数字岩心构建模型。
可选的，所述判别网络包括：判别输入层、判别中间层和判别输出层；
所述生成网络包括：生成输入层、生成中间层和生成输出层；所述生成输入层为全连接层，所述生成中间层和所述生成输出层均为微步卷积层；
所述激活函数包括：判别中间层的激活函数、判别输出层的激活函数、生成中间层的激活函数和生成输出层的激活函数；所述判别中间层的激活函数为LeakyReLu激活函数，所述判别输出层的激活函数为Sigmoid激活函数，所述生成中间层的激活函数为ReLu激活函数，所述生成输出层的激活函数为Tanh激活函数；
所述损失函数包括：判别网络的判别损失函数和生成网络的生成损失函数。
可选的，所述利用所述伪样本集和所述样本集训练所述判别网络，得到判别网络模型，具体包括：
从所述样本集中抽取N个子样本输入所述判别网络，计算第一判别损失函数；
利用所述第一判别损失函数计算所述判别网络每层的第一判别梯度；
从所述伪样本集中抽取N个所述第一伪数字岩心图像输入所述判别网络，计算第二判别损失函数；
利用所述第二判别损失函数计算所述判别网络每层的第二判别梯度；
将所述第一判别损失函数与所述第二判别损失函数相加得到所述判别损失函数；
利用所述第一判别梯度、所述第二判别梯度和小批量梯度下降算法优化所述判别损失函数，得到最优判别网络参数，根据所述最优判别网络参数得到判别网络模型；判别网络参数为所述判别网络每层的权重和偏置。
可选的，所述将所述随机样本噪声作为输入，利用所述判别网络模型对所述生成网络进行训练，得到生成网络模型，具体包括：
将所述随机样本噪声输入所述生成网络，生成第一伪样本；
将所述第一伪样本输入所述判别网络模型，根据公式Loss_S1＝lg(D(G(z,θ),α))计算第一损失函数；上式中，Loss_S1表示所述第一损失函数，D(·)表示所述判别网络模型，G(·)表示所述生成网络，z表示所述随机样本噪声，α表示所述判别网络参数，θ表示生成网络参数；
利用所述第一损失函数计算所述生成网络每层的生成梯度；
利用所述生成梯度和小批量梯度下降算法优化所述生成损失函数，返回迭代“将所述随机样本噪声输入所述生成网络，生成第一伪样本”，直至达到预设迭代次数或所述判别网络模型的真假概率值为预设真假概率值，将达到预设迭代次数时或所述真假概率值为预设真假概率值时对应的生成网络参数确定为最优生成网络参数，根据所述最优生成网络参数得到 生成网络模型；所述生成网络参数为所述生成网络每层的权重和偏置。
一种构建数字岩心的系统，包括：
数字岩心训练图像模块，用于获取数字岩心训练图像；所述数字岩心训练图像为岩心已知的岩石的数字岩心样本图像；
样本集模块，用于将所述数字岩心训练图像分割为多个子样本，将所有所述子样本存储为样本集；
数字岩心模型模块，用于利用所述样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型；所述数字岩心构建模型为通过所述样本集和所述随机样本噪声训练好的生成对抗网络，所述数字岩心构建模型用于构建目标数字岩心图像；
获取模块，用于获取目标随机噪声；
目标数字岩心图像模块，用于将所述目标随机噪声输入所述数字岩心构建模型，得到目标数字岩心图像。
可选的，所述数字岩心训练图像模块，具体包括：
扫描单元，用于利用图像扫描技术对岩心已知的岩石进行扫描，得到岩心已知的岩石灰度图像；
平滑处理单元，用于提取所述岩心已知的岩石灰度图像中心位置的表征单元体，并对所述表征单元体进行平滑处理，得到平滑数字岩心图像；
分割单元，用于采用分水岭分割方法对所述平滑数字岩心图像进行分割，得到数字岩心训练图像。
可选的，所述数字岩心模型模块，具体包括：
获取单元，用于获取生成对抗网络的激活函数和损失函数；所述生成 对抗网络包括生成网络和判别网络；
伪样本集单元，用于将随机样本噪声输入至所述生成网络，得到伪样本集；所述伪样本集包括多个第一伪数字岩心图像；
判别网络模型单元，用于利用所述伪样本集和所述样本集训练所述判别网络，得到判别网络模型；所述判别网络模型为训练好的判别网络，所述判别网络模型的输入为第一伪数字岩心图像，输出为所述第一伪数字岩心图像的真假概率值；
数字岩心构建模型单元，用于将所述随机样本噪声作为输入，利用所述判别网络模型对所述生成网络进行训练，得到生成网络模型；所述生成网络模型为训练好的生成网络，所述生成网络模型的输出为所述目标数字岩心图像；所述判别网络模型和所述生成网络模型构成数字岩心构建模型。
可选的，所述判别网络模型单元，具体包括：
第一判别损失函数子单元，用于从所述样本集中抽取N个子样本输入所述判别网络，计算第一判别损失函数；
第一判别梯度子单元，用于利用所述第一判别损失函数计算所述判别网络每层的第一判别梯度；
第二判别损失函数子单元，用于从所述伪样本集中抽取N个所述第一伪数字岩心图像输入所述判别网络，计算第二判别损失函数；
第二判别梯度子单元，用于利用所述第二判别损失函数计算所述判别网络每层的第二判别梯度；
判别损失函数子单元，用于将所述第一判别损失函数与所述第二判别 损失函数相加得到所述判别损失函数；
判别网络模型子单元，用于利用所述第一判别梯度、所述第二判别梯度和小批量梯度下降算法优化所述判别损失函数，得到最优判别网络参数，根据所述最优判别网络参数得到判别网络模型；判别网络参数为所述判别网络每层的权重和偏置。
根据本发明提供的具体实施例，本发明公开了以下技术效果：
本发明提供一种构建数字岩心的方法及系统。该方法包括：获取数字岩心训练图像；数字岩心训练图像为岩心已知的岩石的数字岩心样本图像；将数字岩心训练图像分割为多个子样本，将所有子样本存储为样本集；利用样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型；数字岩心构建模型为通过样本集和随机样本噪声训练好的生成对抗网络，数字岩心构建模型用于构建目标数字岩心图像；获取目标随机噪声；将目标随机噪声输入数字岩心构建模型，得到目标数字岩心图像。该方法利用生成对抗网络(Generative Adversarial Network，GAN)建立目标数字岩心图像，将训练后的用于构建目标数字岩心图像的数字岩心构建模型存储起来，可以直接利用存储的数字岩心构建模型快速构建目标数字岩心图像，大大减小了计算成本，降低了获取高分辨率样本图像成本和耗时。
结合CT扫描技术和聚焦离子束扫描电子显微镜(Focused Ion Beam Scanning Electron Microscope，FIB-SEM)等图像扫描技术，提高了构建的目标数字岩心图像的真实性与准确性，对于像页岩和深层油气藏这种难以取心的岩心样本实验而言，可以减少取心的成本，在实际应用中有很重要的价值。
说明书附图
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出 创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1为本发明实施例所提供的构建数字岩心的方法的流程图；
图2为本发明实施例所提供的岩石二维灰度图；图2(a)为岩石二维灰度图像；图2(b)为岩石三维灰度图像；
图3为本发明实施例所提供的表征单元体图；图3(a)为提取400*400*400像素的表征单元体图；图3(b)为提取后的表征单元体的三维展示图；
图4为本发明实施例所提供的二维展示图；图4(a)为提取后的表征单元体的二维展示图；图4(b)为平滑处理后的表征单元体的二维展示图；
图5为本发明实施例所提供的分割后的表征单元体的二维展示图；
图6为本发明实施例所提供的目标数字岩心图像；图6(a)为构建的400*400*400像素目标数字岩心图像的三维展示图；图6(b)为构建的400*400*400像素目标数字岩心图像的二维展示图；
图7为本发明实施例所提供的构建数字岩心的系统的结构图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
本发明的目的是提供一种构建数字岩心的方法及系统，以解决现有构建数字岩心的方法获取高分辨率样本图像成本高和耗时长的问题。
为使本发明的所述目的、特征和优点能够更加明显易懂，下面结合附图和具体实施方式对本发明作进一步详细的说明。
本实施例提供一种构建数字岩心的方法，图1为本发明实施例所提供的构建数字岩心的方法的流程图。参见图1，构建数字岩心的方法包括：
步骤101，获取数字岩心训练图像；数字岩心训练图像为岩心已知的岩石的数字岩心样本图像。
步骤101具体包括：
利用图像扫描技术对岩心已知的岩石进行扫描，得到岩心已知的岩石灰度图像。图像扫描技术包括利用CT扫描设备和聚焦离子束扫描电子显微镜(FIB-SEM)等设备进行图像扫描，本实施例的扫描分辨率为p微米(μm)，岩石灰度图像包括：岩石二维灰度图像和岩石三维灰度图像，参见图2(a)和图2(b)。
提取岩心已知的岩石灰度图像中心位置的表征单元体，并对表征单元体进行平滑处理，得到平滑数字岩心图像。此步骤具体为：在岩石三维灰度图像的中间位置提取400*400*400像素的表征单元体，以提高后期模拟的计算速度，提取的400*400*400像素的表征单元体图参见图3(a)，提取后的表征单元体的三维展示图参见图3(b)，提取后的表征单元体的二维展示图参见图4(a)。采用非局部均值方法对表征单元体进行平滑处理，以提高岩石孔隙和基质接触边缘的颜色对比，以便在下一步骤中更加清晰地区分岩石孔隙和骨架。平滑处理后的表征单元体的二维展示图参见图4(b)。图2(a)、图2(b)、图3(a)、图3(b)、图4(a)和图4(b)均为灰度图。
采用分水岭分割方法对平滑数字岩心图像进行分割，得到数字岩心训练图像。具体采用基于阈值的分水岭分割方法，根据孔隙与基质的灰度值进行分割，分割后的表征单元体的二维展示图参见图5，图5中黑色为孔隙，白色为岩石基质。图2-图5中的1mm表示图像的单位尺寸。将数字 岩心训练图像另存为3Dtif格式。
步骤102，将数字岩心训练图像分割为多个子样本，将所有子样本存储为样本集。步骤102具体包括：为确保每个数字岩心训练图像中包含几个完整的颗粒，设置步长为16像素，子样本尺寸为64*64*64像素，对以.tif格式文件保存的数字岩心训练图像每步长做一次分割，生成10648个子样本；将10648个子样本存储为样本集，并以.hdf5的文件格式保存，便于步骤103训练时读取样本集。
步骤103，利用样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型；数字岩心构建模型为通过样本集和随机样本噪声训练好的生成对抗网络，数字岩心构建模型用于构建目标数字岩心图像。
步骤103具体包括：
获取生成对抗网络的激活函数和损失函数；生成对抗网络包括生成网络和判别网络。判别网络包括：判别输入层、判别中间层和判别输出层。生成网络包括：生成输入层、生成中间层和生成输出层；生成输入层为一层全连接层，生成中间层为三层微步卷积层，生成输出层为一层微步卷积层。生成网络的输入是随机样本噪声z，输出是图像数据。判别输出层的输出值在(0，1)区间，作为二分类器。判别网络的输入是图像，输出是输入的图像为真实图像的可能性，即真假概率值，也是判别输出层的输出值。
激活函数包括：判别中间层的激活函数、判别输出层的激活函数、生成中间层的激活函数和生成输出层的激活函数。判别中间层的激活函数为LeakyReLu激活函数，判别输出层的激活函数为Sigmoid激活函数，生成 中间层的激活函数为ReLu激活函数，生成输出层的激活函数为Tanh激活函数。损失函数包括：判别网络的判别损失函数和生成网络的生成损失函数。判别网络的判别损失函数为：
Loss_D＝lg(D(x,α))+lg(1-D(G(z,θ),α))   (1)
上式中，Loss_D表示判别损失函数，D(·)表示判别网络模型，x表示子样本，α表示判别网络参数，G(·)表示生成网络，z表示随机样本噪声，θ表示生成网络参数。判别网络参数为判别网络每层的判别梯度，生成网络参数为生成网络每层的生成梯度。
生成网络的生成损失函数为：
Loss_G＝lg(1-D(G(z,θ)))    (2)
上式中，Loss_G表示生成损失函数，G(·)表示生成网络。
将随机样本噪声输入至生成网络，得到伪样本集；伪样本集包括多个第一伪数字岩心图像。具体为：首先初始化生成网络，且固定生成网络的参数；然后将随机样本噪声输入初始化后的生成网络，生成多个第一伪数字岩心图像；最后将多个第一伪数字岩心图像存储为伪样本集。本实施例中的随机样本噪声符合(0，1)标准正态分布。
利用伪样本集和样本集训练判别网络，得到判别网络模型；判别网络模型为训练好的判别网络，判别网络模型的输入为第一伪数字岩心图像，输出为第一伪数字岩心图像的真假概率值。这一步骤是为了提高判别网络识别真假图像的能力。样本集中的子样本为真图像，伪样本集中的第一伪数字岩心图像为假图像。
利用伪样本集和样本集训练判别网络，得到判别网络模型，具体包括：
从样本集中抽取N个子样本输入判别网络，计算第一判别损失函数。本实施例中N＝128，N一般为2的幂次方，可以提高计算效率，根据公式(3)计算第一判别损失函数Loss_1：
Loss_1＝lg(D(x,α))   (3)
利用第一判别损失函数计算判别网络每层的第一判别梯度。计算判别网络每层的判别梯度包括：首先计算每一层的净输入
Figure PCTCN2020114481-appb-000001
和激活值
Figure PCTCN2020114481-appb-000002
用最后一层的净输入和激活值计算最后一层的误差项；然后用最后一层(即倒数第一层)的误差项计算倒数第二层的误差项，从后往前反向传播计算前一层的误差项，直至第一层，得到每层的误差项，利用误差项计算判别网络每层的权重和偏置，得到判别网络参数；最后计算第一判别损失函数对判别网络参数的偏导数，得到每层的第一判别梯度。根据公式(4)计算和/或更新判别网络参数，具体根据公式(5)计算和/或更新每层的权重，根据公式(6)计算和/或更新每层的偏置：
Figure PCTCN2020114481-appb-000003
Figure PCTCN2020114481-appb-000004
Figure PCTCN2020114481-appb-000005
上式中，lr表示初始学习率；
Figure PCTCN2020114481-appb-000006
表示判别网络第l层的权重；
Figure PCTCN2020114481-appb-000007
表示判别网络第l层的误差项，即第l层的神经元对最终损失的影响；
Figure PCTCN2020114481-appb-000008
表示判别网络第l-1层的激活值，即净输入经过激活函数作用后的结果；
Figure PCTCN2020114481-appb-000009
表示判别网络第l层的偏置；T表示转置；l表示判别网络的层数，判别网络的层数包括判别输入层、判别中间层和判别输出层。
从伪样本集中抽取N个第一伪数字岩心图像输入判别网络，计算第二判别损失函数。根据公式(7)计算第二判别损失函数Loss_2：
Loss_2＝lg(1-D(G(z,θ)))   (7)
利用第二判别损失函数计算判别网络每层的第二判别梯度，具体计算方法与上述步骤“利用第一判别损失函数计算判别网络每层的第一判别梯度”同理。
将第一判别损失函数与第二判别损失函数相加得到判别损失函数。
利用第一判别梯度、第二判别梯度和小批量梯度下降算法优化判别损失函数，得到最优判别网络参数，根据最优判别网络参数得到判别网络模型；判别网络参数为判别网络每层的权重和偏置。此步骤具体包括：利用小批量梯度下降算法采用Adam优化器优化判别损失函数，设置Adam优化器的初始betas参数betas＝(0.5，0.999)和初始学习率lr＝0.00001，将判别损失函数最大化时对应的判别网络的参数确定为最优判别网络参数，保存最优判别网络参数，同时将判别损失函数最大化时对应的判别网络确定为判别网络模型。
将随机样本噪声作为输入，利用判别网络模型对生成网络进行训练，得到生成网络模型；生成网络模型为训练好的生成网络，生成网络模型的输出为目标数字岩心图像；判别网络模型和生成网络模型构成数字岩心构建模型。
将随机样本噪声作为输入，利用判别网络模型对生成网络进行训练，得到生成网络模型，具体包括：
将随机样本噪声输入生成网络，生成第一伪样本。本实施例中具体给 生成网络输入一个512*1*1*1像素的随机样本噪声z，生成一张1*64*64*64像素的第一伪样本。
将第一伪样本输入判别网络模型，根据公式(8)计算第一损失函数Loss_S1：
Loss_S1＝lg(D(G(z,θ),α))   (8)
上式中，D(·)表示判别网络模型，G(·)表示生成网络，z表示随机样本噪声，θ表示生成网络参数，α表示判别网络参数。
利用第一损失函数计算生成网络每层的生成梯度。计算生成网络每层的生成梯度包括：首先计算每一层的净输入
Figure PCTCN2020114481-appb-000010
和激活值
Figure PCTCN2020114481-appb-000011
用最后一层的净输入和激活值计算最后一层的误差项；然后用最后一层(即倒数第一层)的误差项计算倒数第二层的误差项，从后往前反向传播计算前一层的误差项，直至第一层，得到每层的误差项，利用误差项计算生成网络每层的权重和偏置，得到生成网络参数；最后计算生成损失函数(或第一损失函数)对生成网络参数的偏导数，得到每层的生成梯度。根据公式(9)计算生成网络参数，具体根据公式(10)计算每层的权重，根据公式(11)计算每层的偏置：
Figure PCTCN2020114481-appb-000012
Figure PCTCN2020114481-appb-000013
Figure PCTCN2020114481-appb-000014
上式中，lr表示初始学习率；
Figure PCTCN2020114481-appb-000015
表示生成网络第i层的权重；
Figure PCTCN2020114481-appb-000016
表示生成网络第i层的误差项，即第i层的神经元对最终损失的影响；
Figure PCTCN2020114481-appb-000017
表 示生成网络第i-1层的激活值，即净输入经过激活函数作用后的结果；
Figure PCTCN2020114481-appb-000018
表示生成网络第i层的偏置；T表示转置；i表示生成网络的层数，生成网络的层数包括生成输入层、生成中间层和生成输出层。
利用生成梯度和小批量梯度下降算法优化生成损失函数，返回迭代“将随机样本噪声输入生成网络，生成第一伪样本”，直至达到预设迭代次数或判别网络模型的真假概率值为预设真假概率值，将达到预设迭代次数时或真假概率值为预设真假概率值时对应的生成网络参数确定为最优生成网络参数，根据最优生成网络参数得到生成网络模型；生成网络参数为生成网络每层的权重和偏置。此步骤具体包括：利用小批量梯度下降算法采用Adam优化器优化生成损失函数，优化生成损失函数具体为最大化生成损失函数或最小化第一损失函数；设置Adam优化器的初始betas参数betas＝(0.5，0.999)和初始学习率lr＝0.00001；返回步骤“将随机样本噪声输入生成网络，生成第一伪样本”，进行迭代，直至达到预设迭代次数或判别网络模型的真假概率值为预设真假概率值，本实施例中预设真假概率值为接近或等于0.5的值；将达到预设迭代次数时或真假概率值为预设真假概率值时对应的生成网络参数确定为最优生成网络参数，保存最优生成网络参数，同时将最优生成网络参数对应的生成网络确定为生成网络模型。当达到迭代次数，但真假概率值不等于预设真假概率值时，增大迭代次数；当真假概率值等于预设真假概率值，但未达到迭代次数时，提前停止迭代。本实施例为便于后续需要构建目标数字岩心图像读取参数数据，将最优生成网络参数保存为.pth的文件格式。由于一次训练过程是选择N个样本数据进行的，所以在计算损失函数的时候要对损失函数进行平均， 因此上述步骤中的所有损失函数均为平均损失函数。
步骤104，获取目标随机噪声。
步骤105，将目标随机噪声输入数字岩心构建模型，得到目标数字岩心图像。步骤105具体包括：首先设置构建的目标数字岩心图像的大小，即设置ngf的大小；然后将目标随机噪声输入数字岩心构建模型的生成网络模型中，得到目标数字岩心图像。通过调整ngf的大小可构建不同尺寸的目标数字岩心图像。图6为根据本实施例构建的400*400*400像素的目标数字岩心图像，目标数字岩心图像的三维展示图参见图6(a)，二维展示图参见图6(b)。
本实施例提供一种构建数字岩心的系统，图7为本发明实施例所提供的构建数字岩心的系统的结构图。参见图7，构建数字岩心的系统包括：
数字岩心训练图像模块201，用于获取数字岩心训练图像；数字岩心训练图像为岩心已知的岩石的数字岩心样本图像。
数字岩心训练图像模块201具体包括：
扫描单元，用于利用图像扫描技术对岩心已知的岩石进行扫描，得到岩心已知的岩石灰度图像。图像扫描技术包括利用CT扫描设备和聚焦离子束扫描电子显微镜(FIB-SEM)等设备进行图像扫描，本实施例的扫描分辨率为p微米(μm)，岩石灰度图像包括：岩石二维灰度图像和岩石三维灰度图像。
平滑处理单元，用于提取岩心已知的岩石灰度图像中心位置的表征单元体，并对表征单元体进行平滑处理，得到平滑数字岩心图像。
分割单元，用于采用分水岭分割方法对平滑数字岩心图像进行分割， 得到数字岩心训练图像。
样本集模块202，用于将数字岩心训练图像分割为多个子样本，将所有子样本存储为样本集。
数字岩心模型模块203，用于利用样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型；数字岩心构建模型为通过样本集和随机样本噪声训练好的生成对抗网络，数字岩心构建模型用于构建目标数字岩心图像。
数字岩心模型模块203具体包括：
获取单元，用于获取生成对抗网络的激活函数和损失函数；生成对抗网络包括生成网络和判别网络。判别网络包括：判别输入层、判别中间层和判别输出层。生成网络包括：生成输入层、生成中间层和生成输出层；生成输入层为一层全连接层，生成中间层为三层微步卷积层，生成输出层为一层微步卷积层。生成网络的输入是随机样本噪声z，输出是图像数据。判别输出层的输出值在(0，1)区间，作为二分类器。判别网络的输入是图像，输出是输入的图像为真实图像的可能性，即真假概率值，也是判别输出层的输出值。激活函数包括：判别中间层的激活函数、判别输出层的激活函数、生成中间层的激活函数和生成输出层的激活函数。判别中间层的激活函数为LeakyReLu激活函数，判别输出层的激活函数为Sigmoid激活函数，生成中间层的激活函数为ReLu激活函数，生成输出层的激活函数为Tanh激活函数。损失函数包括：判别网络的判别损失函数和生成网络的生成损失函数。
伪样本集单元，用于将随机样本噪声输入至生成网络，得到伪样本集； 伪样本集包括多个第一伪数字岩心图像。伪样本集单元具体用于：首先初始化生成网络，且固定生成网络的参数；然后将随机样本噪声输入初始化后的生成网络，生成多个第一伪数字岩心图像；最后将多个第一伪数字岩心图像存储为伪样本集。本实施例中的随机样本噪声符合(0，1)标准正态分布。
判别网络模型单元，用于利用伪样本集和样本集训练判别网络，得到判别网络模型；判别网络模型为训练好的判别网络，判别网络模型的输入为第一伪数字岩心图像，输出为第一伪数字岩心图像的真假概率值。
判别网络模型单元具体包括：
第一判别损失函数子单元，用于从样本集中抽取N个子样本输入判别网络，计算第一判别损失函数。本实施例中N＝128，N一般为2的幂次方，可以提高计算效率。
第一判别梯度子单元，用于利用第一判别损失函数计算判别网络每层的第一判别梯度。
第二判别损失函数子单元，用于从伪样本集中抽取N个第一伪数字岩心图像输入判别网络，计算第二判别损失函数。
第二判别梯度子单元，用于利用第二判别损失函数计算判别网络每层的第二判别梯度。
判别损失函数子单元，用于将第一判别损失函数与第二判别损失函数相加得到判别损失函数。
判别网络模型子单元，用于利用第一判别梯度、第二判别梯度和小批量梯度下降算法优化判别损失函数，得到最优判别网络参数，根据最优判 别网络参数得到判别网络模型；判别网络参数为判别网络每层的权重和偏置。判别网络模型子单元具体用于：利用小批量梯度下降算法采用Adam优化器优化判别损失函数，设置Adam优化器的初始betas参数betas＝(0.5，0.999)和初始学习率lr＝0.00001，寻找更新后的判别损失函数最大化时对应的判别网络的参数，即最优判别网络参数，保存最优判别网络参数，同时将更新后的判别损失函数最大化时对应的判别网络确定为判别网络模型。
数字岩心构建模型单元，用于将随机样本噪声作为输入，利用判别网络模型对生成网络进行训练，得到生成网络模型；生成网络模型为训练好的生成网络，生成网络模型的输出为目标数字岩心图像；判别网络模型和生成网络模型构成数字岩心构建模型。
数字岩心构建模型单元具体包括：
第一伪样本子单元，用于将随机样本噪声输入生成网络，生成第一伪样本。
第一损失函数子单元，用于将第一伪样本输入判别网络模型，根据公式(8)计算第一损失函数Loss_S1：
Loss_S1＝lg(D(G(z,θ),α))   (8)
上式中，D(·)表示判别网络模型，G(·)表示生成网络，z表示随机样本噪声，α表示判别网络参数，θ表示生成网络参数。
生成梯度子单元，用于利用第一损失函数计算生成网络每层的生成梯度。
生成网络模型子单元，用于利用生成梯度和小批量梯度下降算法优化 生成损失函数，返回第一伪样本子单元进行迭代，直至达到预设迭代次数或判别网络模型的真假概率值为预设真假概率值，将达到预设迭代次数时或真假概率值为预设真假概率值时对应的生成网络参数确定为最优生成网络参数，根据最优生成网络参数得到生成网络模型；生成网络参数为生成网络每层的权重和偏置。
获取模块204，用于获取目标随机噪声。
目标数字岩心图像模块205，用于将目标随机噪声输入数字岩心构建模型，得到目标数字岩心图像。
本实施例提供的构建数字岩心的方法及系统，基于CT扫描技术、聚焦离子束扫描电子显微镜(FIB-SEM)等图像扫描技术可以建立真实多孔介质的数字岩心图像，相比于其他基于数值的重建方法，更具真实性和代表性。数字岩心技术是将真实多孔介质转化为计算机可以识别的数据体，可以在此基础上添加微裂缝等形态的储集空间，使很多微观渗流机理的模拟研究成为了可能。GAN模型在训练结束后会把超参数，即判别网络参数α和生成网络参数θ保存起来，之后在构建数字岩心图像时只需要修改目标随机噪声的大小就可以快速、大量地生成目标数字岩心图像，提高了效率，节省了计算时间成本。在训练GAN时，训练样本来源于CT扫描技术、聚焦离子束扫描电子显微镜(FIB-SEM)等图像扫描技术建立的真实数字岩心的三维数据体(三维图像)，由于CT扫描技术、聚焦离子束扫描电子显微镜(FIB-SEM)等图像扫描技术可以建立真实多孔介质的数字岩心，保存了多孔介质的真实性质和高精度，所以，基于GAN模型(数字岩心构建模型)建立的数字岩心图像的真实性高，精度高。
对于实施例公开的系统而言，由于其与实施例公开的方法相对应，所以描述的比较简单，相关之处参见方法部分说明即可。本文中应用了具体个例对本发明的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本发明的方法及其核心思想；同时，对于本领域的一般技术人员，依据本发明的思想，在具体实施方式及应用范围上均会有改变之处。综上 所述，本说明书内容不应理解为对本发明的限制。
提供以上实施例仅仅是为了描述本发明的目的，而并非要限制本发明的范围。本发明的范围由所附权利要求限定。不脱离本发明的精神和原理而做出的各种等同替换和修改，均应涵盖在本发明的范围之内。

Claims (10)

  1. 一种构建数字岩心的方法，其特征在于，包括：
    获取数字岩心训练图像；所述数字岩心训练图像为岩心已知的岩石的数字岩心样本图像；
    将所述数字岩心训练图像分割为多个子样本，将所有所述子样本存储为样本集；
    利用所述样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型；所述数字岩心构建模型为通过所述样本集和所述随机样本噪声训练好的生成对抗网络，所述数字岩心构建模型用于构建目标数字岩心图像；
    获取目标随机噪声；
    将所述目标随机噪声输入所述数字岩心构建模型，得到目标数字岩心图像。
  2. 根据权利要求1所述的构建数字岩心的方法，其特征在于，所述获取数字岩心训练图像，具体包括：
    利用图像扫描技术对岩心已知的岩石进行扫描，得到岩心已知的岩石灰度图像；
    提取所述岩心已知的岩石灰度图像中心位置的表征单元体，并对所述表征单元体进行平滑处理，得到平滑数字岩心图像；
    采用分水岭分割方法对所述平滑数字岩心图像进行分割，得到数字岩心训练图像。
  3. 根据权利要求2所述的构建数字岩心的方法，其特征在于，所述利用所述样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模 型，具体包括：
    获取生成对抗网络的激活函数和损失函数；所述生成对抗网络包括生成网络和判别网络；
    将随机样本噪声输入至所述生成网络，得到伪样本集；所述伪样本集包括多个第一伪数字岩心图像；
    利用所述伪样本集和所述样本集训练所述判别网络，得到判别网络模型；所述判别网络模型为训练好的判别网络，所述判别网络模型的输入为第一伪数字岩心图像，输出为所述第一伪数字岩心图像的真假概率值；
    将所述随机样本噪声作为输入，利用所述判别网络模型对所述生成网络进行训练，得到生成网络模型；所述生成网络模型为训练好的生成网络，所述生成网络模型的输出为所述目标数字岩心图像；所述判别网络模型和所述生成网络模型构成数字岩心构建模型。
  4. 根据权利要求3所述的构建数字岩心的方法，其特征在于，所述判别网络包括：判别输入层、判别中间层和判别输出层；
    所述生成网络包括：生成输入层、生成中间层和生成输出层；所述生成输入层为全连接层，所述生成中间层和所述生成输出层均为微步卷积层；
    所述激活函数包括：判别中间层的激活函数、判别输出层的激活函数、生成中间层的激活函数和生成输出层的激活函数；所述判别中间层的激活函数为LeakyReLu激活函数，所述判别输出层的激活函数为Sigmoid激活函数，所述生成中间层的激活函数为ReLu激活函数，所述生成输出层的激活函数为Tanh激活函数；
    所述损失函数包括：判别网络的判别损失函数和生成网络的生成损失函数。
  5. 根据权利要求4所述的构建数字岩心的方法，其特征在于，所述利用所述伪样本集和所述样本集训练所述判别网络，得到判别网络模型，具体包括：
    从所述样本集中抽取N个子样本输入所述判别网络，计算第一判别损失函数；
    利用所述第一判别损失函数计算所述判别网络每层的第一判别梯度；
    从所述伪样本集中抽取N个所述第一伪数字岩心图像输入所述判别网络，计算第二判别损失函数；
    利用所述第二判别损失函数计算所述判别网络每层的第二判别梯度；
    将所述第一判别损失函数与所述第二判别损失函数相加得到所述判别损失函数；
    利用所述第一判别梯度、所述第二判别梯度和小批量梯度下降算法优化所述判别损失函数，得到最优判别网络参数，根据所述最优判别网络参数得到判别网络模型；判别网络参数为所述判别网络每层的权重和偏置。
  6. 根据权利要求5所述的构建数字岩心的方法，其特征在于，所述将所述随机样本噪声作为输入，利用所述判别网络模型对所述生成网络进行训练，得到生成网络模型，具体包括：
    将所述随机样本噪声输入所述生成网络，生成第一伪样本；
    将所述第一伪样本输入所述判别网络模型，根据公式Loss_S1＝lg(D(G(z,θ),α))计算第一损失函数；上式中，Loss_S1表示所述第 一损失函数，D(.)表示所述判别网络模型，G(.)表示所述生成网络，z表示所述随机样本噪声，α表示所述判别网络参数，θ表示生成网络参数；
    利用所述第一损失函数计算所述生成网络每层的生成梯度；
    利用所述生成梯度和小批量梯度下降算法优化所述生成损失函数，返回迭代“将所述随机样本噪声输入所述生成网络，生成第一伪样本”，直至达到预设迭代次数或所述判别网络模型的真假概率值为预设真假概率值，将达到预设迭代次数时或所述真假概率值为预设真假概率值时对应的生成网络参数确定为最优生成网络参数，根据所述最优生成网络参数得到生成网络模型；所述生成网络参数为所述生成网络每层的权重和偏置。
  7. 一种构建数字岩心的系统，其特征在于，包括：
    数字岩心训练图像模块，用于获取数字岩心训练图像；所述数字岩心训练图像为岩心已知的岩石的数字岩心样本图像；
    样本集模块，用于将所述数字岩心训练图像分割为多个子样本，将所有所述子样本存储为样本集；
    数字岩心模型模块，用于利用所述样本集和随机样本噪声训练生成对抗网络，得到数字岩心构建模型；所述数字岩心构建模型为通过所述样本集和所述随机样本噪声训练好的生成对抗网络，所述数字岩心构建模型用于构建目标数字岩心图像；
    获取模块，用于获取目标随机噪声；
    目标数字岩心图像模块，用于将所述目标随机噪声输入所述数字岩心构建模型，得到目标数字岩心图像。
  8. 根据权利要求7所述的构建数字岩心的系统，其特征在于，所述数 字岩心训练图像模块，具体包括：
    扫描单元，用于利用图像扫描技术对岩心已知的岩石进行扫描，得到岩心已知的岩石灰度图像；
    平滑处理单元，用于提取所述岩心已知的岩石灰度图像中心位置的表征单元体，并对所述表征单元体进行平滑处理，得到平滑数字岩心图像；
    分割单元，用于采用分水岭分割方法对所述平滑数字岩心图像进行分割，得到数字岩心训练图像。
  9. 根据权利要求8所述的构建数字岩心的系统，其特征在于，所述数字岩心模型模块，具体包括：
    获取单元，用于获取生成对抗网络的激活函数和损失函数；所述生成对抗网络包括生成网络和判别网络；
    伪样本集单元，用于将随机样本噪声输入至所述生成网络，得到伪样本集；所述伪样本集包括多个第一伪数字岩心图像；
    判别网络模型单元，用于利用所述伪样本集和所述样本集训练所述判别网络，得到判别网络模型；所述判别网络模型为训练好的判别网络，所述判别网络模型的输入为第一伪数字岩心图像，输出为所述第一伪数字岩心图像的真假概率值；
    数字岩心构建模型单元，用于将所述随机样本噪声作为输入，利用所述判别网络模型对所述生成网络进行训练，得到生成网络模型；所述生成网络模型为训练好的生成网络，所述生成网络模型的输出为所述目标数字岩心图像；所述判别网络模型和所述生成网络模型构成数字岩心构建模型。
  10. 根据权利要求9所述的构建数字岩心的系统，其特征在于，所述判别网络模型单元，具体包括：
    第一判别损失函数子单元，用于从所述样本集中抽取N个子样本输入所述判别网络，计算第一判别损失函数；
    第一判别梯度子单元，用于利用所述第一判别损失函数计算所述判别网络每层的第一判别梯度；
    第二判别损失函数子单元，用于从所述伪样本集中抽取N个所述第一伪数字岩心图像输入所述判别网络，计算第二判别损失函数；
    第二判别梯度子单元，用于利用所述第二判别损失函数计算所述判别网络每层的第二判别梯度；
    判别损失函数子单元，用于将所述第一判别损失函数与所述第二判别损失函数相加得到所述判别损失函数；
    判别网络模型子单元，用于利用所述第一判别梯度、所述第二判别梯度和小批量梯度下降算法优化所述判别损失函数，得到最优判别网络参数，根据所述最优判别网络参数得到判别网络模型；判别网络参数为所述判别网络每层的权重和偏置。
PCT/CN2020/114481 2020-03-13 2020-09-10 一种构建数字岩心的方法及系统 WO2021179558A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/295,493 US11934488B2 (en) 2020-03-13 2020-09-10 Method and system for constructing digital rock

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010175137.6A CN111402266A (zh) 2020-03-13 2020-03-13 一种构建数字岩心的方法及系统
CN202010175137.6 2020-03-13

Publications (1)

Publication Number Publication Date
WO2021179558A1 true WO2021179558A1 (zh) 2021-09-16

Family

ID=71432490

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/114481 WO2021179558A1 (zh) 2020-03-13 2020-09-10 一种构建数字岩心的方法及系统

Country Status (4)

Country Link
US (1) US11934488B2 (zh)
CN (1) CN111402266A (zh)
AU (1) AU2021101142A4 (zh)
WO (1) WO2021179558A1 (zh)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111402266A (zh) * 2020-03-13 2020-07-10 中国石油大学(华东) 一种构建数字岩心的方法及系统
CN112132959B (zh) * 2020-09-25 2023-03-24 东北石油大学 数字岩心图像处理方法、装置、计算机设备及存储介质
CN112381845B (zh) * 2020-12-02 2023-04-18 中国石油大学(华东) 岩心图像生成方法、模型训练方法及装置
CN113310877B (zh) * 2021-06-07 2022-12-06 中国石油大学(华东) 一种区域结构差异化的岩石多尺度孔隙网络模型构建方法
CN114677281B (zh) * 2022-04-12 2024-05-31 西南石油大学 一种基于生成对抗网络的fib-sem超分辨率方法
CN114708463A (zh) * 2022-05-10 2022-07-05 重庆科技学院 一种岩石标本图像智能识别方法
CN115049781B (zh) * 2022-05-11 2023-05-23 西南石油大学 一种基于深度学习的页岩数字岩心三维重建方法
CN115115783B (zh) * 2022-07-08 2023-08-15 西南石油大学 一种模拟页岩基质纳微米孔隙的数字岩心构建方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107609587A (zh) * 2017-09-11 2018-01-19 浙江工业大学 一种基于深度卷积生成对抗网络的多类别多视图数据生成方法
CN108876901A (zh) * 2018-05-14 2018-11-23 中国石油大学(华东) 一种基于二维图像和多点统计学的数字岩心重建方法
CN109584337A (zh) * 2018-11-09 2019-04-05 暨南大学 一种基于条件胶囊生成对抗网络的图像生成方法
US20190198156A1 (en) * 2017-12-21 2019-06-27 International Business Machines Corporation Medical Image Classification Based on a Generative Adversarial Network Trained Discriminator
CN110009706A (zh) * 2019-03-06 2019-07-12 上海电力学院 一种基于深层神经网络和迁移学习的数字岩心重构方法
CN111402266A (zh) * 2020-03-13 2020-07-10 中国石油大学(华东) 一种构建数字岩心的方法及系统

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107862668A (zh) * 2017-11-24 2018-03-30 河海大学 一种基于gnn的文物图像复原方法
CN108062789B (zh) * 2017-12-20 2021-06-01 中国石油天然气股份有限公司 岩心样品选取方法和装置
CN108267466B (zh) * 2017-12-27 2020-11-03 中国石油天然气股份有限公司 一种构建数字岩心的方法及装置
CN109143353B (zh) * 2018-08-20 2019-10-01 电子科技大学 一种基于深度卷积生成对抗网络的叠前地震波形分类方法
US11520077B2 (en) * 2018-12-11 2022-12-06 Exxonmobil Upstream Research Company Automated reservoir modeling using deep generative networks
CN110074813B (zh) * 2019-04-26 2022-03-04 深圳大学 一种超声图像重建方法及系统
US11320357B2 (en) * 2019-12-23 2022-05-03 Chevron U.S.A. Inc. System and method for estimation of rock properties from core images

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107609587A (zh) * 2017-09-11 2018-01-19 浙江工业大学 一种基于深度卷积生成对抗网络的多类别多视图数据生成方法
US20190198156A1 (en) * 2017-12-21 2019-06-27 International Business Machines Corporation Medical Image Classification Based on a Generative Adversarial Network Trained Discriminator
CN108876901A (zh) * 2018-05-14 2018-11-23 中国石油大学(华东) 一种基于二维图像和多点统计学的数字岩心重建方法
CN109584337A (zh) * 2018-11-09 2019-04-05 暨南大学 一种基于条件胶囊生成对抗网络的图像生成方法
CN110009706A (zh) * 2019-03-06 2019-07-12 上海电力学院 一种基于深层神经网络和迁移学习的数字岩心重构方法
CN111402266A (zh) * 2020-03-13 2020-07-10 中国石油大学(华东) 一种构建数字岩心的方法及系统

Also Published As

Publication number Publication date
US11934488B2 (en) 2024-03-19
AU2021101142A4 (en) 2021-05-06
US20220101055A1 (en) 2022-03-31
CN111402266A (zh) 2020-07-10

Similar Documents

Publication Publication Date Title
WO2021179558A1 (zh) 一种构建数字岩心的方法及系统
Wickramasinghe et al. Voxel2mesh: 3d mesh model generation from volumetric data
CN109493308B (zh) 基于条件多判别生成对抗网络的医疗图像合成与分类方法
CN110276316B (zh) 一种基于深度学习的人体关键点检测方法
WO2022089391A9 (zh) 模型训练方法、人体姿态检测方法、装置、设备及存储介质
CN109635883A (zh) 基于深度堆叠网络的结构信息指导的中文字库生成方法
CN112070658B (zh) 一种基于深度学习的汉字字体风格迁移方法
CN111489357A (zh) 一种图像分割方法、装置、设备及存储介质
CN112734931B (zh) 一种辅助点云目标检测的方法及系统
CN111652864A (zh) 一种基于条件式生成对抗网络的铸件缺陷图像生成方法
CN116206185A (zh) 一种基于改进YOLOv7的轻量级小目标检测方法
CN113269256A (zh) 一种MiSrc-GAN模型的构建方法及应用
CN114118012A (zh) 一种基于CycleGAN的个性字体生成方法
Eschweiler et al. Towards annotation-free segmentation of fluorescently labeled cell membranes in confocal microscopy images
CN117333637B (zh) 三维场景的建模及渲染方法、装置及设备
Yang Application and development of digital enhancement of traditional sculpture art
CN111881919B (zh) 一种基于追踪式格网剖分的线要素智能化简方法与装置
CN112766381A (zh) 有限样本下属性引导的sar图像生成方法
CN110866866A (zh) 图像仿色处理方法、装置、电子设备及存储介质
CN113111906B (zh) 一种基于单对图像训练的条件生成对抗网络模型的方法
Dash et al. A scalable pipeline for gigapixel whole slide imaging analysis on leadership class hpc systems
CN115293983A (zh) 一种融合多层级互补特征的自适应图像超分辨率复原方法
CN114758123A (zh) 一种遥感影像目标样本增强方法
Li et al. Real-time image carrier generation based on generative adversarial network and fast object detection
Haripriya et al. A STUDY ON CNN ARCHITECTURE OF VGGNET AND RESNET FORDICOM IMAGE CLASSIFICATION

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20924542

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20924542

Country of ref document: EP

Kind code of ref document: A1