WO2020015086A1

WO2020015086A1 - Porous medium permeability prediction method based on intelligent machine image learning

Info

Publication number: WO2020015086A1
Application number: PCT/CN2018/104640
Authority: WO
Inventors: 刘江峰; 曹栩楼; 陈师杰; 黄炳香; 陈浙锐; 宋帅兵; 倪宏阳
Original assignee: 中国矿业大学; 徐州佑学矿业科技有限公司
Priority date: 2018-07-18
Filing date: 2018-09-07
Publication date: 2020-01-23
Also published as: CN109191423A; AU2018424207A1; CN109191423B; AU2018424207B2

Abstract

Disclosed is a porous medium permeability prediction method based on intelligent machine image learning. The method comprises: selecting multiple groups of porous medium materials of the same type but different dry densities, and determining the real permeability of each group of porous medium materials; scanning each group of porous medium materials using an SEM to obtain SEM images thereof, and then calculating a gray average, a gray variance, an image energy, an image entropy and a fractal dimension of each of the SEM images; using an extreme learning machine neural network model to perform training and learning on five image feature parameters of each of the SEM images and the real permeability corresponding thereto to determine a change relationship between the five image feature parameters and the real permeability; and during prediction, inputting parameters of the SEM image of a porous medium material of unknown permeability, the extreme learning machine neural network model being able to predict thereby the permeability of the porous medium material. The present invention is simple to operate and has a short test period, achieves a high accuracy of predicted permeability and is low in test costs.

Description

Prediction method of porous medium permeability based on machine image intelligent learning

Technical field

The invention relates to a porous medium permeability prediction method based on intelligent learning of machine images.

Background technique

Permeability is a key technical indicator in many engineering fields, such as the exploitation of coalbed methane and shale gas, the problem of gas migration in the barrier system during nuclear waste disposal, and the deep geological storage of CO ₂ . At present, the testing methods for the permeability of porous media materials mainly include: mercury injection method, gas steady state method and gas transient method. Although the test results of these more mature macro testing methods are very accurate, the test process is cumbersome and it is not easy for non-professional operators to operate. In addition, the test cycle of each of the above methods varies from several days to several months depending on the porous media material; And the equipment required for testing is expensive.

Summary of the invention

In view of the problems existing in the prior art, the present invention provides a porous medium permeability prediction method based on machine image intelligent learning, which has simple operation, short test period, high accuracy of predicted permeability, and low test cost.

In order to achieve the above objective, the technical solution adopted by the present invention is: a method for predicting the permeability of porous media based on intelligent learning of machine images, the specific steps are:

A. Establish a porous media material permeability database:

a. In the same porous medium material, select more than 30 groups by measuring the dry density (the more test groups selected, the better, the more the number of results, the more accurate the results). Porous media materials with different dry densities (porous media materials are mainly Rock and soil particles are cemented with each other, including the soil skeleton and pores, and the permeability of porous media materials is mainly determined by the connected pores between these soils. Therefore, the mass of solid rock and soil particles is the same as the total volume of the soil. The ratio, that is, the dry density, reflects the porosity ratio of rock and soil. The dry density can be used to judge the permeability of porous media materials. The larger the dry density, the smaller the permeability; the smaller the dry density, the greater the permeability. Existing macroscopic detection methods of permeability test each group of porous media materials to obtain the true permeability value of each group of porous media materials;

b. Scan each group of porous media materials by SEM (ie, scanning electron microscope) to obtain SEM images of each group of porous media materials. Compare the SEM images of each group of porous media materials with the true permeability value of the group of porous media materials. correspond;

The order of steps a and b is interchangeable;

B, SEM image feature information extraction and analysis:

Because the SEM image is a grayscale image, MATLAB software is used to identify the grayscale value of each pixel in the SEM image to form a grayscale matrix of each SEM image, and then each of the SEM images is calculated based on the grayscale matrix of the SEM image. The SEM image's image gray average, image gray variance, image energy, image entropy, and fractal dimension are used to characterize the SEM image;

C. The computer neural network is used to perform deep learning on the five image feature parameters of each SEM image obtained:

A known extreme learning machine neural network model is used to train and learn the five image feature parameters of each SEM image and their corresponding true permeability. After training and learning, the neural network model determines the five image feature parameters and Logical regression relationship between the real penetration rate and logistic regression fitting; since the extreme learning machine neural network model runs, the weights and hidden layer thresholds are generated. During the data training and learning process, it can be automatically adjusted to determine the uniqueness. Optimal solution;

D. Permeability prediction for the same porous media material of unknown permeability:

① Scanning the same porous medium material of unknown permeability with an SEM electron microscope to obtain an SEM image of the porous medium material, and then calculating the image gray average value, image gray variance, image energy, and image entropy value of the SEM image And fractal dimensions;

② The five image feature parameters of the SEM image obtained in step ① are input into a computer neural network that has been deeply learned. The computer neural network determines the change relationship between the feature parameters of the five SEM images and the actual permeability according to the input SEM image and After analysis, the computer neural network predicts the permeability corresponding to the SEM image and displays it through a display device.

Further, the existing macroscopic detection method of permeability includes a mercury injection method, a gas steady state method, and a gas transient method.

Further, the specific calculation process of the five image feature parameters in step B is:

I. Normalize the gray value matrix first to obtain a gray histogram. The gray histogram is a discrete function of the gray level, which can reflect the number of pixels with the gray level in the image to the total pixels of the image. Percentage, that is, the frequency of pixels with gray level i in the image, as follows:

Among them, i represents the gray level, N represents the total number of image pixels, n _i represents the total number of pixels of the gray level i in the image, and L represents the number of types of gray levels;

II. Average gray value of the image: The average gray value of the image is a measure that reflects the average brightness of the entire gray image texture. The higher the permeability of the porous media material, the more pores in the SEM image, and the darker the image as a whole. The lower the average gray value will be; normalize the gray value matrix to get the gray histogram. The average gray value of the image is as follows:

In the formula, m is the average value of the gray level of the image, i is the gray level, p (i) is the discrete function of the gray level, and L is the number of types of gray level;

III. Image Gray Variance: The image gray variance is a measure of the discreteness of the gray value of a gray image and is also a measure of the average contrast of the image texture. The image gray variance is as follows:

In the formula, σ ² is the gray variance of the image, i is the gray level, m is the average gray level of the image, p (i) is the discrete function of the gray level, and L is the number of types of gray level;

IV. The image energy reflects the uniformity of the gray value of the gray image. Generally, the more uniform the gray value distribution of the image is, the greater the energy of the image is; otherwise, the image energy will be smaller. The image energy can reflect the uniformity of the distribution of rock and soil pores and soil skeleton in the SEM image of porous media. The image energy is as follows:

In the formula, U is the image energy, p (i) is a discrete function of gray levels, and L is the number of types of gray levels;

V. Image entropy: The image entropy reflects the uniformity of gray histogram distribution. The larger the image entropy, the greater the randomness of the image. On the contrary, the smaller the randomness of the image, the image entropy is as follows:

In the formula, e is the entropy value of the image, p (i) is a discrete function of gray levels, and L is the number of types of gray levels;

VI. Fractal Dimension: The fractal dimension reflects the relationship between the pore structure of the porous material and the pore surface. It is related to the complexity, non-uniformity, surface roughness, and regularity of the pore structure. The higher the fractal dimension, the more pore The more irregular the surface, the stronger the non-uniformity of the pore structure. The fractal dimension is defined as follows: Let A be any non-empty bounded subset of R ⁿ space. For any r> 0, the side length required to cover A is r. The minimum number of n-dimensional cubes is N _r (A). Let d exist so that when r → 0:

N _r (A) ∝1 / r ^d

Then call d the box dimension of A. There is only one positive number k, such that:

Taking the logarithm of the left and right sides of the above formula, we can get:

In the calculation process, according to the actual situation, the number of boxes N _r (A) required to cover A when different r values are counted, and the logarithmic coordinate system with lgr as the abscissa and lgN _r (A) as the ordinate Draw in

Finally, the absolute value is obtained by the slope of the fitted line of these points, and the fractal dimension of the set A is obtained.

Compared with the prior art, the present invention adopts a combination of image recognition and neural network to obtain the true permeability of each group of porous media materials by first testing the permeability of multiple groups of the same porous media material with different dry densities. Then, the SEM images of each group of porous media materials were scanned to obtain the SEM images of each group, and then each SEM image was identified and calculated by MATLAB software for each pixel in the SEM image to obtain the gray of the SEM image. Degree average, gray variance, image energy, image entropy value and fractal dimension; the extreme image learning machine neural network model is used to train and learn the five image feature parameters of each SEM image and their corresponding true permeability, and determine five The relationship between the characteristics of an image and its true permeability. When predicting, the SEM image of the same porous medium material with unknown permeability is input, and the image gray average, image gray variance, and image are calculated. Energy, image entropy value and fractal dimension, the extreme learning machine neural network model can predict the permeability of the porous media material. The method has the advantages of simple operation, short test period, high accuracy of predicted permeability, and low test cost.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of device connection used in the present invention;

2 is a structural diagram of a neural network of an extreme learning machine in the present invention;

FIG. 3 is an SEM image of a porous medium material in the present invention;

FIG. 4 is a comparison chart between the predicted permeability value and the actual permeability value in the test certificate of the present invention.

detailed description

The invention will be further described below.

As shown in the figure, a method for predicting the permeability of porous media based on intelligent learning of machine images, the specific steps are:

A. Establish a porous media material permeability database:

a. In the same porous media material, select more than 30 groups by measuring the dry density (the more test groups selected, the better, the more the number of results, the more accurate the results). Porous media materials with different dry densities, and then use the existing The macroscopic detection method of permeability tests each group of porous media materials to obtain the true permeability value of each group of porous media materials;

b. Scan the SEM electron microscopy of each group of porous media materials to obtain the SEM images of each group of porous media materials, and associate the SEM images of each group of porous media materials with the true permeability value of the group of porous media materials;

B, SEM image feature information extraction and analysis:

The known neural network model of the extreme learning machine is used to train and learn the five image feature parameters of each SEM image and its corresponding true permeability. The extreme learning machine is a new algorithm for feedforward neural networks. After training and learning, the network model determines the logical change relationship between the five image feature parameters and the true permeability, and performs logistic regression fitting. Since the random neural network model runs, the weights and hidden layer thresholds are generated. During the training and learning process, the only optimal solution can be determined automatically after adjustment; as shown in Figure 2, if the connection weight of the input layer and the hidden layer is set to w, as shown in the following formula:

Let the connection weight of the hidden layer and the output layer be β, as shown in the following formula:

It can be obtained that the output of the extreme learning machine neural network is T when the input sample set is Q, as shown in the following formula:

② The five image feature parameters of the SEM image obtained in step ① are input into a computer neural network that has been deeply learned. The computer neural network determines the change relationship between the feature parameters of the five SEM images and the actual permeability according to the input SEM image and After the analysis, the computer neural network predicts the permeability corresponding to the SEM image and displays it through a display device.

Further, it is characterized in that the existing macroscopic detection method of permeability includes a mercury injection method, a gas steady state method, and a gas transient method.

Further, it is characterized in that the specific calculation process of the five image feature parameters in step B is:

I. Normalize the gray value matrix first to obtain the gray histogram. The gray histogram is a discrete function of gray levels, as shown in the following formula:

II. Average gray value of the image: The average gray value of the image is a measure that reflects the average brightness of the entire gray image texture. The average value of the image gray is as follows:

IV. The image energy reflects the uniformity of the gray value of the gray image. The image energy is as follows:

V. Image entropy value: The image entropy value reflects the uniformity of the gray histogram distribution. The image entropy value is as follows:

VI. Fractal Dimension: The fractal dimension reflects the relationship between the pore structure and the surface of the porous medium. The fractal dimension is defined as follows: Let A be any non-empty bounded subset of R ⁿ space. , The minimum number of n-dimensional cubes with side length r required to cover A is N _r (A). If d is present, when r → 0:

N _r (A) ∝1 / r ^d

Finally, the absolute value is obtained by the slope of the fitted line of these points, that is, the fractal dimension of the set A is obtained.

The test proves that the bentonite (GMZ bentonite) from Gaomiaozi in Inner Mongolia is selected. This porous medium material can be used as one of the buffer barrier materials for deep geological repositories of radioactive nuclear waste. Its permeability index is related to the tightness of the entire repository. ; Select 34 groups of GMZ bentonite blocks with different dry densities, complete 30 groups of bentonite blocks according to the steps A to C of the present invention to complete the training and learning process of the extreme learning machine neural network model, and then take the other 4 groups of bentonite blocks as required For porous media materials with predicted permeability, SEM images of 4 groups of bentonite blocks are obtained by scanning SEM electron microscopy, and then according to step D of the present invention, the neural network model of the extreme learning machine outputs the predicted permeability of 4 groups of bentonite blocks through the result output display device. The four groups of bentonite blocks were then tested for their true permeability using the permeability macro-detection method. Finally, the four groups of bentonite blocks were numbered sequentially, and the predicted results and the tested permeability were plotted into Figure 4;

It can be known from FIG. 4 that after actual testing, the error between the permeability of the four groups of bentonite blocks predicted by the present invention and the actual permeability obtained after the actual test is less than 5%, so the permeability of the present invention to porous media materials is explained. The prediction accuracy is high.

Claims

A method for predicting the permeability of porous media based on intelligent learning of machine images is characterized in that the specific steps are:

A. Establish a porous media material permeability database:

a. In the same porous medium material, select more than 30 groups of porous medium materials with different dry densities by measuring the dry density, and then test the porous medium materials of each group using a known macroscopic detection method of permeability to obtain each group of porous medium. True permeability value of the material;

b. Scan the SEM electron microscopy of each group of porous media materials to obtain the SEM images of each group of porous media materials, and associate the SEM images of each group of porous media materials with the true permeability value of the group of porous media materials;

B, SEM image feature information extraction and analysis:

Because the SEM image is a grayscale image, MATLAB software is used to identify the grayscale value of each pixel in the SEM image to form a grayscale matrix of each SEM image, and then each of the SEM images is calculated based on the grayscale matrix of the SEM image. The SEM image's image gray average, image gray variance, image energy, image entropy, and fractal dimension are used to characterize the SEM image;

C. The computer neural network is used to perform deep learning on the five image feature parameters of each SEM image obtained:

A known extreme learning machine neural network model is used to train and learn the five image feature parameters of each SEM image and their corresponding true permeability. After training and learning, the neural network model determines the five image feature parameters and Logical change relationship between real permeability, and perform logistic regression fitting;

D. Permeability prediction for the same porous media material of unknown permeability:

① Scanning the same porous medium material of unknown permeability with an SEM electron microscope to obtain an SEM image of the porous medium material, and then calculating the image gray average value, image gray variance, image energy, and image entropy value of the SEM image And fractal dimensions;

② The five image feature parameters of the SEM image obtained in step ① are input into the computer neural network that has been deeply learned. The computer neural network determines the relationship between the feature parameters of the five SEM images and the actual permeability based on the input SEM image and After the analysis, the computer neural network predicts the permeability corresponding to the SEM image and displays it through a display device.
The method for predicting the permeability of porous media based on intelligent learning of machine images according to claim 1, characterized in that the existing macroscopic detection method of permeability comprises a mercury injection method, a gas steady state method, and a gas transient method .
The method for predicting the permeability of porous media based on intelligent learning of machine images according to claim 1, wherein the specific calculation process of the five image characteristic parameters in step B is:

Ⅰ. Normalize the gray value matrix first to obtain a gray histogram. The gray histogram is a discrete function of gray levels, as follows:

Among them, i represents the gray level, N represents the total number of image pixels, n i represents the total number of pixels of the gray level i in the image, and L represents the number of types of gray levels;

Ⅱ. The average value of the gray level of the image: The average value of the gray level of the image is a measure that reflects the average brightness of the entire gray image texture. The average value of the gray level of the image is as follows:

In the formula, m is the average value of the gray level of the image, i is the gray level, p (i) is the discrete function of the gray level, and L is the number of types of gray level;

Ⅲ. Image Gray Variance: The image gray variance is a measure of the dispersion of the gray value of a gray image and is also a measure of the average contrast of the image texture. The image gray variance is as follows:

In the formula, σ 2 is the gray variance of the image, i is the gray level, m is the average gray level of the image, p (i) is the discrete function of the gray level, and L is the number of types of gray level;

Ⅳ. The image energy reflects the uniformity of the gray value of the gray image. The image energy is as follows:

In the formula, U is the image energy, p (i) is a discrete function of gray levels, and L is the number of types of gray levels;

Ⅴ. Image entropy value: The image entropy value reflects the uniformity of the gray histogram distribution. The image entropy value is as follows:

In the formula, e is the entropy value of the image, p (i) is a discrete function of gray levels, and L is the number of types of gray levels;

Ⅵ. Fractal dimension: The fractal dimension reflects the relationship between the pore structure of the porous media material and the pore surface. The fractal dimension is defined as follows: Let A be any non-empty bounded subset of R n space. For any r> 0 , The minimum number of n-dimensional cubes with side length r required to cover A is N r (A). If d is present, when r → 0:

N r (A) ∝1 / r d

Then call d the box dimension of A. There is only one positive number k, such that:

Taking the logarithm of the left and right sides of the above formula, we can get:

In the calculation process, according to the actual situation, the number of boxes N r (A) required to cover A when different r values are counted, and the logarithmic coordinate system with lgr as the abscissa and lgN r (A) as the ordinate Draw in
Finally, the absolute value is obtained by the slope of the fitted line of these points, and the fractal dimension of the set A is obtained.