CN111753958A - Microfacies identification method of Dengying Formation microbial rock based on deep learning of logging data - Google Patents

Abstract

The invention discloses a lamp shadow group microorganism rock microphase identification method based on logging data deep learning, which comprises the following steps of: s1, establishing a sample database according to the multi-type well logging data of the determined lithology of the microorganism; s2, checking the integrity of the logging curve in the sample database, and processing the missing segment data by using a principal component analysis and linear regression method to realize data supplement; s3, homogenizing the sample data, and dividing the sample data into a training set and a verification set; s4, establishing a neural network model based on TensorFlow/Playground standard spiral data. The invention establishes the neural network model based on the TensorFlow/Playground standard spiral data by analyzing various well logging data of determined microorganism lithology, carries out microorganism rock microphase identification through the neural network model, has stronger objectivity and systematicness, and improves the identification accuracy and the identification efficiency.

Description

Microfacies identification method of Dengying Formation microbial rock based on deep learning of logging data

technical field

The invention relates to a method for identifying microfacies of Dengying Formation microbial rock based on deep learning of logging data.

Background technique

Well logging, the full name of geophysical logging or mine geophysics, referred to as well logging, is a method of measuring geophysical parameters using the electrochemical properties, electrical conductivity, acoustic properties, radioactivity and other geophysical properties of rock formations. In the process of oil and gas exploration, the formation lithology can be effectively divided by the curve fluctuation characteristics of various logging curves such as acoustic wave, resistance, spontaneous potential, and natural gamma, and then the formation rock particle size, fluid-bearing properties and saturation, formation porosity can be effectively divided. Calculation of degree and formation permeability, and finally complete the identification of rock types.

In theory, different types of rocks will show obvious fluctuation anomalies on different logging curves. However, in the actual logging process, many factors such as the type of drilling mud and the collapse of the wellbore make it extremely difficult to manually identify and judge the logging curves. In the process of manual determination, the subjective factors of the discriminating personnel will also produce large errors in the identification results, which makes the identification of rock types lack objectivity and systematicness, and it is even difficult to provide a basis for discrimination after judging the rock types. Therefore, it is particularly important to establish an objective, systematic and automatic identification mode combining multiple logging curves.

Existing methods for identifying lithology based on logging data are mainly aimed at the properties of coal and the division of sand and mudstone in coalfield geology, and there is no logging identification technology for microbial rocks. And the existing technology is only for a certain curve, for example, only the GR curve is used in coal property prediction and sand-mudstone division, and there is no identification method for multiple curves as the judgment basis at the same time.

The existing logging identification technology of sedimentary microfacies recognizes the manifestations of known sedimentary microfacies on the GR curve, so as to achieve the purpose of identifying lithology and sedimentary microfacies. The manifestations of each sedimentary microphase on the GR curve are shown in Fig. 1.

Extract the statistical characteristics of the morphology on the GR logging curve, including: average amplitude, amplitude difference, relative center of gravity, relative number of sign changes, variance, average median, variance root of variation, etc., and transform the morphological difference on the logging data. into mathematical language.

Determine the abnormal shape on the unknown curve. Common judgment methods are: fuzzy mean clustering method, bayes discrimination method. The basic principle of the so-called fuzzy mean clustering method is to statistically calculate the clustering centers of various sedimentary microfacies, and then calculate the distance from the clustering centers of various microfacies for the samples that need to be identified, and attribute them to the lowest distance. which type of microphase. The bayes discriminant method is actually a modification of the shortest distance discrimination method. While using the distance to discriminate, it takes into account the difference in the prior probabilities of various microphases, and performs clustering by the expected distance.

Disadvantages of the prior art:

(1) The existing logging identification technology of lithology and sedimentary microfacies is only for sand and mudstone sections, and it is difficult to apply to carbonate rocks to which microbial rocks belong.

(2) The existing logging identification technology can only complete the identification and judgment of a single log curve, and compared with the comprehensive identification of multiple curves, it is easy to cause identification errors.

(3) The existing logging identification technology needs to extract all the mathematical features of the curve shape, and the extraction time is long and the efficiency is low.

SUMMARY OF THE INVENTION

The object of the present invention is to overcome the deficiencies of the prior art, and provide a kind of neural network model based on the standard spiral data of TensorFlow/Playground by analyzing the various types of logging data of the determined microbial lithology. The identification method of microbial rock microfacies in Dengying Formation based on deep learning of logging data is used to identify microbial rock microfacies, which not only improves the accuracy of identification but also improves the efficiency of identification work.

The object of the present invention is achieved through the following technical solutions: a method for identifying microfacies of Dengying Formation microbial rock based on deep learning of logging data, comprising the following steps:

S1. Establish a sample database according to various types of logging data that have determined microbial lithology;

S2. Check the integrity of the logging curves in the sample database, and use the principal component analysis and linear regression methods to process the missing data to realize data supplementation;

S3. The sample data is equalized, and the equalized sample data is divided into a training set and a verification set;

S4. Establish a neural network model based on TensorFlow/Playground standard spiral data.

Further, the specific implementation method of step S1 is as follows: using a polarized light microscope to identify the drilling core slices, clarifying the microbial rock fabric, and then determining the microbial rock type and sedimentary microfacies type; matching the logging data with the microbial rock type, Build a sample database.

Further, the specific implementation method of the step S2 is: using the built-in function of the Tesorflow/pandas processing module to quickly traverse and count each logging curve to check the integrity of the logging curve; the specific method is as follows: for each logging curve Perform numerical grayscale conversion, convert the value of non-mutated pixel points to 0 (black) according to the blurred boundary of the image, and convert the value of abrupt pixel points to 1 (white) according to the blurred boundary of the image; then use the Pandas.isnull function to judge whether the curve is complete.

Further, the specific implementation method of step S3 is: performing normalization/positive value operation on each geophysical logging parameter, and unifying each logging parameter to a scale of 0 to 1:

Among them, P _i is the scale value after normalization/positive value, Li is the _ith logging parameter, L _min represents the minimum value among all logging parameters, and L _max is the maximum value among all logging parameters.

Further, the specific implementation method of step S4 is: using the Python/Keras interface in combination with the one-hot/Softmax function interface of one-hot encoding to construct a neural network model for the logging curve, the network model includes an input layer, an output layer and 4 hidden layers, of which, hidden layer 1 contains 600 neurons, and the activation function uses elu; hidden layer 2 contains 128 neurons, and the activation function uses Relu; hidden layer 3 contains 32 neurons, and the activation function uses sigmoid; hidden Layer 4 contains 8 neurons, and the activation function uses softsign.

The beneficial effects of the present invention are as follows: the present invention establishes a neural network model based on TensorFlow/Playground standard spiral data by analyzing various types of well logging data that have determined microbial lithology, and uses the neural network model to identify microbial rock microfacies , has strong objectivity and systematicness, and improves the efficiency of identification work while improving the accuracy of identification.

Description of drawings

Figure 1 shows the manifestations of each sedimentary microphase on the GR curve;

2 is a flow chart of a method for identifying microfacies of Dengying Formation microbial rock according to the present invention;

FIG. 3 is a diagram of microbial rock types and sedimentary microfacies types of this embodiment;

FIG. 4 is a result diagram of the identification of the logging curve according to the present invention.

Detailed ways

Neural network is a mathematical model that is inspired by the operation of biological (human or other animal) neural network functions and applies the structure similar to the synaptic connection of the brain for information processing. A neural network is an operational model that consists of a large number of nodes (or neurons) and connections between them. Each node represents a specific output function, called the activation function. The connection between each two nodes represents a weighted value for the signal passing through the connection, called the weight, which is equivalent to the memory of the artificial neural network. The output of the network is different according to the connection mode of the network, the weight value and the excitation function. The neural network is usually an optimized combination of a learning method based on the mathematical statistics type, which is an extended application of the mathematical statistics method. In the field of artificial perception of artificial intelligence, neural networks can make decisions about artificial perception through the application of mathematical statistics (that is to say, through statistical methods, artificial neural networks can have the same simple decision-making ability and simplicity as human beings. ability to judge).

The complexity of logging curves and the regularity of fluctuations caused by changes in lithology make the work of using neural networks to identify logging curves more objective and systematic, which not only improves the recognition accuracy but also improves the Identify work efficiency.

GR curve: Spontaneous potential logging, which is a part of electrical logging, is mainly used for sand and mudstone sections. Spontaneous potential logging measures the curve of spontaneous potential as a function of well depth. Spontaneous potential logging is one of the important methods to divide and evaluate reservoirs because of the obvious abnormal display in the permeable layer.

The technical solutions of the present invention are further described below with reference to the accompanying drawings.

As shown in FIG. 2 , a method for identifying microfacies of Dengying Formation microbial rocks based on deep learning of logging data of the present invention includes the following steps:

S1. Establish a sample database according to various types of logging data that have determined microbial lithology;

The specific implementation method is as follows: using polarized light microscope to identify drilling core slices to clarify microbial rock fabric, and then determine microbial rock types and sedimentary microfacies types; match well logging data with microbial rock types to establish a sample database. As shown in Figure 3, in the figure, (A) is the microscopic characteristics of the silicified laminar bond rock, MX105, the fourth member of the lamp, 5360.58m, 5 times the cross polarized light; (B) is the silicified laminar bond rock Microscopic features, MX105, Deng 4th member, 5321.4m, 5x crossed polarized light; (C) Microscopic features of the clot-bonded rock, MX105, Deng 4th member, 5303.7m, 4X single polarized light; (D) Microscopic features of clotted bond rock, MX105-3, Deng 4 member, 5x single polarized light; (E) Microscopic features of laminar bond rock, MX13-5, Deng 4 Section, 5 times single polarized light; (F) is the microscopic characteristics of micrite bond rock, MX11-0041, 5485.69m, the fourth section of the lamp, 2 times single polarized light. The logging data includes acoustic AC, well diameter CAL, compensated neutron CNL, compensated density DEN, natural gamma GR, photoelectric absorption cross section index PE, bottom gradient resistivity RT and flushing zone resistivity RXO, etc. parameter.

S2. Check the integrity of the logging curves in the sample database, and use the principal component analysis and linear regression methods to process the missing data to realize data supplementation;

As one of the important steps of preprocessing, the integrity test of logging curve and data processing of missing segments are the keys to using neural network model to carry out prediction work. The specific inspection method is: use the built-in function of the Tesorflow/pandas processing module to quickly traverse and count each logging curve to check the integrity of the logging curve; the specific method is as follows: perform numerical grayscale transformation on each logging curve, The blurred boundary of the image converts the value of the non-mutated pixel point to 0 (black), and the value of the mutation pixel of the blurred boundary of the image is converted to 1 (white); then use the Pandas.isnull function to judge whether the curve is complete, the closer the value of the pixel matrix is to 255 closer to white. After the fuzzy image processing is completed, the recognition samples of the neural network are changed from a discrete data set with a huge variation range to a combination of a certain color interval, which realizes the natural classification of discrete sample data.

The method of principal component analysis and linear regression is used for processing, that is, the missing data is supplemented, and the average value of the logging curve is filled in the missing data.

The implementation process using Python is as follows:

#Filter feature AC->RXO

selected_cols=['AC','CAL','CNL','DEN','GR','PE','RT','RXO',

'Laminated lamellae', 'clumps', 'siliceous', 'micrite']

selected_learn_data=learn_data[selected_cols]

#Integrity check of logging data in coring section

#Determine which logging curves are incomplete, True indicates that there are incomplete items, and count the number of nulls

selected_learn_data.isnull().sum()

# Fill in the value for the missing PE curve and fill in the average value in the vacancy

PE_mean_value=selected_learn_data['PE'].mean()

selected_learn_data['PE']=selected_learn_data['PE'].fillna(PE_mean_value)

S3. The sample data is equalized, and the equalized sample data is divided into a training set and a verification set;

The application can also use the shuffle function of the Pandas database to shuffle the data. The purpose of shuffling the data is: the logging data of the same lithology often show similar data characteristics, so when using the gradient descent method for sample training, the learned data may fall on the local optimal solution, rather than the global optimal solution. optimal solution. After shuffling the data, the data isolation is significantly improved, effectively avoiding the "trap of local optimal solution".

In the process of logging production and construction, different geophysical parameters have different value ranges, and the value ranges of different parameters are usually quite different. / Positive value operation, unify all logging parameters to 0~1 scale:

Among them, P _i is the scale value after normalization/positive value, Li is the _ith logging parameter, L _min represents the minimum value among all logging parameters, and L _max is the maximum value among all logging parameters.

After the normalization/normalization operation, the training caused by the parameter numbers is eliminated. To ensure the accuracy of the model, the samples in the neural network model need to be divided into training sets and validation sets. The function of the training set is to extract the characteristics of the sample data, and complete the training of the neural network model through the feedback/excitation function and the back-propagation mechanism. The function of the validation set is to test the trained fully connected neural network, and to judge the accuracy of the model by identifying and judging the data that did not participate in the training. The data ratio of the training set and the validation set in this embodiment is 8:2.

The implementation process using Python is as follows:

#shuffle shuffle data

shuffled_learn_data=selected_learn_data.sample(frac=1)

#Separate samples and labels

#Convert to ndarray array

ndarray_data=shuffled_learn_data.values

#The first 8 columns are learning samples

learn_sample=ndarray_data[:,:8]

#The last 4 columns are the label data

labels=ndarray_data[:,8:14]

# Learning sample normalization processing

from sklearn import preprocessing

minmax_scale=preprocessing.MinMaxScaler(feature_range=(0,1))

norm_learn_sample=minmax_scale.fit_transform(learn_sample)

# Divide the learning sample into training set and test set

x_data=norm_learn_sample

y_data=labels

train_size=int(len(x_data)*0.8)

x_train=x_data[:train_size]

y_train=y_data[:train_size]

x_test=x_data[train_size:]

y_test=y_data[train_size:]

S4. Establish a neural network model based on TensorFlow/Playground standard spiral data.

The neural network model structure test based on TensorFlow/Playground standard spiral data shows that the four-layer fully connected neural network with relu-elu-sogmoid-softsign has a recognition accuracy of 75%-80%. Using Python/Keras interface combined with one-hot coding (one-hot)/Softmax function interface to build a neural network model for logging curves: the network model includes an input layer, an output layer and 4 hidden layers, of which hidden layer 1 It contains 600 neurons, and the activation function uses elu; the hidden layer 2 contains 128 neurons, and the activation function uses Relu; the hidden layer 3 contains 32 neurons, and the activation function uses sigmoid; the hidden layer 4 contains 8 neurons, and the activation function Use softsign.

The implementation process using Python is as follows:

#Build a neural network architecture based on Keras

model=tf.keras.models.Sequential()

#1st Neural Training Layer

model.add(tf.keras.layers.Dense(units=600,

input_dim=8,

use_bias=True,

kernel_initializer='uniform',

bias_initializer='zeros',

activation='relu'))

#2nd Neural Training Layer

model.add(tf.keras.layers.Dense(units=128,

activation='elu'))

#3rd Neural Training Layer

model.add(tf.keras.layers.Dense(units=32,

activation='sigmoid'))

#4th Neural Training Layer

model.add(tf.keras.layers.Dense(units=8,

activation='softsign'))

#output layer

model.add(tf.keras.layers.Dense(units=4,

activation='softmax'))

#model structure print

model.summary()

#optimizer configuration

model.compile(optimizer=tf.keras.optimizers.Adam(0.005),

loss='categorical_crossentropy',

metrics=['accuracy'])

#optimizer configuration

model.compile(optimizer=tf.keras.optimizers.Adam(0.005),

loss='categorical_crossentropy',

metrics=['accuracy'])

#train model

train_history=model.fit(x=x_train,

y=y_train,

validation_split=0.2,

epochs=500,

batch_size=100,

verbose=2)

Model training visualization

import matplotlib.pyplot as plt

def visual_train(train_history,train_metric,validation_metric):

plt.plot(train_history.history[train_metric])

plt.plot(train_history.history[validation_metric])

plt.title('Train History')

plt.ylabel(train_metric)

plt.xlabel('epochs')

plt.legend(['train','validation'],loc='upper left')

plt.show

The procedure for making predictions is as follows:

import numpy

import pandas as pd

import math

# Predict unpredicted well sections

data_file_path="C:\\Users\\74086\\Desktop\\data\\simulation data of the well to be tested.xlsx"

pred_data=pd.read_excel(data_file_path,sheet_name=22)

#Filter unpredicted well section features AC->RXO

selected_pred_cols=['AC','CAL','CNL','DEN','GR','PE','RT','RXO']

selected_pred_data=pred_data[selected_pred_cols]

#Fill values for missing PE/RT/RXO curves, and fill in regression values in the vacancies

PE_mean_value=selected_pred_data['PE'].mean()

selected_pred_data['PE']=selected_pred_data['PE'].fillna(PE_mean_value)

RT_mean_value=selected_pred_data['RT'].mean()

selected_pred_data['RT']=selected_pred_data['RT'].fillna(RT_mean_value)

RXO_mean_value=selected_pred_data['RXO'].mean()

selected_pred_data['RXO']=selected_pred_data['RXO'].fillna(RXO_mean_value)

#The last 8 columns are the samples to be predicted

ndarray_pred_data=selected_pred_data.values

#standardization

from sklearn import preprocessing

minmax_scale=preprocessing.MinMaxScaler(feature_range=(0,1))

norm_learn_sample=minmax_scale.fit_transform(ndarray_pred_data)

#Predict the uncored section

pred_result=model.predict(norm_learn_sample)

The neural network constructed according to the logging curve data is used for multiple tests to identify various logging curves of different wells. The test curve shown in Figure 4 is obtained. In the figure, (a) represents a schematic diagram of the variation of the accuracy with the training, and (b) is a schematic diagram of the variation of the error with the training.

Those of ordinary skill in the art will appreciate that the embodiments described herein are intended to assist readers in understanding the principles of the present invention, and it should be understood that the scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations without departing from the essence of the present invention according to the technical teachings disclosed in the present invention, and these modifications and combinations still fall within the protection scope of the present invention.

Claims (5)

1. A biological rock microphase identification method based on logging data deep learning is characterized by comprising the following steps:

s1, establishing a sample database according to the multi-type well logging data of the determined lithology of the microorganism;

s2, checking the integrity of the logging curve in the sample database, and processing the missing segment data by using a principal component analysis and linear regression method to realize data supplement;

s3, homogenizing the sample data, and dividing the homogenized sample data into a training set and a verification set;

s4, establishing a neural network model based on TensorFlow/Playground standard spiral data.

2. The method for identifying the microphase of the microbial rock of the lamp shadow group based on the deep learning of the logging data of claim 1, wherein the step S1 is realized by the following steps: identifying the drilling rock core slice by using a polarizing microscope, determining the microbial rock structure, and further determining the type of the microbial rock and the type of the sedimentary microfacies; and matching the logging data with the type of the microbial rock to establish a sample database.

3. The method for identifying the microphase of the microbial rock of the lamp shadow group based on the deep learning of the logging data of claim 1, wherein the step S2 is realized by the following steps: rapidly traversing and counting all the logging curves by using a built-in function of the Tesorflow/pandas processing module, and checking the integrity of the logging curves; the specific method comprises the following steps: performing numerical gray level conversion on each logging curve, converting the value of a non-mutation pixel point into 0 according to the fuzzy boundary of the image, and converting the value of a mutation pixel point of the fuzzy boundary of the image into 1; the pandas isnull function is then used to determine if the curve is complete.

4. The method for identifying the microphase of the microbial rock of the lamp shadow group based on the deep learning of the logging data of claim 1, wherein the step S3 is realized by the following steps: normalizing/correcting the physical logging parameters of each earth, unifying the logging parameters to 0-1 scale:

wherein, P_iFor scale values after normalization/positification, L_iFor the ith logging parameter, L_minRepresents the minimum of all logging parameters, L_maxIs the maximum of all logging parameters.

5. The method for identifying the microphase of the microbial rock of the lamp shadow group based on the deep learning of the logging data of claim 1, wherein the step S4 is realized by the following steps: a neural network model for a well logging curve is constructed by combining a one-hot/Softmax function interface with a Python/Keras interface, wherein the network model comprises an input layer, an output layer and 4 hidden layers, the hidden layer 1 comprises 600 neurons, and an activation function is elu; hidden layer 2 contains 128 neurons, and Relu is used as an activation function; the hidden layer 3 contains 32 neurons, and sigmoid is used as an activation function; the hidden layer 4 contains 8 neurons, and the activation function uses softsign.

Images