CN115880455A - Three-dimensional intelligent interpolation method based on deep learning - Google Patents

Abstract

The invention provides a three-dimensional intelligent interpolation method based on deep learning, which comprises the following steps: step 1, carrying out quantitative evaluation on characteristics and matching relations of multi-type geological heterogeneous data; step 2, performing intelligent well logging lithology identification based on combination of class-removing non-equalization and deep learning; and 3, performing three-dimensional intelligent interpolation based on deep learning. The three-dimensional intelligent interpolation method based on deep learning is based on an artificial intelligence technology, a large amount of well seismic data and geological research results are effectively fused, results can be provided for geophysical personnel to carry out three-dimensional model construction and reservoir prediction research, and a solid foundation is laid for next research such as determination of favorable reservoirs by the geological personnel, well position design assistance, reserve calculation and the like.

Description

Three-dimensional intelligent interpolation method based on deep learning

Technical Field

The invention relates to the technical field of petroleum exploration, in particular to a three-dimensional intelligent interpolation method based on deep learning.

Background

Big data are influencing our lives, and have great significance for researching new problems and recovering old problems. The key scientific and technical problems involved in the geological science development in the big data era comprise integrated storage management and processing of structured data and semi-structured unstructured data, big data and small data, mixed data and accuracy data, models and data, static exploration models and dynamic monitoring models and the like, combination of data mining and data analysis, unification of correlation and causal relationship, and deep mining and visualization of geological science big data. The key to the analysis of the large geoscience data is to comprehensively analyze the multi-element heterogeneous geoscience data in multiple angles and multiple dimensions. The left kernel analyzes the weak detection anomaly identification mode, and a local RX method is applied to carry out multivariate dimension reduction and extraction of weak and slow geochemical anomalies, so that a better effect can be achieved. Due to the semantic complexity of the spatiotemporal changes of geologic and geologic phenomena, the nonlinearity and uncertainty of the process, the multi-dimensional, multi-scale and real-time nature of the change information. Wu Chonglong proposes integration and utilization of geological space-time big data, and relates to a series of theories, methods and technical problems, mainly comprising an integrated space reference system of geology and geography; coupling the static structure exploration model with the dynamic change monitoring model; data mining, data fusion and intelligent geological big data processing technology. Developing intelligent data processing methods can catch up with the extraordinary growth of big data, so the developed artificial intelligence geology is an important development direction. In the field of artificial intelligence research, many scientific methods and application techniques such as information extraction, cluster analysis, quantitative interpolation, automatic dimension expansion, etc. have been accumulated through long-term practice.

Accurate reservoir lithology prediction and identification are the premise of successful oil and gas exploration and development, and logging is used as an important oil and gas exploration means and can provide reservoir rock physical response information with high resolution and large data volume for effective reservoir lithology identification. The traditional lithology interpretation of well logging is mostly based on professional knowledge and regional experience of interpreters, and because various well logging data have noise interference in the acquisition process and well logging responses of different lithologies have an overlapping phenomenon, the traditional lithology interpretation method of well logging mostly shows that the efficiency is low and the influence of human factors is large. In recent years, machine learning, as an intelligent and efficient data mining method, is increasingly applied to well logging interpretation, and Dubois et al (2007) and Adrielle et al propose applying Neural Network (NN) to realize reservoir lithofacies classification based on well logging data; using a support vector machine algorithm (SVM) to predict a logging reservoir successively by Anazi et al (2010), sebtosheikh et al (2015), zhang Jun et al (2018); (2015) Li et al (2010), shi et al (2015) performed complex lithology log identification by Decision Tree (DT) algorithms. The actual logging data features are often unbalanced, samples of some classes in a data set are far more than samples of other classes, a single model algorithm has certain limitation on logging interpretation, an Ensemble Learning (EL) method can well solve the problem, and Leso et al (2016) and Chamingrui (2017) apply a lifting algorithm in the ensemble learning method to classification in unbalanced data and logging prediction of components of the sandstone debris by a machine learning method; the Random Forest (RF) classification and regression method is respectively used in reservoir prediction and lithofacies classification by Cracknell et al (2013) and Sunday et al (2017). The extreme gradient lifting algorithm (XGboost) is the optimization of the lifting algorithm (Chen and the like, 2016 and the like, 2017), has higher model precision and generalization capability, is added with a regular term for preventing overfitting, supports parallel computation, and is widely applied to the field of image recognition, and the XGboost algorithm is applied to the classification of logging reservoirs by Yan and the like (2019) to obtain a certain effect. However, with the deep exploration degree and the increase of the oil and gas development cost, the inapplicability of the shallow machine learning model in the complex reservoir lithology identification and the higher requirements of the exploration and development on the reservoir lithology prediction precision and reliability form a clear fall. In the face of a more complex and weaker nonlinear mapping relation shown between logging response and geological lithology information in logging interpretation, an intelligent logging lithology interpretation method with stronger data mining capability is urgently needed to be provided so as to realize high-precision identification of complex reservoir lithology.

To visualize geoscience big data, the data can be constructed through geological modeling techniques. Geological modeling requires the transformation of discrete finite spatial sample point data into a continuously viewable geological profile or geologic volume. In some cases, some data are lost in the history process, the acquisition rate is low, and the like, so that the data cannot realize a large data target. And the processing of the geo-sampling point data is divided into two categories: one type is processed into uniform grid structured data or regular grid structured data, so that the realization of stereoscopic vision is facilitated, the algorithm types are rich, and the real-time rendering generation efficiency of the image is higher; the other is to directly process the estimated values at the interpolated points.

Since geological data are often distributed unevenly and in limited quantity, but some connection and regularity still exist between the geological data, "an interpolation method can be selected by combining geological principles and related geological conditions" to obtain continuous geological profiles or bodies through geological data interpolation. It is proposed that data can be interpolated to missing offsets and azimuths by a least squares offset driven 5D interpolation algorithm that de-rasterizes the current subsurface image. The first is Generalized Cross Validation (GCV), which is a guided method of surface prediction error without prior knowledge of the noise level, by extending the existing fast interpolation framework, a true smooth fit can be constructed to very large datasets typically collected in geophysics. In order to consider the balance of multiple aspects such as geological precision requirement, curved surface continuity, data storage capacity and the like, li Ming super et al propose and realize a complex geological curved surface interpolation approximation fitting construction method based on a NURBS technology. The method adopts an NURBS skin interpolation method for the original data which are concentrated and uniformly distributed in the key area of the engineering, so that the curved surface strictly passes through the data points; for the data with discrete distribution in the peripheral area, a NURBS (non-uniform rational B-spline) approximation fitting method is adopted to enable the curved surface to fully approximate the original data under the given precision; and finally, checking, analyzing and adjusting the rationality, the geometry and the precision of the geological structure of the integral curved surface. For a three-dimensional geologic body, three-dimensional interpolation and two-dimensional interpolation have many differences in details, and fitting operations in multiple directions are added. Spatial interpolation is the conversion of discretely distributed borehole survey data into continuous geologic surface data. And interpolating or extrapolating data values in unknown points and areas according to the known data and the mutual relation thereof by using a geological statistical model or a mathematical function method so as to form a continuous curved surface. The Shepard method, delaunay tetrahedron subdivision method and the representative Kriging method among the spatial statistical methods, which are commonly used in the geometric method. However, this method is not highly accurate. The method for carrying out interpolation construction on the curved surface of the geologic body by utilizing the drilling data is the most effective and reliable method for three-dimensional geologic modeling at present. According to the distribution characteristics of the drilling data and the actual situation of the geologic body, a proper and effective interpolation method is selected to ensure the accuracy of the interpolation result.

The prior art is greatly different from the invention, and the technical problem which is to be solved by the invention cannot be solved, so that a novel three-dimensional intelligent interpolation method based on deep learning is invented.

Disclosure of Invention

The invention aims to provide a three-dimensional intelligent interpolation method based on deep learning, which can efficiently and accurately reflect the change rule and the characteristic of a complex lithology space.

The object of the invention can be achieved by the following technical measures: the three-dimensional intelligent interpolation method based on the deep learning comprises the following steps:

step 1, carrying out quantitative evaluation on characteristics and matching relations of the multi-type geological heterogeneous data;

step 2, logging lithology intelligent identification based on combination of class-removing non-equalization and deep learning is carried out;

and 3, carrying out three-dimensional intelligent interpolation based on deep learning.

The object of the invention can also be achieved by the following technical measures:

the step 1 comprises the following steps:

step 1.1: preparing input data;

step 1.2: constructing a multidisciplinary data analysis and knowledge base;

step 1.3: optimizing and constructing reservoir lithofacies logging data;

step 1.4: establishing a petrophysical relationship between the elastic parameters and the reservoir physical parameters;

step 1.5: and establishing a quantitative evaluation formula of geological elements and reservoir lithology and physical parameters.

In step 1.1, importing the collected multi-type geological heterogeneous data into a database, and carrying out preprocessing such as data cleaning and data transformation on various information in the database; and (4) summarizing the geological heterogeneous data according to three-dimensional coordinates so as to facilitate the matching work of the next step.

In step 1.2, comprehensively analyzing the characteristics of multidisciplinary data such as well logging, well drilling, geophysical prospecting and geology, and researching the characterization connotation of heterogeneous data; and constructing a knowledge base of lithology, reservoir skeleton and reservoir physical property by taking the reservoir target geology as a model.

In step 1.3, carrying out geological reservoir and interlayer logging information analysis, and searching the correlation between the reservoir and multiple disciplines; and carrying out principal component analysis on a large amount of logging information and carrying out optimization and correlation analysis based on a long-time and short-time memory neural network curve to realize optimization and construction of lithofacies logging data.

In step 1.4, a petrophysical relationship between the elastic parameters and the physical parameters is established according to an effective medium theory, a self-adaptive theory, a contact theory and an anisotropic model.

In step 1.5, on the basis of establishing various parameter quantitative evaluation formulas through a large number of statistical analyses, fine dissection is carried out on the development area, and the two are organically combined to establish a matching relation between the macro geological elements of the research area and various reservoir parameters so as to form a quantitative evaluation formula.

The step 2 comprises the following steps:

step 2.1: establishing a learning sample;

step 2.2: performing data de-equalization processing based on a MAHAKIL (kluskal) method;

step 2.3: grouping learning samples, wherein a training set is used for model training, a verification set is used for model hyper-parameter adjustment, and the verification set is used for model optimization;

step 2.4: establishing a deep learning network with a proper scale according to the number of learning samples and carrying out network training;

step 2.5: and carrying out logging lithology or lithology intelligent identification on the unknown well by using the optimized deep learning model.

In step 2.1, learning samples are established based on the logging response curve and the well string interpretation result, and each sample comprises logging data of sound wave time difference, volume density, compensation neutrons, natural gamma, deep induction resistivity, depth and interpretation conclusion.

In step 2.2, a MAHAKIL oversampling method is adopted, a propagation process in the simulation genetics is adopted to generate new samples for a few classes, the diversity of the new samples can be ensured to the maximum extent while the non-equalization is removed, and the process is divided into three steps:

(1) separating the few samples from the data set needing to be processed, recording as Cmin, and calculating the Mahalanobis distance of each few sample in the class;

(2) sorting the Cmin according to the Mahalanobis distance, dividing the Cmin into two parts from the median of the sequence numbers, and respectively recording the two parts as Cmin1 and Cmin2;

(3) selecting data samples from Cmin1 and Cmin2 in sequence in pairs respectively, and calculating the mean value of the pairs as a generated new sample;

if the number of new samples generated in one round is not enough to meet the requirement of the highest probability of occupation ratio, the newly generated samples and the samples in Cmin1 and Cmin2 continue to propagate, if the number of new samples generated in one round is not enough, the newly generated samples and the samples in the classes of the two samples continue to propagate, and the like.

The step 3 comprises the following steps:

step 3.1: preparation of input data is performed, with different reservoir parameters including: taking the speed, density, porosity, shale content, fluid saturation or lithology of the well log intelligently identified in the step 2 as input data;

step 3.2: regularizing input data by adopting a random forest regression algorithm;

step 3.3: inputting different longitudinal target layer interval structure interpretation horizons and determining a three-dimensional volume data interpolation frame model;

step 3.4: performing piecewise interpolation based on Delaunay triangulation;

step 3.5: and carrying out three-dimensional intelligent interpolation based on CNN.

Step 3.4 comprises:

(1) taking different reservoir parameters or intelligently identified well logging lithology data as input, constructing a large triangle which comprises all scattered points and putting the large triangle into a triangle linked list;

(2) sequentially inserting scattered points in the point set, finding out a circumscribed circle in the triangular chain table, deleting a triangle which affects the point, connecting all vertexes of the affected triangle in the insertion point, and completing the insertion of one point in the Delaunay triangular chain table;

(3) and optimizing the local newly formed triangles according to an optimization criterion, and putting the formed triangles into the Delaunay triangle linked list.

Step 3.5 comprises:

(1) extraction of training samples: namely, the Delaunay triangle of the segmented interpolation obtained in the step 3.4 is used as an effective label, and the seismic amplitude value in the triangular range is used as a sample characteristic value;

(2) establishing a network architecture: establishing 100 layers of Unet networks, downsampling the first 50 layers, and adding padding operation to input data before convolution in order to prevent losing part of edge information of seismic data in the sampling process; in order to obtain more low-frequency information during up-sampling of the last 50 layers, channel combination is carried out on the down-sampled low-frequency information and the up-sampled high-frequency information;

(3) training a model: dividing sample data according to the proportion of 8;

(4) and (3) outputting a model: when the loss function of the training model meets the requirement, outputting the model;

(5) application of the model: and (4) popularizing and applying the model to a three-dimensional space to obtain a three-dimensional lithologic body.

The three-dimensional intelligent interpolation method based on deep learning is used for reservoir prediction links in petroleum exploration, and comprises prediction services such as three-dimensional geologic body modeling, seismic inversion, lithology recognition and fluid recognition. The method comprises the following steps: the well logging lithology intelligent identification method based on the combination of the class-removing non-equalization and the deep learning can ensure the diversity of new samples to the maximum extent while removing the non-equalization, and the equalized new samples are used for the deep learning to obtain a complex response relation between a well logging response value and well logging lithology, so that more accurate well logging lithology intelligent identification is realized. A multidisciplinary heterogeneous data big data structure is established, a reservoir parameter interpolation method based on deep learning is established on the geological big data structure, the multi-solution property of earthquake prediction is effectively reduced, and the accuracy of reservoir prediction and the success rate of complex oil reservoir exploration are improved. The method is based on the artificial intelligence technology, effectively integrates a large amount of well seismic data and geological research results, and the results can be provided for geophysical personnel to carry out three-dimensional model construction and reservoir prediction research, so that a solid foundation is laid for the geological personnel to determine favorable reservoirs, assist well position design, calculate reserves and other next researches.

The invention provides an intelligent well logging lithology identification method based on combination of class removal non-equalization and deep learning, aiming at the problem that well logging lithology identification or division is represented by the problems of large classification number, insufficient learning samples, uncontrollable classification imbalance and the like, so that the accuracy of well logging lithology identification through machine learning is limited. A MAHAKIL oversampling method is adopted in the learning sample de-equalization process, and new samples are generated for a small number of classes by simulating the propagation process in genetics, so that the diversity of the new samples can be ensured to the maximum extent while the de-equalization is carried out. And (3) the equalized new sample is used for deep learning to obtain a complex response relation between the logging response value and the logging lithology, so that more accurate intelligent identification of the logging lithology is realized.

The method comprises the steps of researching a deep learning network structure with space-time characteristics, constructing an intelligent relation model of seismic data and reservoir parameters based on a deep learning platform, optimizing network coefficients by adopting an intelligent optimizer of the deep learning platform, and constructing an optimized deep learning network of three-dimensional intelligent interpolation of the reservoir parameters such as speed, density, porosity, shale content, fluid saturation and the like.

The system analyzes the advantages and disadvantages of an interpolation algorithm based on geostatistics, analyzes the internal relationship between reservoir parameters and heterogeneous data by combining multidisciplinary heterogeneous data, performs regularization on the heterogeneous data by adopting a random forest regression algorithm based on different disciplinary geoscience models, and provides high-quality input data for three-dimensional intelligent interpolation.

And (3) constructing a deep learning network model considering the space-time relationship, programming a deep learning network algorithm, designing a CNN deep learning method, and finishing three-dimensional intelligent interpolation.

Drawings

FIG. 1 is a flowchart of an embodiment of a deep learning-based three-dimensional intelligent interpolation method of the present invention;

FIG. 2 is a schematic flow chart of the MAHAKIL method according to an embodiment of the present invention;

FIG. 3 is a diagram of a network architecture for lithologic intelligent identification of well logging in accordance with an embodiment of the present invention.

Detailed Description

It is to be understood that the following detailed description is exemplary and is intended to provide further explanation of the invention as claimed. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of the stated features, steps, operations, and/or combinations thereof, unless the context clearly indicates otherwise.

As shown in fig. 1, fig. 1 is a flowchart of a three-dimensional intelligent interpolation method based on deep learning according to the present invention.

1. Quantitative evaluation of characteristics and matching relation of multi-type geological heterogeneous data

The system analyzes the characteristics of multidisciplinary data such as well logging, well drilling, geophysical prospecting, geology and the like, researches the data representation connotation of heterogeneous data geological knowledge, researches the multi-scale matching relation of multidisciplinary data with different resolutions, constructs a heterogeneous data matching method under a multi-scale frame, and constructs a database (knowledge base) of lithology, reservoir type, reservoir skeleton type and reservoir physical property type by taking the reservoir target geology as a model.

And (3) carrying out logging class, rock physics class and seismic attribute class data characteristic analysis by taking the reservoir target as a research object and combining geological knowledge. Aiming at reservoir rock characteristics, reservoir texture, rock components, rock structures, mineral compositions and the like are analyzed, deep excavation of logging information of a complex geologic body reservoir and an interlayer is carried out, the correlation and the change rule of electrical property in the reservoir are analyzed, principal Component Analysis (PCA) is carried out on a large amount of logging information, and optimization and correlation analysis based on long-time and short-time memory neural network curves are carried out, so that optimization and construction of lithofacies logging data are realized.

The method comprises the steps of determining a main structure of complex lithology through petrophysical data aiming at different reservoir types, and establishing a petrophysical relation between elastic parameters and physical parameters through a series of steps according to an effective medium theory, a self-adaptive theory, a contact theory, an anisotropic model and the like. Based on geological knowledge, on the basis of establishing various parameter quantitative evaluation formulas through a large amount of statistical analysis, fine dissection is carried out on a development area, and the matching relation between macro geological elements of a research area and various reservoir parameters is established through organic combination of the two, so that the quantitative evaluation formulas are formed.

2. Logging lithology intelligent recognition based on combination of class-removing non-equalization and deep learning

With the continuous development of the exploration field, the increase of measurement data causes the workload of lithology identification of the traditional manual logging to be greatly increased, and the intellectualization of logging interpretation is realized to a certain extent by the conventional machine learning method. However, the problem of well logging lithology identification or division is represented by a large number of classifications, insufficient learning samples, uncontrollable classification imbalance and the like, and the conventional machine learning method has low prediction accuracy and even represents inapplicability to the problem. Therefore, research on a logging lithology intelligent identification method based on combination of clashing non-equalization and deep learning is needed. A MAHAKIL oversampling method is adopted in the learning sample de-equalization process, and new samples are generated for a small number of classes by simulating the propagation process in genetics, so that the diversity of the new samples can be ensured to the maximum extent while the de-equalization is carried out. And (3) the equalized new sample is used for deep learning to obtain a complex response relation between the logging response value and the logging lithology, so that more accurate intelligent identification of the logging lithology is realized. The well logging lithology deep learning intelligent identification model is established based on learning samples formed by existing well logging lithology interpretation results, namely data pairs formed by various well logging response values and corresponding well logging lithology labels, and the quantity and quality of the samples determine the scale and the final prediction precision of the deep learning model which can be constructed to a certain extent. In the training process, the learning samples are divided into a training set and a verification set, the training set is used for model training, and the verification set is used for hyper-parameter adjustment and model optimization. In actual prediction, a deep learning network with a proper scale is established according to the number of learning samples, network training is carried out, and finally, the optimal deep learning model is used for carrying out intelligent recognition on the well logging lithology.

3. Three-dimensional intelligent interpolation based on deep learning

For different reservoir parameters, including: the method is characterized in that the deep learning network structure with space-time characteristics is researched by speed, density, porosity, shale content, fluid saturation or intelligently identified well logging lithology data. And regularizing the optimized data, and regularizing heterogeneous data by adopting a random forest regression algorithm to provide high-quality input data for three-dimensional intelligent interpolation. On the basis of constructing three triangulation segmented interpolation grids of reservoir parameters such as speed, density, porosity, shale content, fluid saturation and the like or well logging lithology, a three-dimensional intelligent interpolation method based on a CNN convolutional neural network is adopted to complete three-dimensional intelligent interpolation.

Example 1:

in an embodiment 1 to which the present invention is applied, the three-dimensional intelligent interpolation method based on deep learning of the present invention includes the following steps:

step 1, quantitatively evaluating geological heterogeneous data characteristics and matching relationship of three lower sections of sandy turbid sedimentary rock

Step 1.1: input data preparation

And importing the collected geological heterogeneous data of the third segment of the turbid sedimentary rock into a database, and carrying out pretreatment such as data cleaning, data transformation and the like on various information in the database. And (4) generalizing the geological heterogeneous data according to three-dimensional coordinates so as to facilitate the matching work of the next step.

Step 1.2: multidisciplinary data analysis and knowledge base construction

And comprehensively analyzing the characteristics of multidisciplinary data such as well logging, well drilling, geophysical prospecting and geology, and researching the connotation of heterogeneous data representation. And constructing a knowledge base of lithology, reservoir skeleton and reservoir physical property by taking the reservoir target geology as a combined type.

Step 1.3: optimization and construction of nephelometric rock reservoir lithofacies logging data

And (4) analyzing logging information of the nephelometric rock reservoir, the mudstone layer and the gray mudstone layer, and searching correlation between the reservoir and multiple disciplines. And carrying out Principal Component Analysis (PCA) on a large amount of logging information and analyzing curve optimization and relevance based on a long-time memory neural network, so as to realize lithofacies logging data optimization and construction.

Step 1.4: and establishing a petrophysical relationship between the elasticity parameter and the reservoir physical property parameter.

And establishing a petrophysical relationship between the elasticity parameters and the physical parameters of the nephelometric rock according to an effective medium theory, a self-adaptive theory, a contact theory, an anisotropic model and the like.

Step 1.5: and establishing a quantitative evaluation formula of the geological elements, the lithology of the nephelometric rock reservoir and the physical parameters.

On the basis of establishing various parameter quantitative evaluation formulas through a large amount of statistical analysis, fine dissection of a development area is added, and the matching relation between the macro geological elements of the research area and the parameters of the nephelometric rock reservoir is established through organic combination of the development area and the nephelometric rock reservoir to form a quantitative evaluation formula.

Step 2, logging lithology intelligent identification based on combination of class-removing non-equalization and deep learning

Step 2.1: learning sample creation

And establishing learning samples based on the logging response curve and the well string interpretation result, wherein each sample comprises logging data such as acoustic time difference (AC), volume Density (DEN), compensation Neutron (CNL), natural Gamma (GR), deep induction Resistivity (RILD), depth and interpretation conclusion (nepheloid rock type).

Step 2.2: data de-equalization processing based on MAHAKIL method

The method for solving class imbalance mainly comprises three methods, namely undersampling, oversampling and threshold value moving, wherein the undersampling is to acquire a larger class by a little, and the core problem is how to prevent information loss caused by neglecting some samples; the oversampling and the threshold value moving are to add sample points in the minority class set to realize dynamic balance with the majority class set, so that the accuracy of the sample can be ensured to the maximum extent. The threshold shifting method is not mechanical to make the number of the minority classes equal to that of the majority classes, but makes the classification easier and more accurate. The MAHAKIL oversampling method is adopted, a propagation process in genetics is simulated to generate new samples for a few classes, the diversity of the new samples can be ensured to the maximum extent while the de-equalization is carried out, and one of the purposes of the MAHAKIL is to reduce the proportion (Pfp) with high possibility of misjudgment. The process comprises three steps:

(1) the minority samples are separated from the dataset that needs to be processed and are denoted as Cmin, and for each minority sample in the class, the mahalanobis distance is calculated.

(2) And sorting the Cmin according to the Mahalanobis distance, and dividing the Cmin into two parts from the median of the sequence numbers, and respectively marking the two parts as Cmin1 and Cmin2.

(3) Data samples are selected from Cmin1 and Cmin2, respectively, in sequential pairs, and the mean of the pair is calculated as the new sample generated.

If the number of new samples generated in one round is not enough for the requirement of Pfp, the newly generated samples and the samples in Cmin1 and Cmin2 continue to propagate, if the number of new samples generated in one round is not enough, the newly generated samples and the samples in the classes of the two samples continue to propagate, and the like. The algorithm flow diagram is shown in fig. 2, wherein the gray dots represent the newly generated samples for each round.

Step 2.3: learning sample grouping

The training set is used for model training, the verification set is used for model hyper-parameter adjustment, and the verification set is used for model optimization.

Step 2.4: and establishing a deep learning network with a proper scale according to the number of the learning samples and carrying out network training.

Step 2.5: and carrying out logging lithology or lithology intelligent identification on the unknown well by using the optimized deep learning model.

The structure of a well logging lithology intelligent recognition network based on the combination of clashing non-equalization and deep learning is shown in FIG. 3.

Step 3, three-dimensional intelligent interpolation based on deep learning

Step 3.1: input data preparation

Different reservoir parameters of the nephelometric rock include: velocity, density, porosity, shale content, fluid saturation, or lithology of the well log intelligently identified in step 2 as input data.

Step 3.2: regularization of input data

The method mainly adopts a random forest regression algorithm to carry out regularization on input data.

Step 3.3: and inputting different vertical sand layer group structure interpretation horizons and determining a three-dimensional volume data interpolation frame model.

Step 3.4: the process of the piecewise interpolation of the triangulation based on Delaunay comprises three steps:

(1) different reservoir parameters or intelligently identified well logging lithology data are used as input, a large triangle is constructed, all scattered points are contained, and the large triangle is placed into a triangle linked list.

(2) And sequentially inserting scattered points in the point set, finding out a circumscribed circle in the triangular chain table, deleting the common edges of the affected triangles including the inserted points (called affected triangles of the point), and connecting all the vertexes of the affected triangles affected by the inserted points to complete the insertion of one point in the Delaunay triangular chain table.

(3) And optimizing the local newly formed triangles according to an optimization criterion, and putting the formed triangles into the Delaunay triangle linked list.

Step 3.5: CNN-based three-dimensional intelligent interpolation

(1) Extraction of training samples: namely, the Delaunay triangle of the segmented interpolation obtained in step 3.4 is used as an effective label, and the seismic amplitude value in the triangle range is used as a sample characteristic value.

(2) Establishing a network architecture: the disadvantage of creating a 100-layer Unet network, down-sampling the first 50 layers, adding padding to the input data before convolution in order to prevent losing part of the edge information of the seismic data during sampling, and making the effect of losing information or more precisely information of corners or image edges less is impaired. In order to obtain more low frequency information when up-sampling the last 50 layers, the down-sampled low frequency information is channel-combined with the up-sampled high frequency information.

(3) Training a model: sample data was divided according to the ratio of 8.

(4) And (3) outputting a model: and when the loss function of the training model meets the requirement, outputting the model.

(5) Application of the model: and (4) popularizing and applying the model to a three-dimensional space to obtain a three-dimensional nephelometric rock lithosomal body.

Example 2:

in the specific embodiment 2 to which the present invention is applied, the three-dimensional intelligent interpolation method based on deep learning of the present invention includes the following steps:

step 1, quantitatively evaluating geological heterogeneous data characteristics and matching relation of sand four upper-section gravel rock mass

Step 1.1: input data preparation

And importing the geological heterogeneous data of the upper sand gravel rock mass of the four collected sands into a database, and carrying out pretreatment such as data cleaning, data transformation and the like on various information in the database. And (4) generalizing the geological heterogeneous data according to three-dimensional coordinates so as to facilitate the matching work of the next step.

Step 1.2: multidisciplinary data analysis and knowledge base construction

And comprehensively analyzing the characteristics of multidisciplinary data such as well logging, well drilling, geophysical prospecting, geology and the like, and researching the characterization connotation of heterogeneous data. And constructing a knowledge base of lithology, reservoir skeleton and reservoir physical property by taking the target geology of the steep slope gravel rock reservoir as a combination type.

Step 1.3: optimization and construction of lithofacies logging data of gravel rock reservoir

And analyzing logging information of the gravel rock reservoir and the mudstone interlayer, and searching the correlation between the reservoir and multiple disciplines. And carrying out Principal Component Analysis (PCA) on a large amount of logging information and analyzing curve optimization and relevance based on a long-time memory neural network, so as to realize lithofacies logging data optimization and construction.

Step 1.4: and establishing a petrophysical relationship between the elasticity parameter and the reservoir physical property parameter.

And establishing the rock physical relationship between the elastic parameters and the physical parameters of the gravel rock mass according to an effective medium theory, a self-adaptive theory, a contact theory, an anisotropic model and the like.

Step 1.5: and establishing a quantitative evaluation formula of the geological elements, the lithology of the conglomerate reservoir and the physical parameters.

On the basis of establishing various parameter quantitative evaluation formulas through a large amount of statistical analysis, fine dissection of a development area is added, and the matching relation between the macro geological elements of the research area and the gravel rock reservoir parameters is established through organic combination of the two, so that the quantitative evaluation formula is formed.

Step 2, logging lithology intelligent identification based on combination of class-removing non-equalization and deep learning

Step 2.1: learning sample creation

And establishing learning samples based on the logging response curve and the well column interpretation result, wherein each sample comprises logging data such as acoustic time difference (AC), volume Density (DEN), compensation Neutrons (CNL), natural Gamma (GR), deep induction Resistivity (RILD), depth, interpretation conclusion (gravel rock type) and the like.

Step 2.2: data de-equalization processing based on MAHAKIL method

The method for solving class imbalance mainly comprises three methods, namely undersampling, oversampling and threshold value moving, wherein the undersampling is to acquire a larger class by a little, and the core problem is how to prevent information loss caused by neglecting some samples; the oversampling and the threshold value moving are to add sample points in the minority class set to realize dynamic balance with the majority class set, so that the accuracy of the sample can be ensured to the maximum extent. Instead of mechanically equating the number of minority classes to majority classes, the threshold shifting method makes classification easier and more accurate. The MAHAKIL oversampling method is adopted, a propagation process in genetics is simulated to generate new samples for a few classes, the diversity of the new samples can be ensured to the maximum extent while the de-equalization is carried out, and one of the purposes of the MAHAKIL is to reduce the proportion (Pfp) with high possibility of misjudgment. The process comprises three steps:

(1) the minority samples are separated from the dataset that needs to be processed and are denoted as Cmin, and for each minority sample in the class, the mahalanobis distance is calculated.

(2) And sorting the Cmin according to the Mahalanobis distance, and dividing the Cmin into two parts from the median of the sequence numbers, and respectively recording the two parts as Cmin1 and Cmin2.

(3) Data samples are selected from Cmin1 and Cmin2, respectively, in sequential pairs, and the mean of the pair is calculated as the new sample generated.

If the number of the newly generated samples is not enough for Pfp, the newly generated samples and the samples in Cmin1 and Cmin2 continue to be propagated, if the number of the newly generated samples is not enough, the newly generated samples and the samples in the respective classes continue to be propagated, and the like. The algorithm flow diagram is shown in fig. 2, wherein the gray dots represent the newly generated samples for each round.

Step 2.3: learning sample grouping

The training set is used for model training, the verification set is used for model hyper-parameter adjustment, and the verification set is used for model optimization.

Step 2.4: and establishing a deep learning network with a proper scale according to the number of the learning samples and carrying out network training.

Step 2.5: and carrying out logging lithology or lithology intelligent identification on the unknown well gravel rock body by using the optimized deep learning model.

The structure of the well logging lithology intelligent recognition network based on the combination of the classmatic non-equalization and the deep learning is shown in fig. 3.

Step 3, three-dimensional intelligent interpolation based on deep learning

Step 3.1: input data preparation

Different reservoir parameters of a gravel rock mass include: velocity, density, porosity, shale content, fluid saturation, or lithology of the well log intelligently identified in step 2 as input data.

Step 3.2: regularization of input data

The method mainly adopts a random forest regression algorithm to carry out regularization on input data.

Step 3.3: inputting the longitudinal multi-stage gravel rock layer structure interpretation horizon on the sand four, and determining a three-dimensional volume data interpolation frame model.

Step 3.4: the process of the piecewise interpolation of triangulation based on Delaunay comprises three steps:

(1) different reservoir parameters or intelligently identified well logging lithology data are used as input, a large triangle is constructed, all scatter points are contained, and the large triangle is placed into a triangle linked list.

(2) And sequentially inserting scattered points in the point set, finding out a circumscribed circle in the triangle linked list, deleting the common edges of the influenced triangles of the triangle (called the influenced triangle of the point), and connecting all the vertexes of the influenced triangles influenced by the inserted points to finish the insertion of one point in the Delaunay triangle linked list.

(3) And optimizing the local newly formed triangles according to an optimization criterion, and putting the formed triangles into a Delaunay triangle linked list.

Step 3.5: CNN-based three-dimensional intelligent interpolation

(1) Extraction of training samples: namely, the Delaunay triangle of the segmented interpolation obtained in step 3.4 is used as an effective label, and the seismic amplitude value in the triangle range is used as a sample characteristic value.

(2) Establishing a network architecture: the disadvantage of creating a 200-layer Unet network, the first 100 layers of downsampling, adding padding to the input data before convolution in order to prevent losing part of the edge information of the seismic data during sampling, and the loss of information or more precisely the information at corners or image edges playing a smaller role is diminished. In order to obtain more low frequency information when the last 100 layers are upsampled, the downsampled low frequency information and the upsampled high frequency information are channel combined.

(3) Training a model: the sample data is divided according to the proportion of 8.

(4) And (3) outputting a model: and when the loss function of the training model meets the requirement, outputting the model.

(5) Application of the model: and (5) popularizing and applying the model to a three-dimensional space to obtain a three-dimensional gravel rock lithologic body.

Example 3:

in specific embodiment 3 to which the present invention is applied, the three-dimensional intelligent interpolation method based on deep learning of the present invention includes the following steps:

step 1, quantitative evaluation of characteristics and matching relationship of geologic heterogeneous data of a red layer of a Sanxia-Kongshu group step 1.1: input data preparation

And importing the collected geologic heterogeneous data of the red layer of the Sanxia-Kong shou group into a database, and carrying out preprocessing such as data cleaning, data transformation and the like on various information in the data. And (4) generalizing the geological heterogeneous data according to three-dimensional coordinates so as to facilitate the matching work of the next step.

Step 1.2: multidisciplinary data analysis and knowledge base construction

And comprehensively analyzing the characteristics of multidisciplinary data such as well logging, well drilling, geophysical prospecting, geology and the like, and researching the characterization connotation of heterogeneous data. And constructing a lithologic knowledge base, a reservoir framework knowledge base and a reservoir physical property knowledge base by taking the thin mutual reservoir target geology in the red layer as a combined type.

Step 1.3: optimization and construction of lithofacies logging data of gravel rock reservoir

And analyzing logging information of the red-layer thin interbedded layer, the mudstone, the igneous rock and other complex interbedded layers, and searching for correlation between the reservoir and multiple disciplines. And carrying out Principal Component Analysis (PCA) on a large amount of logging information and analyzing curve optimization and relevance based on a long-time memory neural network, so as to realize lithofacies logging data optimization and construction.

Step 1.4: and establishing a petrophysical relationship between the elasticity parameter and the reservoir physical property parameter.

According to an effective medium theory, a self-adaptive theory, a contact theory, an anisotropic model and the like, a petrophysical relationship between the elasticity parameter and the physical property parameter of the red layer thin interbed is established.

Step 1.5: and establishing a quantitative evaluation formula of lithology and physical property parameters of the geological elements and the thin mutual reservoir of the red layer.

On the basis of establishing various parameter quantitative evaluation formulas through a large amount of statistical analysis, fine dissection is carried out on a development area, and the two are organically combined to establish a matching relation between the macroscopic geological elements of a research area and the reservoir parameters of the red horizon thin reservoir so as to form the quantitative evaluation formula.

Step 2, logging lithology intelligent identification based on combination of class-removing non-equalization and deep learning

Step 2.1: learning sample creation

And establishing learning samples based on the logging response curve and the well column interpretation result, wherein each sample comprises logging data such as acoustic time difference (AC), volume Density (DEN), compensated Neutron (CNL), natural Gamma (GR), deep induction Resistivity (RILD), depth, interpretation conclusion (red layer type) and the like.

Step 2.2: data de-equalization processing based on MAHAKIL method

The method for solving class unbalance mainly comprises three methods, namely under-sampling, over-sampling and threshold value moving, wherein the under-sampling is to collect more classes a little bit less, and the core problem is how to prevent information loss caused by neglecting some samples; the oversampling and the threshold value moving are both realized by adding sample points to the minority class set to realize dynamic balance with the majority class set, so that the accuracy of the sample can be ensured to the maximum extent. The threshold shifting method is not mechanical to make the number of the minority classes equal to that of the majority classes, but makes the classification easier and more accurate. A MAHAKIL oversampling method is adopted, a small number of new samples are generated through a reproduction process in the simulation genetics, the diversity of the new samples can be ensured to the maximum extent while the non-equilibrium is removed, and one of the purposes of the MAHAKIL is to reduce the percentage of error judgment (Pfp) with high possibility of error judgment. The process comprises three steps:

(1) the minority samples are separated from the dataset that needs to be processed and are denoted as Cmin, and for each minority sample in the class, the mahalanobis distance is calculated.

(2) And sorting the Cmin according to the Mahalanobis distance, and dividing the Cmin into two parts from the median of the sequence numbers, and respectively marking the two parts as Cmin1 and Cmin2.

(3) Data samples are selected from Cmin1 and Cmin2, respectively, in sequential pairs, and the mean of the pair is calculated as the new sample generated.

If the number of the newly generated samples is not enough for Pfp, the newly generated samples and the samples in Cmin1 and Cmin2 continue to be propagated, if the number of the newly generated samples is not enough, the newly generated samples and the samples in the respective classes continue to be propagated, and the like. The algorithm flow diagram is shown in fig. 2, wherein the gray dots represent the newly generated samples for each round.

Step 2.3: learning sample grouping

The training set is used for model training, the verification set is used for model hyper-parameter adjustment, and the verification set is used for model optimization.

Step 2.4: and establishing a deep learning network with a proper scale according to the number of the learning samples and carrying out network training.

Step 2.5: and carrying out intelligent identification on the lithology or lithology of the log of the red layer and the thin interbed of the unknown well by using the optimized deep learning model.

The structure of a well logging lithology intelligent recognition network based on the combination of clashing non-equalization and deep learning is shown in FIG. 3.

Step 3, three-dimensional intelligent interpolation based on deep learning

Step 3.1: input data preparation

And (3) analyzing different reservoir parameters of the red layer, including: velocity, density, porosity, shale content, fluid saturation, or lithology of the well log intelligently identified in step 2 as input data.

Step 3.2: regularization of input data

The method mainly adopts a random forest regression algorithm to carry out regularization on input data.

Step 3.3: inputting a vertical multi-sand layer group structure interpretation horizon of a sand four-down-hole shop, and determining a three-dimensional volume data interpolation frame model.

Step 3.4: the process of the piecewise interpolation of the triangulation based on Delaunay comprises three steps:

(1) different reservoir parameters or intelligently identified well logging lithology data are used as input, a large triangle is constructed, all scatter points are contained, and the large triangle is placed into a triangle linked list.

(2) And sequentially inserting scattered points in the point set, finding out a circumscribed circle in the triangle linked list, deleting the common edges of the influenced triangles of the triangle (called the influenced triangle of the point), and connecting all the vertexes of the influenced triangles influenced by the inserted points to finish the insertion of one point in the Delaunay triangle linked list.

(3) And optimizing the local newly formed triangles according to an optimization criterion, and putting the formed triangles into a Delaunay triangle linked list.

Step 3.5: CNN-based three-dimensional intelligent interpolation

(1) Extraction of training samples: namely, the Delaunay triangle of the segmented interpolation obtained in step 3.4 is used as an effective label, and the seismic amplitude value in the triangle range is used as a sample characteristic value.

(2) Establishing a network architecture: the disadvantage of creating a 240-tier Unet network, downsampling the first 120 tiers, adding padding to the input data before convolution in order to prevent losing part of the edge information of the seismic data during sampling, and making the effect of losing information or more precisely information of corners or image edges smaller is impaired. In order to obtain more low frequency information when the last 120 layers are upsampled, the downsampled low frequency information and the upsampled high frequency information are channel combined.

(3) Training a model: the sample data is divided according to the proportion of 7.

(4) And (3) outputting a model: and when the loss function of the training model meets the requirement, outputting the model.

(5) Application of the model: and (4) popularizing and applying the model to a three-dimensional space to obtain a three-dimensional red layer thin interbed lithologic body.

Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

In addition to the technical features described in the specification, the technology is known to those skilled in the art.

Claims (13)

1. The three-dimensional intelligent interpolation method based on the deep learning is characterized by comprising the following steps of:

step 1, carrying out quantitative evaluation on characteristics and matching relations of the multi-type geological heterogeneous data;

step 2, performing intelligent well logging lithology identification based on combination of class-removing non-equalization and deep learning;

and 3, carrying out three-dimensional intelligent interpolation based on deep learning.

2. The three-dimensional intelligent interpolation method based on deep learning of claim 1, wherein step 1 comprises:

step 1.1: preparing input data;

step 1.2: constructing a multidisciplinary data analysis and knowledge base;

step 1.3: optimizing and constructing reservoir lithofacies logging data;

step 1.4: establishing a petrophysical relationship between the elastic parameters and the reservoir physical parameters;

step 1.5: and establishing a quantitative evaluation formula of geological elements, reservoir lithology and physical parameters.

3. The three-dimensional intelligent interpolation method based on deep learning of claim 2, wherein in step 1.1, the collected multi-type geological heterogeneous data is imported into a database, and various information in the database is preprocessed by data cleaning and data transformation; and (4) summarizing the geological heterogeneous data according to three-dimensional coordinates so as to facilitate the matching work of the next step.

4. The three-dimensional intelligent interpolation method based on deep learning of claim 2, wherein in step 1.2, multidisciplinary data characteristics of logging, drilling, geophysical prospecting and geology are comprehensively analyzed, and heterogeneous data characterization connotations are researched; and constructing a knowledge base of lithology, reservoir framework and reservoir physical property by taking the reservoir target geology as a model.

5. The three-dimensional intelligent interpolation method based on deep learning of claim 2, wherein in step 1.3, geologic reservoir and interlayer logging information analysis is performed to find correlations between the reservoir and multidisciplinary; and carrying out principal component analysis on a large amount of logging information and carrying out optimization and correlation analysis based on a long-time memory neural network curve, thereby realizing optimization and construction of lithofacies logging data.

6. The deep learning-based three-dimensional intelligent interpolation method according to claim 2, wherein in step 1.4, a petrophysical relationship between the elastic parameters and the physical parameters is established according to an effective medium theory, a self-adaptive theory, a contact theory and an anisotropic model.

7. The three-dimensional intelligent interpolation method based on deep learning of claim 2, wherein in step 1.5, based on the quantitative evaluation formula of various parameters established through a large number of statistical analyses, the matching relationship between the macro geological elements of the research area and various reservoir parameters is established through the organic combination of the fine dissection of the development area and the development area, so as to form the quantitative evaluation formula.

8. The three-dimensional intelligent interpolation method based on deep learning of claim 1, wherein the step 2 comprises:

step 2.1: establishing a learning sample;

step 2.2: carrying out data de-equalization processing based on the MAHAKIL method;

step 2.3: grouping learning samples, wherein a training set is used for model training, a verification set is used for model hyper-parameter adjustment, and the verification set is used for model optimization;

step 2.4: establishing a deep learning network with a proper scale according to the number of learning samples and carrying out network training;

step 2.5: and carrying out logging lithology or lithology intelligent identification on the unknown well by using the optimized deep learning model.

9. The three-dimensional intelligent interpolation method based on deep learning of claim 8, wherein in step 2.1, learning samples are established based on the log response curve and well string interpretation results, and each sample comprises log data of acoustic time difference, volume density, compensated neutrons, natural gamma, deep induction resistivity, depth and interpretation conclusion.

10. The three-dimensional intelligent interpolation method based on deep learning of claim 8, wherein in step 2.2, a MAHAKIL oversampling method is adopted to generate new samples for a small number of classes by simulating a propagation process in genetics, so that the diversity of the new samples can be ensured to the maximum extent while the de-equalization is performed, and the process is divided into three steps:

(1) separating a few classes of samples from a data set needing to be processed, recording the samples as Cmin, and calculating the Mahalanobis distance of each few class of samples in the class;

(2) sorting the Cmin according to the Mahalanobis distance, dividing the Cmin into two parts from the median of the sequence numbers, and respectively recording the two parts as Cmin1 and Cmin2;

(3) selecting data samples from Cmin1 and Cmin2 in sequence in pairs respectively, and calculating the mean value of the pairs as a generated new sample;

if the number of new samples generated in one round is not enough to meet the requirement of the largest probability ratio, the newly generated samples and the samples in Cmin1 and Cmin2 continue to be propagated, if not enough, the newly generated samples and the samples in the respective classes continue to be propagated, and the like.

11. The three-dimensional intelligent interpolation method based on deep learning of claim 1, wherein step 3 comprises:

step 3.1: preparation of input data is performed, taking into account different reservoir parameters, including: taking the speed, density, porosity, shale content, fluid saturation or lithology of the well log intelligently identified in the step 2 as input data;

step 3.2: regularizing input data by adopting a random forest regression algorithm;

step 3.3: inputting different longitudinal target layer interval structure interpretation horizons and determining a three-dimensional volume data interpolation frame model;

step 3.4: performing piecewise interpolation based on Delaunay triangulation;

step 3.5: and performing three-dimensional intelligent interpolation based on CNN.

12. The three-dimensional intelligent interpolation method based on deep learning of claim 11, wherein step 3.4 comprises:

(1) taking different reservoir parameters or intelligently identified well logging lithology data as input, constructing a large triangle which comprises all scattered points and putting the large triangle into a triangle linked list;

(2) sequentially inserting scattered points in the point set, finding out a circumscribed circle in the triangular chain table, deleting a triangle which affects the point, connecting all vertexes of the affected triangle in the insertion point, and completing the insertion of one point in the Delaunay triangular chain table;

(3) and optimizing the local newly formed triangles according to an optimization criterion, and putting the formed triangles into a Delaunay triangle linked list.

13. The three-dimensional intelligent interpolation method based on deep learning of claim 11, wherein step 3.5 comprises:

(1) extraction of training samples: namely, the Delaunay triangle of the segmented interpolation obtained in the step 3.4 is used as an effective label, and the seismic amplitude value in the triangle range is used as a sample characteristic value;

(2) establishing a network architecture: establishing 100 layers of Unet networks, sampling the first 50 layers, and adding padding operation to input data before convolution in order to prevent losing part of edge information of seismic data in the sampling process; in order to obtain more low-frequency information during up-sampling of the last 50 layers, channel combination is carried out on the down-sampled low-frequency information and the up-sampled high-frequency information;

(3) training a model: dividing sample data according to the proportion of 8;

(4) and (3) outputting a model: when the loss function of the training model meets the requirement, outputting the model;

(5) application of the model: and (4) popularizing and applying the model to a three-dimensional space to obtain a three-dimensional lithologic body.

Images