CN113130018A - Lithology identification method based on reservoir element target invariant feature description - Google Patents

Abstract

The invention provides a lithology identification method, and particularly relates to a lithology identification method based on invariant feature description of reservoir meta-target. According to the method, by extracting and describing the invariant features of the logging curves, the invariant features of the meta-target are embedded into a machine learning model for lithology recognition while the automatic layering of the meta-target is realized, finally, lithology prediction is realized in unknown well machine model application, the individuality of a local reservoir is kept, the generalization capability of reservoir description is greatly enhanced, the bottleneck problem that cross-well generalization capability limits introduction of a machine learning method when the lithology recognition is carried out by using the logging curves at present is solved, and the reservoir description precision and reliability are improved.

Description

Lithology Identification Method Based on Reservoir Element Object Invariant Feature Description

technical field

The invention proposes a lithology identification method, in particular to a lithology identification method based on the invariant feature description of a reservoir element target.

Background technique

As most of the old oilfields in my country have entered the middle and late stage of development, the exploration and development of unconventional oil and gas resources has become the most important way to achieve stable production, increase production and prolong production life in old oilfields. Whether it is the exploration and development of unconventional oil and gas resources in old oilfields, or the accurate prediction and development of new oil and gas resources, they are unavoidable problems faced by major oilfield companies in my country today.

Geophysical logging data is one of the most important information sources for obtaining description information of oil and gas reservoir resources. Since the use of different types of geophysical logging curve data sets for target reservoir stratification is an important basis for subsequent lithology identification, logging facies analysis, reservoir division and oil-bearing inter-well parameter prediction, etc., it will directly affect these Therefore, in order to effectively solve the above-mentioned problems, the first task is to be able to describe the geological changes of the target reservoir more accurately. The physical essence of using logging curves for stratification is to divide the target reservoir into multiple sub-layers with the same geological characteristics, which can reduce the amount of data required for analysis of reservoir description and reduce the influence of non-stratigraphic factors to a certain extent. influences. Obviously, accurate and reliable lithology identification is of great practical significance to the exploration and development of oil and gas resources.

At present, domestic and foreign related enterprises and academic research institutions have carried out a large number of researches on reservoir description with logging big data. In terms of theoretical research methods, three main types of methods are currently used: deterministic or stochastic geological modeling and reservoir characterization methods based on geostatistics; classical machine learning methods; ensemble learning/deep learning methods. However, the development speed of classical geological and geophysical technologies is slow and the cost of manpower and time is high, which cannot keep up with the rapid development of communication, storage, and computing power; a single machine learning method has a high risk of overfitting and poor promotion ability; although the integrated learning method has It has a certain vitality, but it cannot fully utilize the growing multi-source heterogeneous well seismic big data; although deep learning methods have demonstrated a strong ability to tap the potential value of related big data in many other fields, the current well seismic data Multi-source heterogeneous big data has not yet reached the applicable conditions of deep learning methods and cannot give full play to its capabilities. However, the actual storage state of oil and gas resources is becoming more and more complex, changing rapidly, and the models are scattered and changeable. Single or simple multi-model descriptions cannot adapt to this actual situation. Especially in the face of the late development of old oilfields in the middle and late stages, and the effective production of new oil and gas resources and reserves, the role of large-scale high-precision logging curves is very prominent. In addition, the statistical characteristics of logging data between different wells often have non-negligible differences, and simple curve normalization processing is difficult to make up for this difference or even bring about new information loss.

Therefore, aiming at the layered properties of thin sand bodies in logging curves, target-oriented multi-geological object samples, and object-oriented effective identification of interwell lithology and lithofacies through the correlation and matching of object-oriented test data and sample data are Oilfields are in urgent need of key technologies that are urgently needed to be conquered.

SUMMARY OF THE INVENTION

The present invention solves the problem that the original spatial amplitude characteristic of the existing logging curve does not have inter-well invariance. The invention provides a lithology identification method based on the invariant feature description of the reservoir element target.

It includes the following steps:

Step S1: Obtain the correlation feature of the reservoir by taking the correlation measure of the adjacent points in the vertical direction for each depth sampling vector, and obtain the corresponding correlation difference feature by taking the difference of the measured distance of the correlation feature, so as to realize the correlation between multiple logging curves. Correlation invariant feature extraction;

Step S2: extracting multi-curve reservoir structure tensor features by performing lateral singular value decomposition on the neighborhood vector set of each depth sampling vector, and extracting vertical local binary mode LBP texture features for each curve, Realize the feature extraction of structure invariance among multiple logging curves;

Step S3: Obtain local statistical features through the microscopic invariant moment features obtained by describing the statistical information of the logging curve data set, and obtain global statistical features with the help of the macroscopic grayscale invariant texture features, and obtain the invariant features between wells, so as to achieve multiple Statistical invariance feature extraction between logging curves;

Step S4: combining the correlation difference feature obtained in step S1 and the tensor feature obtained in step S2 to obtain fine and accurate geological edge stratification points of the reservoir element target, so as to realize automatic stratification;

Step S5: realize the lithology identification of each fine sublayer by fully describing the available information and invariant features that can be obtained from the logging curve data set in the sublayer;

Step S6: use the multi-channel integrated machine learning method to apply the description information obtained in step 5 to perform lithology identification or coding, and build a lithology identification machine learning model;

Step S7: Realizing lithology prediction in the application of the unknown well machine model.

Preferably, the method for extracting the correlation invariant features between multiple logging curves in the step S1 is as follows: the correlation features include: Pearson correlation coefficient and cosine correlation coefficient, which are respectively obtained by the following formula (1) and formula (2) ) to calculate:

Wherein, S _i represents the i-th depth sampling vector on the depth axis; Cov(S _i , S _i-1 ) represents the covariance of adjacent depth sampling vectors; σ(S _i ) represents the standard deviation of the depth sampling vector S _i ;

Distance measures include: Euclidean distance measure, Chebyshev distance measure and urban block distance measure, which are calculated by formulas (3) (4) (5) respectively:

Preferably, in the step S2, the method for extracting the structure-invariant features between multiple logging curves is:

In order to obtain the structural tensor characteristics between logging curves, for the local depth sampling point vector set N(S _i ), perform singular value decomposition on it:

Among them, λ ₁ ≥ λ ₂ is the eigenvalue of the local depth sampling point vector set N(S _i ), then the structure tensor feature of the corresponding depth sampling vector S _i is taken as:

For a given curve X, the LBP texture feature calculation formula is as follows:

Among them, N _i is the local neighborhood of the i-th depth sample point of the logging curve; f _j is a binary code value, and its encoding rules are as follows:

Preferably, the method for extracting statistical invariant features between multiple logging curves in step S3 specifically includes:

For the local depth sampling point vector set N(S _i ), the invariant moment is expressed as:

φ ₁ =η ₂₀ +η ₀₂ (10)

φ ₃ =(η ₃₀ -3η ₁₂ ) ² +(3η ₂₁ -η ₀₃ ) ² (12)

φ ₄ =(η ₃₀ +η ₁₂ ) ² +(η ₂₁ +η ₀₃ ) ² (13)

φ ₆ =(η ₂₀ -η ₀₂ )[(η ₃₀ +η ₁₂ ) ² -(η ₂₁ +η ₀₃ ) ² ]+4η ₁₁ (η ₃₀ +η ₁₂ )(η ₂₁ +η ₀₃ ) (15)

in,

M represents the horizontal dimension of the depth sampling vector set, that is, the number of logging curves; N represents the vertical dimension of the depth sampling vector set, that is, the number of local depth sampling points; N _si (i, j) represents the ith curve in the depth sampling vector set The magnitude of the jth depth sampling point;

The gray-scale co-occurrence invariance texture feature description is given from the log curve to the macro-interaction statistical information; for the differential log curve pair (dX, dY), each curve is quantified or gray-scaled; gray-scale co-occurrence invariance Texture feature expression:

T _GLCM (i)=GLCM(dX(i),dY(i)) (17)

in,

At this time, N(dX=i, dY=j) represents the number of dX=i, dY=j in the same depth sampling;

TN=L*L is the number of all possible grayscale pairs, where L is the number of grayscale levels.

Preferably, in step S4, the correlation difference feature obtained in step S1 and the tensor feature obtained in step S2 are combined to obtain fine and accurate geological edge stratification points of the reservoir element target, so that the method for realizing automatic stratification specifically includes:

Using the correlation difference feature obtained in step S1, the following candidate edge points can be obtained:

Among them, ZCross(dCorr) represents the rising zero-crossing point of the _dCorr correlation difference feature; N(pi) represents the local neighborhood of the current point _{pi; T dCorr is} a threshold constant;

Using the reservoir structure tensor feature obtained in step S2, the following candidate edge points are obtained:

P _Ten ={pi |( _{pi ∈Peak} (Ten))} (19)

Among them, Peak(Ten) represents the peak point of the Ten feature;

The total edge candidates are:

P _EC =∪(P _dCorr , P _Ten ). (20)

Preferably, in step S5, by fully describing the available information and invariant characteristics that can be obtained from the logging curve data set in the sublayer, the method for realizing the lithology identification of each fine sublayer specifically includes:

The available information for describing the logging curve data set in the sublayer, including: the thickness of the sublayer, the relative height information of the curve in the layer, the absolute height information of the curve in the layer, and the shape/morphological information of each curve in the sublayer and the sublayer Inner data and context information of the upper and lower adjacent layers.

Preferably, the thickness information of the small layer can be directly obtained by using the depth difference between the top and bottom of the small layer, namely:

Thick=Depth _bottom -Depth _top (21)

Relative height information of curves within layers:

表示第i个小层的第j条曲线；in,

represents the jth curve of the ith sublayer;

Absolute height information for curves within layers:

The shape/shape information of each curve in the small layer:

Use the structure tensor and LBP texture feature mentioned in step S2 to describe;

Context information of data in a small layer and adjacent layers:

The sub-layer context information is described from the layer thickness contrast relationship, the absolute amplitude relationship, and the layer structure similarity relationship. The layer structure similarity relationship is calculated by the correlation of layer structure tensor, invariant distance and LBP texture feature information.

The lithology identification method based on the invariant feature description of the reservoir element target provided by the invention constructs a lithology machine learning model, successfully solves the problem of large deviation of the statistical characteristics of the existing inter-well logging curve features, and realizes the machine learning model. The interwell generalization ability of the learning model is improved. The method extracts and describes the invariant features of the logging curve, and at the same time realizes the automatic stratification of the meta-target, embeds the invariant features of the meta-target into the machine learning model of lithology identification, and finally uses the machine model of the unknown well. Lithology prediction is realized in the application, which greatly enhances the generalization ability of reservoir description while retaining the local reservoir individuality. The introduced bottleneck problem also improves the accuracy and reliability of reservoir description.

Description of drawings

Fig. 1 is the construction diagram of the invariant feature system of the interwell domain of the logging curve;

FIG. 2 is a schematic flowchart of the lithology identification method based on the description of the invariant characteristics of the reservoir element object according to the first embodiment of the present invention;

3 is an example diagram of a log data correlation characteristic curve of a certain well of the lithology identification method based on the invariant feature description of reservoir elements according to Embodiment 2 of the present invention;

FIG. 4 is an example diagram of a log data correlation difference characteristic curve of a certain well of the lithology identification method based on the invariant feature description of the reservoir element object according to the second embodiment of the present invention;

FIG. 5 is an example diagram of a well logging data structure tensor characteristic curve of the lithology identification method based on the invariant feature description of reservoir elements according to Embodiment 3 of the present invention;

FIG. 6 is an example diagram of the LBP texture characteristic curve of a well logging data of the lithology identification method based on the invariant feature description of the reservoir element object according to the third embodiment of the present invention;

7 is an example diagram of the invariant moment feature of a logging curve set of a well logging curve set according to the lithology identification method based on the invariant feature description of reservoir elements according to Embodiment 4 of the present invention;

FIG. 8 is an example diagram of a logging curve set GLCM feature of a certain well of the lithology identification method based on the invariant feature description of the reservoir element object according to the fifth embodiment of the present invention;

FIG. 9 is a schematic diagram of a complete description of the small layer logging facies characteristics of the lithology identification method based on the description of the invariant characteristics of reservoir elements according to Embodiment 6 of the present invention;

Fig. 10 is a schematic diagram of the quantitative description vector structure of intra-layer data and context-related information of the lithology identification method based on the description of the invariant feature of reservoir element target according to Embodiment 6 of the present invention;

FIG. 11 is a schematic diagram of a lithology identification machine learning model of the lithology identification method based on the description of the invariant characteristics of reservoir elements according to Embodiment 7 of the present invention;

Fig. 12 is the experimental result that the present invention carries out lithology prediction to Jin 98 well;

Figure 13 is a detailed diagram of the experimental results of the present invention's lithology prediction for Well Jin 98;

Fig. 14 is the experimental result that the present invention carries out lithology prediction to Jin 392 well;

Figure 15 is a detailed diagram of the experimental results of the present invention's lithology prediction for Well Jin 392;

Fig. 16 is the experimental result that the present invention carries out lithology prediction to Jin 50 well;

FIG. 17 is a detailed diagram of the experimental results of the lithology prediction of Well Jin 50 according to the present invention.

Detailed ways

Example 1:

The present embodiment will be described below with reference to FIG. 1. The lithology identification method based on the description of the invariant feature of the reservoir element target described in the present embodiment includes the following steps:

Step S1: Obtain the correlation feature of the reservoir by taking the correlation measure of the adjacent points in the vertical direction for each depth sampling vector, and obtain the corresponding correlation difference feature by taking the difference of the measured distance of the correlation feature, so as to realize the correlation between multiple logging curves. Correlation invariant feature extraction;

Step S2: Extracting multi-curve reservoir structure tensor features by performing lateral singular value decomposition on the neighborhood vector set of each depth sampling vector and performing vertical local binary patterns (LBP, Local Binary Patterns) on each curve. ) Texture feature extraction to achieve structure invariance feature extraction between multiple logging curves;

Step S3: Obtain local statistical features through the microscopic invariant moment features obtained by describing the statistical information of the logging curve data set, and obtain global statistical features with the help of the macroscopic grayscale invariant texture features, and obtain the invariant features between wells, so as to achieve multiple Statistical invariance feature extraction between logging curves;

Step S4: combining the correlation difference feature obtained in step 1 and the tensor feature obtained in step S2 to obtain fine and accurate geological edge stratification points of the reservoir element target, so as to realize automatic stratification;

Step S5: realize the lithology identification of each fine sublayer by fully describing the available information and invariant features that can be obtained from the logging curve data set in the sublayer;

Step S6: use the multi-channel integrated machine learning method to apply the description information obtained in step 5 to perform lithology identification or coding, and build a lithology identification machine learning model;

Step S7: Realizing lithology prediction in the application of the unknown well machine model.

Embodiment 2:

A further limitation of the lithology identification method based on the description of the invariant feature of the reservoir element in the first embodiment,

The method for extracting the relevant invariant features in step 1 is as follows:

The traditional logging features do not fully consider the lateral correlation and related transformation information of multiple logging curves at the same depth, and this information just has the invariance ability of inter-well reservoir description.

However, specifically, for a data set composed of multiple curves of a well, since each depth sampling vector is composed of multiple values from different curves, the depth sampling vector of the well logging curve can be longitudinally Reservoir correlation characteristics can be obtained by the correlation measurement of the upper adjacent points. Among them, the correlation features mentioned in this patent mainly include: Pearson correlation coefficient and cosine correlation coefficient. In addition, correlation can also be calculated by calculating vector irrelevance, such as Euclidean distance measure, Chebyshev distance measure, urban block distance measure, etc.

At the same time, after the above-described correlation features are obtained, the corresponding correlation difference features can be obtained by taking the difference. In order to better illustrate the extraction of the relevant invariant features of the logging curve, Figure 1 and Figure 2 respectively give the relevant features and correlation difference characteristic curves of the target reservoir of a well as an example.

Embodiment three:

A further limitation of the lithology identification method based on the description of the invariant feature of the reservoir element in the first embodiment,

The method for extracting the structure invariant features in step 2 is as follows:

Structural invariant features refer to features that remain invariant to geometric transformations for detection or description of local structures. The basic idea is to extract the essential properties of local structures for description. Specifically, the structure-invariant information involved in this patent mainly includes texture feature descriptions such as structure tensor and local binary pattern.

For a logging curve set, for a certain depth sampling vector S _i , let N(S _i ) represent the local neighborhood centered on the depth sampling vector S _i (the neighborhood radius is generally set to be about 0.5 meters). Then the feature extraction of structure tensor can be realized by analyzing the neighborhood vector set of each depth sampling vector. The structure tensor refers to the information that reflects the local structure invariance derived from the gradient change information of the logging curve. By performing singular value decomposition on a local neighborhood, it is possible to find the main direction and coherent information of the direction of numerical changes in the neighborhood.

In addition to using the local depth sampling vector set to extract the tensor features of the multi-curve reservoir structure in the lateral direction, the local vertical structural features of each curve can also be described invariantly. Among them, Local Binary Mode (LBP) is a feature description method for invariant encoding of local structure.

In order to better illustrate the extraction of the structural invariant features of the logging curve, Figure 3 and Figure 4 respectively give an example graph of the logging data structure tensor characteristic curve of a well and an example graph of the LBP texture characteristic curve of the logging data of a certain well. Example.

Embodiment 4:

A further limitation of the lithology identification method based on the description of the invariant feature of the reservoir element in the first embodiment,

The method for extracting the local statistical invariant moment feature in the statistical invariant feature in step 3 is:

Statistical invariant features are purely statistical features obtained by starting with the statistical information of the logging curve data set. These features have good inter-well invariance and have good description ability in logging facies identification. Specifically, the statistical invariant features are divided into local statistical invariant moment features and global statistical features. The global statistical features are obtained through the symbiotic relationship of logging data values.

Moment invariant feature is a feature that is invariant to translation, rotation and scale, which is composed of first-order, second-order and higher-order statistical features of data.

Specifically, for the local depth sampling point vector set N(S _i ), its two-dimensional moment can be defined as:

表示深度采样向量集中第i条曲线的第j个深度采样点的幅值。对应的，的中心矩可表示为：Among them, M represents the horizontal dimension of the depth sampling vector set, that is, the number of logging curves; N represents the vertical dimension of the depth sampling vector set, that is, the number of local depth sampling points;

Represents the magnitude of the jth depth sample point of the ith curve in the depth sample vector set. Correspondingly, the central moment of , can be expressed as:

In order to make the central moment geometrically invariant, Equation (2) can be rewritten as the following normalized central moment form:

Furthermore, the invariant moment, that is, the Hu invariant moment, can be obtained by using the above normalized central moments.

In order to better illustrate the extraction of local statistical invariance features of logging curves, Figure 5 shows an example graph of moment invariant features of a logging curve set for a well as an example.

Embodiment 5:

A further limitation of the lithology identification method based on the description of the invariant feature of the reservoir element in the first embodiment,

The method for extracting the global statistical invariant moment feature in the statistical invariant feature in step 3 is as follows:

The invariant moment information described in the fourth embodiment is proposed from the perspective of local micro-statistical information mining. Obviously, in addition to the mining of statistical information on the microscopic level, reservoir-related information mining can also be performed on the macroscopic level. Therefore, with the help of the idea of Gray-Level Co-occurrence Matrix (GLCM) in image processing, a macroscopic texture feature description can be given from the logging curve to the macroscopic interaction statistics. Specifically, for the differential logging curve pair (dX, dY), quantify (or grayscale) each curve respectively, such as performing 256-level grayscale processing, even if the amplitude of the differential curve is specified to be between 0-255 between. Next, perform the following grayscale co-occurrence matrix calculation:

Among them, N(dX=i, dY=j) represents the number of dX=i, dY=j in the same depth sampling; TN=L*L is the number of all possible grayscale pairs (L is the number of grayscale levels) . After obtaining the gray-level co-occurrence matrix, the gray-level co-occurrence invariant texture feature corresponding to the difference curve pair (dX, dY) can be calculated:

T _GLCM (i)=GLCM(dX(i),dY(i)) (5)

In order to better illustrate the extraction of the global statistical invariance feature of the logging curve, Figure 6 shows an example graph of the GLCM feature of a logging curve set for a well as an example.

Embodiment 6:

A further limitation of the lithology identification method based on the description of the invariant feature of the reservoir element in the first embodiment,

In step 5, by fully describing the available information and invariant characteristics that can be obtained from the logging curve data set within the sublayer, the method for realizing the lithology identification of each fine sublayer is as follows:

The available information for describing the logging curve data set in the sublayer, including: the thickness of the sublayer, the relative height information of the curve in the layer, the absolute height information of the curve in the layer and other basic physical properties; the shape/shape of each curve in the sublayer information; and the data in the small layer and the context information of the upper and lower adjacent layers, etc., as shown in FIG. 7 .

(1) Layer thickness information: It can be directly obtained by using the depth difference between the top and bottom of the small layer, namely:

Thick=Depth _bottom -Depth _top (6)

(2) Relative height information of curves in layers:

表示第i个小层的第j条曲线；in,

represents the jth curve of the ith sublayer;

(3) The absolute height information of the curve in the layer:

(4) In-layer curve shape/shape information:

It is described by the method proposed in step 2 - structure tensor and LBP texture feature.

(5) The context information of the data in the layer and the adjacent layers:

The sub-layer context information is described from the aspects of layer thickness contrast relationship, absolute amplitude relationship, layer structure similarity relationship, etc. The layer thickness contrast information and absolute amplitude relationship are the values of layer thickness and absolute amplitude recorded in this layer and adjacent layers. ; The similarity relation of layer structure can be calculated by the correlation of information such as layer structure tensor, invariant distance and LBP texture. To this end, a quantitative description vector of intra-layer data and context-related information as shown in FIG. 8 can be obtained.

Embodiment 7:

A further limitation of the lithology identification method based on the description of the invariant feature of the reservoir element in the first embodiment,

In step 6, the description information obtained in step 5 is applied by means of multi-channel integrated machine learning for lithology identification or coding, and the method for constructing a lithology identification machine learning model is as follows:

After a complete description of the information, multiple basic learning machines are used to form a multi-channel integrated machine learning, to construct the meta-features of the logging curve, and to optimize its prediction results through meta-machine learning and formation geological knowledge. The schematic diagram of the machine learning recognition framework constructed by the present invention is shown in FIG. 9 .

In order to verify the effectiveness of the method of the present invention, the method was verified by using 24 wells within the Jin 3 to Jin 392 working areas in the Qijia working area in the central sag of the Songyuan Basin. The experimental results show that the prediction result of the method of the present invention has higher precision than the existing conventional logging interpretation, is closer to the core lithology, can accurately reflect the change of the sand-mud ratio of the small layer, and has a good promotion effect.

Among them, 11 wells including Jin 3, Jin 12, Jin 23, Jin 27, Jin 31, Jin 37, Jin 38, Jin 54, Jin 58, Jin 80, and Jin 391 were randomly selected as model wells. 10 wells such as Jin 30, Jin 34, Jin 40, Jin 45, Jin 51, Jin 55, Jin 56 and Jin 59 are used as model calibration wells, and 3 wells such as Jin 50, Jin 98 and Jin 392 have fine lithology described by cores Wells are used as test wells.

In the selection of logging curves, 7 conventional curves of GR, DEN, SP, CAL, AC, LLD and LLS are selected, and the target lithology is: mudstone, siltstone, argillaceous siltstone, silty mudstone, fine sandstone, ossomorphic Insect layer, oil shale. Figures 6-11 show the overall and detailed results of lithology prediction for the three test wells, respectively.

However, the experimental results also show that this method also has certain shortcomings. Since the proportion of fine sandstone in the training sample and the test sample is very small, the prediction effect is not good. By focusing on the detection of oil shale, the effective detection rate is guaranteed. , the false detection rate is increased. To this end, in the follow-up research, it is necessary to increase the processing capacity of small samples such as fine sandstone and oil shale to further enhance the system performance.

The feature of the present invention is to construct an invariant feature system of inter-well domain that can effectively improve the logging curve formation information expression capability as shown in FIG. 1, and based on this invariant feature system, construct a The machine learning model realizes lithology identification.

In order to solve the problem that the original spatial amplitude characteristics of the existing logging curves do not have inter-well invariance, the present invention takes full excavation of the invariant characteristics that can effectively characterize the reservoir characteristics as a starting point, and designs a combination of correlation invariance characteristics and structural invariance characteristics. The log curve invariant feature system composed of characteristics and statistical invariant features solves the problem of inconsistent log curve characteristics between wells to a certain extent, and can express invariantly the essential properties of the target reservoir for subsequent processing and analysis. Foundation.

Aiming at the problem of large deviation of the statistical characteristics of the existing inter-well logging curve features, the present invention proposes an invariant feature description method oriented to the reservoir element target, and constructs a logging curve inter-well domain invariant feature system. At the same time, in order to better To achieve the task of lithology identification, the available information that can be obtained from the small-layer logging curve data set is fully described. On the basis of the two, the integrated machine learning model is used to achieve accurate lithology identification of geological reservoirs. . To this end, the method of the present invention constructs a logging curve inter-well domain invariant feature system, and conducts the logging curve inter-well domain invariant feature from three aspects: correlation invariant feature, structural invariant feature and statistical invariant feature. Therefore, it effectively solves the problem of using the intrinsic properties of the target reservoir for invariant expression, and realizes a method that effectively utilizes the invariant features of the reservoir meta-target and multi-channel integrated machine learning to realize lithology identification. At the same time, the generalization ability of reservoir description is greatly enhanced.

In order to solve the bottleneck problem that the cross-well generalization ability restricts the introduction of the machine learning method when using the logging curve to identify the lithology, the present invention proposes a lithology identification method based on the invariant feature description of the reservoir element target. The method adapted to the present invention has good adaptability and robustness in the application of lithology identification using logging curves of different wells in different regions. Compared with the conventional methods of qualitative analysis such as logging facies analysis and simple curves such as local spatial data amplitude characteristic curves for reservoir description, the method of the present invention greatly improves the accuracy and reliability of target reservoir description. Fig. 2 presents a structural block diagram of the method of the present invention. The key technical content of the present invention includes two parts: establishing the invariant characteristic system of the inter-well domain of the logging curve and the complete expression of the knowledge of the inter-well sub-layer.

The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited to this, any person familiar with the technology can understand the transformation or replacement within the technical scope given by the present invention. , should be covered within the protection scope of the claims of the present invention.

Claims (7)

1. A lithology identification method based on reservoir element target invariant feature description is characterized by comprising the following steps:

step S1: obtaining reservoir correlation characteristics by taking correlation measurement of adjacent points in the longitudinal direction of each depth sampling vector and obtaining corresponding correlation difference characteristics by taking difference of measure distances of the correlation characteristics, thereby realizing the extraction of correlation invariance characteristics among multiple logging curves;

step S2: extracting the tensor features of the multi-curve reservoir structure by performing horizontal singular value decomposition on a neighborhood vector set of each depth sampling vector and extracting the local binary pattern LBP texture features of each curve in the longitudinal direction to extract the structural invariance features among the multi-logging curves;

step S3: obtaining local statistical characteristics through microscopic invariant moment characteristics obtained by describing statistical information of a logging curve data set, obtaining global statistical characteristics by means of macroscopic gray level symbiotic invariance texture characteristics, obtaining inter-well invariance characteristics of the local statistical characteristics, and realizing extraction of the statistical invariance characteristics among multiple logging curves;

step S4: obtaining a precise and accurate geological edge layering point of the reservoir element target by combining the related difference characteristics obtained in the step S1 and the tensor characteristics obtained in the step S2, thereby realizing automatic layering;

step S5: the lithology recognition of each fine and fine layer is realized by completely describing available information and invariant features which can be obtained by a logging curve data set in the small layer;

step S6: performing lithology recognition or coding by using the description information obtained in the step five in a multi-channel integrated machine learning mode, and constructing a lithology recognition machine learning model;

step S7: lithology prediction is implemented in unknown well machine model applications.

2. The lithology identification method based on the reservoir meta target invariant feature description as claimed in claim 1, wherein the method for extracting the relevant invariant features among the multiple well logging curves in the step S1 is as follows: the correlation features include: pearson correlation coefficient and cosine correlation coefficient, which are calculated by the following equation (1) and equation (2), respectively:

wherein ,S_iRepresenting the ith depth sample vector on the depth axis; cov (S)_i,S_i-1) Representing the covariance of neighboring depth sample vectors; sigma (S)_i) Representing a depth sampling vector S_iStandard deviation of (d);

the distance measure includes: euclidean distance measure, chebyshev distance measure, and city block distance measure, which are calculated by formulas (3) (4) (5), respectively:

3. the lithology identification method based on reservoir element target invariant feature description as claimed in claim 1, wherein the method for extracting the structure invariant features among the multiple well logging curves in step S2 is as follows:

to obtain structure tensor features between log curves, a set of vectors N (S) is sampled for local depths_i) And carrying out singular value decomposition on the obtained product:

wherein ,λ₁≥λ₂For a set of local depth sample point vectors N (S)_i) Then corresponding to the depth sample vector S_iThe structure tensor characteristics of (a) are taken as:

for a given curve X, the LBP texture feature calculation formula is as follows:

wherein ,N_iA local neighborhood of the ith depth sampling point of the logging curve is obtained; f. of_jThe code rule is as follows:

4. the lithology identification method based on reservoir element target invariant feature description as claimed in claim 1, wherein the method for extracting the statistical invariant features among the multiple well logs in step S3 specifically comprises:

vector set N (S) for local depth sampling points_i) The invariant moment is expressed as:

φ₁＝η₂₀+η₀₂ (10)

φ₃＝(η₃₀-3η₁₂)²+(3η₂₁-η₀₃)² (12)

φ₄＝(η₃₀+η₁₂)²+(η₂₁+η₀₃)² (13)

φ₆＝(η₂₀-η₀₂)[(η₃₀+η₁₂)²-(η₂₁+η₀₃)²]+4η₁₁(η₃₀+η₁₂)(η₂₁+η₀₃) (15)

wherein ,

m represents the transverse dimension of the depth sampling vector set, namely the number of logging curves; n represents the longitudinal dimension of the depth sampling vector set, namely the number of local depth sampling points;

representing the amplitude of the jth depth sampling point of the ith curve in the depth sampling vector set;

giving out gray level symbiotic invariance texture feature description from the aspect of logging curve to macroscopic interaction statistical information; for the differential logging curve pair (dX, dY), quantizing or graying each curve respectively; gray level symbiotic invariance texture characteristic expression:

T_GLCM(i)＝GLCM(dX(i),dY(i)) (17)

wherein ,

in this case, N (dX ═ i, dY ═ j) indicates the number of dX ═ i, dY ═ j in the same depth sample;

TN L is the number of all possible pairs of gray levels, where L is the number of gray levels.

5. The lithology identification method based on the invariant feature description of the reservoir element target as claimed in claim 1, wherein the step S4 combines the correlation difference features obtained in step S1 and the tensor features obtained in step S2 to obtain the precise geological edge layering points of the reservoir element target, so as to implement the automatic layering method specifically comprising:

using the correlation difference features obtained in step S1, the following candidate edge points can be obtained:

wherein, ZCross (dCorr) represents the rising zero crossing of the correlation difference characteristic of dCorr; n (p)_i) Indicates the current point p_iA local neighborhood of; t is_dCorrIs a threshold constant;

using the tensor features of the reservoir structure obtained in step S2 to obtain the following candidate edge points:

P_Ten＝{p_i|(p_i∈Peak(Ten))} (19)

wherein peak (Ten) represents peak points of Ten characteristics;

the total edge candidate points are:

P_EC＝∪(P_dCorr,P_Ten) (20) 。

6. the lithology identification method based on the reservoir meta-target invariant feature description as claimed in claim 1, wherein the method for identifying the lithology of each fine and fine layer by completely describing the available information and invariant features that can be obtained by the logging curve data set in the small layer in step S5 specifically comprises:

information available for describing a well log data set within a sub-zone, including: the thickness of the small layer, the relative height information of the curve in the layer, the absolute height information of the curve in the layer, the shape/form information of each curve in the small layer, and the context information of the data in the small layer and the upper and lower adjacent layers.

7. The lithology identification method based on the reservoir meta target invariant feature description according to claim 1, comprising:

wherein, the thickness information of the small layer: the depth difference of the top and the bottom of the small layer can be directly obtained, namely:

Thick＝Depth_bottom-Depth_top (21)

relative height information of curves within layers:

wherein ,

a jth curve representing an ith sublayer;

absolute height information of the curve within the layer:

shape/shape information of each curve in the small layer:

describing by using the structure tensor and LBP texture characteristics mentioned in step S2;

context information of data in small layer and adjacent layer:

and describing the context information of the small layer from the layer thickness contrast relation, the absolute amplitude relation and the layer structure similarity relation, wherein the layer structure similarity relation is calculated by utilizing the layer structure tensor, the invariant distance and the correlation of the LBP texture feature information.

Images