CN116975669A - Automatic layering method for logging curves based on constraint integrated clustering - Google Patents

Abstract

An integrated clustering automatic layering method of logging curves based on energy constraint aims to solve the problem that an existing logging curve automatic layering algorithm is difficult to be qualified for a single-sand-level unconventional oil gas resource fine description task. According to the method, geological description invariant features are constructed by utilizing logging curves to effectively excavate inter-well multi-curve invariant associated information, and further, under the support of the invariant features, an unsupervised integrated clustering region growth model with layering energy constraint is designed, so that automatic layering of fine description industrial application requirements of a reservoir can be achieved, and the method has remarkable popularization potential and can be used for fine description of a follow-up unconventional oil and gas reservoir.

Description

Automatic layering method for logging curves based on constraint integrated clustering

Technical Field

The invention belongs to the field of petroleum engineering, and particularly relates to an automatic layering method of a logging curve based on constraint integrated clustering.

Background

Fine reservoir descriptions are one of the key challenges in achieving efficient mobilization of unconventional hydrocarbon resources. Logging technology utilizes sensors of different physical mechanisms such as electricity, magnetism, sound, radioactivity and the like to acquire continuous response data about a target reservoir, and is the most main data for realizing reservoir description. Reservoir description based on logging curves can realize effective depiction of geological conditions of a target reservoir through interpretation of logging data, and objective cognition of typical geological modes such as lithology, physical properties, oil-gas properties and the like of the target reservoir is formed, so that effective support for design of exploration and development schemes is realized.

The key step of reservoir description by using logging data is to realize automatic layering of a target reservoir, and in the process of oil and gas resource exploration and development in the past decades, various logging curve automatic layering methods aiming at conventional oil reservoirs are formed and mainly divided into the following three types of methods: (1) Statistical methods based on classical geological geophysical techniques; (2) a machine learning method; (3) an integration/deep learning method.

The geostatistical-based method has relatively reliable theoretical support and physical sense support, and reservoir information quantitative analysis of geological geophysical technology typified by cross plots in amplitude space is the most common method for current geological interpreters. However, classical geological geophysical techniques exhibit a strong prior dependence in the actual production process and have poor parameter applicability. In recent years, as machine learning technology becomes a popular tool for various research fields, a reservoir description method based on machine learning is increasing. The machine learning method has good data adaptability and large development space. However, the single machine learning method has the problem of overfitting to training data, and meanwhile, the parameter generalization capability, the applicability and the popularization are low, and the performance after integration is generally superior to that of a single clustering method and has better stability. Although deep learning has achieved some success, the quality and effective labels of current practical application data are somewhat unsatisfactory for deep learning.

Various clustering-based methods are applied to the current research to perform automatic layering of a target reservoir by using logging data, however, in the application of layering of a thin sand body stratum, the methods do not effectively solve the existing problems, and the problems that the unconventional reservoir cannot realize fine automatic layering and the like still exist. Based on the analysis, the invention provides an energy constraint-based logging curve integrated clustering automatic layering model: firstly, constructing an invariable characteristic system with high inter-well consistency by utilizing original multi-attribute logging data; and then, an integrated clustering region growth algorithm based on layered energy constraint is designed by utilizing the built logging data invariant features so as to achieve the purpose of fine description of the final unconventional oil and gas reservoir.

Disclosure of Invention

The invention aims to analyze logging data by using an integrated clustering method to obtain automatic reservoir division. Therefore, the invention provides a logging curve automatic layering method based on constraint integrated clustering, which comprises the following steps:

step S1: the method for analyzing the association rule is adopted to realize the effective mining of the association rule among the input multi-attribute logging data;

step S2: screening logging data by using the association rule established in the first step, and extracting relevant difference features and tensor features of the screened data;

step S3: acquiring a zero crossing point of the correlation difference characteristic and a tensor characteristic peak point in the second step, and taking the zero crossing point and the tensor characteristic peak point as seed points for region growth;

step S4: generating a plurality of clustering results by adopting a k-means clustering and spectral clustering method, constructing NMI matrixes based on all the clustering results, and selecting cluster members meeting a certain degree of difference;

step S5: performing integrated clustering on the cluster members obtained by screening in the step four, arranging clustering classification results according to the depths of the wells, and taking the classification mutation points as reservoir candidate layering points;

step S6: combining the 'pseudo layers' by using the region growing process based on energy constraint for the candidate layered points in the step five;

step S7: and arranging the results after the region growth according to the depths of the wells, and taking the result mutation points as final reservoir layering points to realize fine layering of the reservoir.

The beneficial effects of the invention are that

The method is characterized in that after simple rock stratum layering is realized by utilizing integrated clustering to cluster logging data, a model of regional growth based on layering energy constraint is constructed, so that the problem of pseudo layering after the integrated clustering can be effectively reduced, and the accuracy of lithologic reservoir layering is obviously improved.

Preferably, in the step S4, a k-means clustering and spectral clustering method is adopted to generate a plurality of clustering results, and the method for constructing the canonical mutual information NMI matrix based on all the clustering results is as follows:

for one logging data set, clustering is carried out by adopting a K-Means and a spectral clustering method, wherein the clustering times are T times (T is a positive integer, and the values are 1,2, … and N). Two kinds of clustering result sets omega= { G can be obtained respectively _ij ,i＝1,2,3…,T；j＝1,2,3…,T},G _ij And (3) representing the jth class result of the ith clustering algorithm, wherein i and j are positive integers, and the values are 1,2, … and T. For each algorithm, an NMI matrix of T-degree results can be constructed by analyzing canonical mutual information NMI values of each clustering result. Taking two clustering results G _a and G_b P (a, b) (where a, b are positive integers, values 1,2, …, t.), and the edge distribution is p (a) and p (b), the mutual information MI between the two can be defined as:

correspondingly, normalized mutual information NMI is introduced:

wherein ,H(G_a) and H(G_b ) Respectively represent G _a and G_b Entropy values of two random variables, wherein the entropy values are defined as follows:

NMI _ab the larger the value of (c) is, the smaller the difference between the clustering results is, and when the two clusters are completely equal, the value is 1. Based on NMI measure, NMI interaction matrix between different clustering results can be established.

Next, summing the NMI matrix to obtain sum vector V of NMI values of different clustering results and other clustering results _sum ：

wherein ,is the sum of NMI values of i clustering results. In order to select a result with reliable clustering result precision and a certain divergence between results for final integration, the following selection scheme may be used to select a clustering result Ids for final integration analysis:

wherein ,V_low and V_high Respectively represent the order V _sum In (a)Divide into two parts and->A value of the split bit.

After the clustering result for integration is selected, counting the times that each sample point is at the clustering edge, obtaining lithology interface candidate boundary points given by the integration clustering method through threshold processing, and completing assignment of clustering labels among the boundary points through alignment of the clustering result.

Preferably, the modeling method for constructing the region growth based on the layered energy constraint in the step S6 is as follows:

the growth criterion is the core of region growth, and the invention converts the problem of region growth into a data segmentation problem of minimum layering energy iteration by introducing layering energy constraint so as to realize the purpose of automatic layering. To log data sets(K is a positive integer, the maximum value is the number of sample points of the logging curve set, n represents the number of the whole logging data set), the corresponding integrated clustering result and the region growth seed point are region growth basic data, and a layered energy function E (L) in each iteration process is defined as follows:

wherein P is a certain sample point in the integrated clustering result, P is the integrated clustering result,representing data cost components, D _p (L _p ) Representing point p to label L _p The Euclidean distance of the corresponding cluster center;representing a smoothed cost component for penalizing the tag discontinuity; w (w) _pq ＝αe ^-d(p,q) Is the weight assigned to the point pair (p, q) in the same neighborhood with different labels, d (p, q) is the Euclidean distance of the point pair (p, q), δ (L) _p ≠L _q ) Indicating the function discontinuously for labels, i.e. at L only _p ≠L _q When the value is 1, α=max (max (d (p, L _p )),max(d(q,L _q )))，max(d(p,L _p ) P to the intra-neighborhood label L) _p The maximum euclidean distance of the points, max (d (p, L _p ) For p to the label L in the neighborhood _p The maximum distance of the points, max (d (q, L _q ) Indication)q to in-neighborhood label L _q The maximum distance of the points, max (d (P, L _p ) For P to L in the neighborhood label _p The maximum distance of the points.

After the construction of the energy function is completed, the cluster labels of the integrated clusters are used as initial segmentation, region growth analysis is carried out from the region growth seed points, region growth is carried out iteratively under the constraint of the minimization of the energy function, and the automatic layering task of the logging data can be completed.

The invention provides an automatic clustering layering method for logging curve integration based on layering energy constraint. The universality of the algorithm among different data statistics characteristic wells is effectively ensured by utilizing association rule mining and invariant feature extraction; candidate edge seed points are given by combining integrated cluster analysis with invariant features, and the fine layering of a target reservoir is finally realized by utilizing reservoir layering sensitive logging data through considering the growth processing of multi-level energy constraint areas of global and local segmentation. The method has the advantages of few super parameters, high efficiency and strong robustness, and has better popularization potential in practical application of various reservoir fine descriptions.

Drawings

FIG. 1 is a schematic flow chart of an automatic layering method of logging curves based on constraint integrated clustering.

Fig. 2 is a schematic diagram of a NMI matrix composition of a clustering result according to a fourth embodiment of the present invention.

Fig. 3 is an NMI interaction matrix of 100K-means and spectral clustering methods according to a fourth embodiment of the present invention.

FIG. 4 is a visual view of quantitative evaluation of an A1 well as an example using the automatic layering method of the logging curves based on the constraint integrated clustering and other comparison methods of the present invention.

FIG. 5 shows the results of an automatic layering experiment using an A1 well as an example, by using the automatic layering method of the logging curves based on the constraint integrated clustering.

FIG. 6 is an example of an automated stratification test result for an A2 well using the constraint-integrated-clustering-based automatic stratification method of log curves of the present invention.

Detailed Description

Embodiment one:

referring to fig. 1, the method for automatically layering logging curves based on constraint integrated clustering according to the present embodiment is described, and includes the following steps:

step S1: the method comprises the steps of adopting a correlation rule analysis mode to realize effective mining of the correlation rule among the input multi-attribute logging data;

step S2: screening logging data by using the association rule established in the first step, and extracting relevant difference features and tensor features of the screened data;

step S3: acquiring a zero crossing point of the correlation difference characteristic and a tensor characteristic peak point in the second step, and taking the zero crossing point and the tensor characteristic peak point as seed points for region growth;

step S4: generating a plurality of clustering results by adopting a k-means clustering and spectral clustering method, constructing NMI matrixes based on all the clustering results, and selecting cluster members meeting a certain degree of difference;

step S5: performing integrated clustering on the cluster members obtained by screening in the step four, arranging clustering classification results according to the depths of the wells, and taking the classification mutation points as reservoir candidate layering points;

step S6: combining the 'pseudo layers' by using the region growing process based on energy constraint for the candidate layered points in the step five;

step S7: and arranging the results after the region growth according to the depths of the wells, and taking the result mutation points as final reservoir layering points to realize fine layering of the reservoir.

Embodiment two:

further defining the automatic layering method of logging curves based on constraint integrated clustering according to the first embodiment,

the method for effectively mining the association rule among the input multi-attribute logging data by adopting the association rule analysis mode in the first step comprises the following steps:

in this embodiment, the association rule is applied to mining the association relationship between different logging attributes, and therefore, a discretization level of 5 data processing as shown in table 1 is required to be performed on the logging curve.

Table 1 log data discretized coded representation example

Based on the discrete coding data set, the association relation among the multi-attribute logging curves in the logging data set is mined by utilizing a classical Apriori algorithm. The establishment of the association rule between logging data mainly comprises the following two steps: firstly, finding out frequent sets of all logging data according to a minimum support threshold; and secondly, generating association rules according to the frequent set and the minimum confidence threshold. By mining the association relation, the full utilization of reliable association attribute data can be realized.

Embodiment III:

further defining the automatic layering method of logging curves based on constraint integrated clustering according to the first embodiment,

the extraction method of the correlation difference features and tensor features in the second step comprises the following steps:

specifically, for a log data set s= { S composed of K log curves of N samples _i ∈R ^N×1 } _i＝1,…,K For a certain depth sampling vectorThe correlation coefficient of the sample is calculated as follows:

wherein ,S^j-1 Representing depth sample vector S ^j The last adjacent sampling point in the depth direction, cov (S ^j ,S ^j-1 ) Representing covariance of adjacent depth sample vectors; sigma (S) ^j ) Representing depth sample vector S ^j Standard deviation of sigma (S) ^j-1 ) Representing depth sample vector S ^j-1 Standard deviation of (2). In order to emphasize the significance of mutation of the related characteristic values, the related difference characteristic can be obtained by carrying out first-order difference on the related characteristic:

dCorr(S ^j )＝Corr(S ^j )-Corr(S ^j-1 ) (8)

wherein Corr (S) ^j-1 ) For depth sample vector S ^j-1 Is used for the correlation characteristic value of (a). If S ^j Inside a certain lithology unit, dCorr (Sj) is approximately equal to 0, which indicates that the point and the neighborhood point are in a certain lithology smooth area; if S ^j-1 At the upper lithology interface, i.e. the transition from one continuous lithology unit to the next, there is a dCorr (S) ^j ) < 0; similarly, if S ^j-1 At the lower lithology interface, i.e. from the lithology interface into the next successive lithology unit, there is dCorr (S) ^j ) > 0. The rapid change from the continuous trough to the crest forms a Z-shaped lithology interface mark, namely a zero crossing point of a Z-shaped transition zone, and the positioning of lithology unit junction points can be realized by utilizing the zero crossing point of the correlation difference characteristic.

Texture structure descriptions based on structure tensors are expressed by eigenvalues of a common local data covariance matrix, and tensor eigenvalues change greatly when successive reservoir lithology changes. In particular, for a well-logging dataset,expressed as S ^j Is a local neighborhood (N _Nbr Representing the local neighborhood sample number) data set, then one can determine the local neighborhood sample number for N (S ^j ) Singular value decomposition is carried out to obtain:

wherein ,λ₁ ≥λ ₂ ≥L≥λ _K For the sorted feature values, the depth sampling vector S is correspondingly _i The structure tensor feature of (c) can be taken as:

wherein when lambda is ₁ ≈λ ₂ At the time, the structure tensor is takenThe value is smaller, which indicates that the structural consistency at the point is stronger, namely, the point is positioned in a smooth area inside the lithology unit; when lambda is ₁ >>λ ₂ At this point, the structure Zhang Liangqu value will increase dramatically, indicating that the lithology of the point is abrupt and the sample point belongs to lithology unit interface points. For this purpose, the location of lithology unit junctions can be achieved with maxima points of the structure tensor.

Embodiment four:

further defining the automatic layering method of logging curves based on constraint integrated clustering according to the first embodiment,

the construction method and the clustering result generation method for the NMI matrix provided in the step four are as follows:

for one logging data set, clustering is carried out by adopting a K-Means and a spectral clustering method, wherein the clustering times are T times (T is a positive integer, and the values are 1,2, … and N). Two kinds of clustering result sets omega= { G can be obtained respectively _ij ,i＝1,2,3…,T；j＝1,2,3…,T},G _ij And (3) representing the jth class result of the ith clustering algorithm, wherein i and j are positive integers, and the values are 1,2, … and T. For each algorithm, an NMI matrix of T-degree results can be constructed by analyzing canonical mutual information NMI values of each clustering result. Taking two clustering results G _a and G_b Is of the joint distribution of (a) _p (a, b) (wherein a, b are positive integers, take the values 1,2, …, t.), the edge distribution is _p(a) and _p (b) The mutual information MI between the two can be defined as:

correspondingly, normalized mutual information NMI is introduced:

wherein ,H(G_a) and H(G_b ) Respectively represent G _a and G_b Entropy values of two random variables, wherein the entropy values are defined as follows:

NMI _ab the larger the value of (c) is, the smaller the difference between the clustering results is, and when the two clusters are completely equal, the value is 1. Based on NMI measure, NMI interaction matrix between different clustering results can be established.

Next, summing the NMI matrix to obtain sum vector V of NMI values of different clustering results and other clustering results _sum ：

wherein ,is the sum of NMI values of i clustering results. In order to select a result with reliable clustering result precision and a certain divergence between results for final integration, the following selection scheme may be used to select a clustering result Ids for final integration analysis:

wherein ,V_low and V_high Respectively represent the order V _sum In (a)Divide into two parts and->A value of the split bit.

Fifth embodiment:

further defining the automatic layering method of logging curves based on constraint integrated clustering according to the first embodiment,

the method for combining the candidate layering points by using the region growth based on energy constraint in the step six comprises the following steps:

under the constraint of the prior information of lithology unit junction points given by the invariant features and the integrated clusters, the reliability of automatic layering of logging data is further improved by adopting a region growing mode. Generally, region growing is mainly divided into two steps, 1) manually selecting one or more seed points in a target region; 2) And comparing all the growing points with the neighborhood points according to the growing criteria, merging the neighborhood points meeting the growing criteria with the seed points, and attributing the merged neighborhood points to the same area until the number of the new growing points is zero.

For a certain logging dataset, the correlation difference feature dCorr and the reservoir structure tensor feature Ten are first obtained using equations (8) and (10). Then, the following region growing seed point P can be obtained by using the correlation difference characteristics _dCorr ：

Wherein ZCrosss (dCorr) represents the zero crossing of the dCorr feature; n (p) _i ) Representing the current point p _i Is a local neighborhood of (b); t (T) _dCorr Is a threshold constant. Similarly, the following region growing seed point P can be obtained using the structure tensor feature _Ten ：

P _Ten ＝{p _i |(p _i ∈Peak(Ten))} (12)

Wherein Peak (Ten) represents the Peak point of the Ten feature. Thus, the total region growing seed point P can be obtained _ec The method comprises the following steps:

P _ec ＝U(P _dcorr ,P _Ten ) (13)

where U (X, Y) represents the union of sets X and Y.

The invention adopts energy minimization as a growth criterion, and for the proposed regional growth seed points, the energy model of the local region is utilized, and the energy is the optimal state when the energy reaches the minimum after iterating for a certain number of times, and the edge structure can be effectively represented.

And the regional growth is realized by combining energy minimization constraint, so that the accuracy of a pseudo layer of a geological reservoir is effectively improved. And (3) taking the reservoir candidate layering points integrated with the clustering classification result in the step (V) as input data, comprehensively considering two principles of constraint of layering energy minimization and layering point depth space connectivity, effectively avoiding the problem of dislocation of layering points, and enhancing the stability of algorithms in geological reservoirs with different complexity.

In order to verify the effectiveness of the method provided by the invention, an automatic layering comparison experiment is carried out on a plurality of wells in a region of a well in a Daqing oilfield land depression basin with more thin layers and thin interbeds. Specifically, 7 conventional curves ubiquitous in the work area are selected to form a logging data set: natural Gamma (GR), natural potential (SP), sonic jet lag (AC), shallow lateral resistivity (LLS), deep lateral resistivity (LLD), density (DEN), borehole diameter (CAL). The work area mainly comprises five lithologies of mudstone, siltstone, argillite siltstone, siltstone and oil shale, so that the clustering number of single cluster members is set to be 5. In addition to the invention, two kinds of comparison of K-means and integrated K-means are adopted to conduct automatic layering comparison test.

In the experiment, a large number of comparison tests are carried out on target reservoirs with different depths and different geological complexity degrees, and the clustering frequency T of the single-view clustering method in the experiment is set to be 100. In order to quantitatively evaluate the performance of the method, lithology boundary points in lithology section comprehensive interpretation results of geological interpretation personnel are used as references, each path of logging data is used as a reference, and the purity of the method is used as a quantitative evaluation index to automatically and hierarchically detect the well in the work area:

for a certain sample size of N log data sets, c= { C ₁ ,C ₂ ,…,C _j The automatic hierarchical division result is the division result after the automatic hierarchical division, and the real manual division result is W= { W ₁ ,W ₂ ,…,W _k }. Table 2 shows the quantitative evaluation results of the detection results of the methods exemplified by the A1 well, wherein the purity has a value of [0,1]The better the clustering result, the closer the purity value is to 1. FIG. 4 shows the A1 well asThe quantitative evaluation of the examples is visual.

TABLE 2 purity of layered results and artificial layered results for different clustering methods

Logging curve	k-means	Integration of k-means	Text clustering method
				GR	0.78	0.82	0.85
CAL	0.61	0.66	0.70
				SP	0.72	0.75	0.75
AC	0.69	0.74	0.77
				LLS	0.73	0.76	0.83
LLD	0.73	0.76	0.83
				DEN	0.70	0.77	0.79

As can be seen from the quantitative evaluation results given in table 2, the purity of the automatic layering result and the artificial interpretation result of the present invention is significantly better than that of the comparison method, and the automatic layering performance meeting the application requirements can be achieved. Fig. 5-6 show visual comparisons of automated layering results for two wells. The matching degree of the automatic layering result and the manual interpretation result can be more clearly seen from the visual comparison chart. It should be noted that, the invention not only can realize the automatic fine layering of the single sand level reservoir, but also can realize the optimization of the artificial interpretation result to a certain extent, such as the horizontal dotted line position penetrating through each data in fig. 5-6 is the layering point given by the invention, and has a certain difference from the artificial interpretation result, but the layering point can be better positioned to the half-amplitude point position of the typical lithology sensitive curve through the comparison with the corresponding logging curve, and can be considered as the optimization of the artificial interpretation result. In summary, the algorithm provided by the invention essentially requires logging vector information of three consecutive samples to represent the reservoir structure change, and therefore, for conventional logging data with longitudinal resolution of 0.125 m, the invention can realize automatic layering of reservoirs with more than 3 sample levels, namely, a thin layer with a minimum layering unit of 0.375 m.

While the invention has been described with respect to particular embodiments thereof, it will be appreciated that the invention is not limited thereto but may be practiced with modification and alteration within the spirit and scope of the appended claims.

Aiming at the problem of realizing fine layering of complex unconventional geological reservoirs by using logging data, the invention researches the prior clustering-based geological layering model, however, the prior method only aims at a specific rock stratum, cannot adapt to rock stratum environments with different complexity degrees, and particularly does not go deep in the problem of layering of a thin sand body rock stratum. Therefore, the method of the invention increases different attributes of the original data by generating the association relation between the logging data by using the association rule, and based on the lithology jump point information provided by the correlation difference feature and the tensor feature, the pseudo jump point influence caused by the interference information in the logging data is effectively reduced by adopting the region growth algorithm of the layering energy function through the supplement of the integrated clustering information, the description capability of the logging data to the fine reservoir is improved, and the accurate and reliable automatic layering is finally realized.

The invention provides an automatic clustering layering method for logging curve integration based on layering energy constraint. The invention is suitable for the actual well area with different data distribution characteristics. Compared with classical geostatistical methods and general machine learning methods, the method provided by the invention can improve layering precision, can effectively realize automatic layering of reservoir fine description industrial application requirements, and has remarkable popularization potential. Fig. 1 shows a block diagram of the structure of the present invention. The key technical content of the invention comprises two parts, namely one part, a reservoir layer segment segmentation model based on the fusion of local energy minimization and region growth is provided; and secondly, dividing different lithologic intervals of the reservoir by combining an integrated clustering result and a region growing algorithm.

Claims (3)

1. The invention provides an automatic layering method of a logging curve based on constraint integrated clustering, which comprises the following steps:

step S1: the method comprises the steps of adopting a correlation rule analysis mode to realize effective mining of the correlation rule among the input multi-attribute logging data;

step S2: screening logging data by using the association rule established in the first step, and extracting relevant difference features and tensor features of the screened data;

step S3: acquiring a zero crossing point of the correlation difference characteristic and a tensor characteristic peak point in the second step, and taking the zero crossing point and the tensor characteristic peak point as seed points for region growth;

step S4: generating a plurality of clustering results by adopting a k-means clustering and spectral clustering method, constructing NMI matrixes based on all the clustering results, and selecting cluster members meeting a certain degree of difference;

step S5: performing integrated clustering on the cluster members obtained by screening in the step four, arranging clustering classification results according to the depths of the wells, and taking the classification mutation points as reservoir candidate layering points;

step S6: combining the 'pseudo layers' by using the region growing process based on energy constraint for the candidate layered points in the step five;

step S7: and arranging the results after the region growth according to the depths of the wells, and taking the result mutation points as final reservoir layering points to realize fine layering of the reservoir.

2. The automatic layering method of the logging curves based on the constraint integrated clustering according to claim 1 is characterized in that in the fifth step, an NMI matrix based on clustering results is constructed to better screen out suitable candidate reservoir layering points with the effect of improving the integrated clustering.

For one logging data set, clustering the logging data set by adopting a K-Means and a spectral clustering method respectively, wherein the clustering times are T times (T is a positive integer, and the values are 1,2, … and N); two kinds of clustering result sets omega= { G can be obtained respectively _ij ,i＝1,2,3…,T；j＝1,2,3…,T},G _ij Representing the j-th class result of the i-th clustering algorithm, wherein i and j are positive integers, and the values are 1,2, … and T; for each algorithm, an NMI matrix of T times of results can be constructed by analyzing the canonical mutual information NMI value of each clustering result; taking two clustering results G _a and G_b Is p (a)B) (where a, b are positive integers, take on values 1,2, …, T) and the edge distribution is p (a) and p (b), then the mutual information MI between the two can be defined as:

correspondingly, normalized mutual information NMI is introduced:

wherein ,H(G_a) and H(G_b ) Respectively represent G _a and G_b Entropy values of two random variables, wherein the entropy values are defined as follows:

NMI _ab the larger the value of (2) is, the smaller the difference between the clustering results is, and when the two clusters are completely equal, the value is 1; based on NMI measure, NMI interaction matrix between different clustering results can be established.

Next, summing the NMI matrix to obtain sum vector V of NMI values of different clustering results and other clustering results _sum ：

wherein ,NMI values for i clustering results;

in order to select a result with reliable clustering result precision and a certain divergence between results for final integration, the following selection scheme may be used to select a clustering result Ids for final integration analysis:

wherein ,V_low and V_high Respectively represent the order V _sum In (a)Divide into two parts and->A value of the split bit;

after the clustering result for integration is selected, counting the times that each sample point is at the clustering edge, obtaining lithology interface candidate boundary points given by the integration clustering method through threshold processing, and completing assignment of clustering labels among the boundary points through alignment of the clustering result.

3. The automatic layering method of logging curves based on constraint integrated clustering according to claim 1, wherein in the sixth step, a model of region growth based on layering energy constraint is constructed:

the growth criterion is the core of region growth, and the invention converts the problem of region growth into a data segmentation problem of minimum layering energy iteration by introducing layering energy constraint so as to realize the purpose of automatic layering; to log data sets(K is a positive integer, the maximum value is the number of sample points of the logging curve set, n represents the number of the whole logging data set), the corresponding integrated clustering result and the region growth seed point are region growth basic data, and a layered energy function E (L) in each iteration process is defined as follows:

wherein P is a certain sample point in the integrated clustering result, P is the integrated clustering result,representing data cost components, D _p (L _p ) Representing point p to label L _p The Euclidean distance of the corresponding cluster center; />Representing a smoothed cost component for penalizing the tag discontinuity; w (w) _pq ＝αe ^-d(p,q) Is the weight assigned to the point pair (p, q) in the same neighborhood with different labels, d (p, q) is the Euclidean distance of the point pair (p, q), δ (L) _p ≠L _q ) Indicating the function discontinuously for labels, i.e. at L only _p ≠L _q When the value is 1, α=max (max (d (p, L _p )),max(d(q,L _q )))，max(d(p,L _p ) P to the intra-neighborhood label L) _p The maximum euclidean distance of the points, max (d (p, L _p ) For p to the label L in the neighborhood _p The maximum distance of the points, max (d (q, L _q ) Q to the label L in the neighborhood _q The maximum distance of the points, max (d (P, L _p ) For P to L in the neighborhood label _p The maximum distance of the points.

After the construction of the energy function is completed, the cluster labels of the integrated clusters are used as initial segmentation, region growth analysis is carried out from the region growth seed points, region growth is carried out iteratively under the constraint of the minimization of the energy function, and the automatic layering task of the logging data can be completed.