See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/344313893

Adaptive multi-resolution graph-based clustering algorithm for electrofacies

analysis

Article in Applied Geophysics · March 2020

DOI: 10.1007/s11770-020-0806-x

CITATIONS READS

36 149

6 authors, including:

Binsen Xu
China University of Petroleum, Beijing
9 PUBLICATIONS 147 CITATIONS

SEE PROFILE

All content following this page was uploaded by Binsen Xu on 29 September 2024.

The user has requested enhancement of the downloaded file.

APPLIED GEOPHYSICS, Vol. 17, No. 1 (March 2020), P. 13―25, 13 Figures.
DOI:10.1007/s11770-020-0806-x

Adaptive multi-resolution graph-based clustering

algorithm for electrofacies analysis*
Wu Hongliang1, Wang Chen3, Feng Zhou1, Yuan Ye1, Wang Hua-Feng♦2,3, and Xu Bin-Sen1

Abstract: Logging facies analysis is a significant aspect of reservoir description. In particular,

as a commonly used method for logging facies identification, Multi-Resolution Graph-based
Clustering (MRGC) can perform depth analysis on multidimensional logging curves to
predict logging facies. However, this method is very time-consuming and highly dependent
on the initial parameters in the propagation process, which limits the practical application
effect of the method. In this paper, an Adaptive Multi-Resolution Graph-based Clustering
(AMRGC) is proposed, which is capable of both improving the efficiency of calculation
process and achieving a stable propagation result. More specifically, the proposed method,
1) presents a light kernel representative index (LKRI) algorithm which is proved to need
less calculation resource than those kernel selection methods in the literature by exclusively
considering those “free attractor” points; 2) builds a Multi-Layer Perceptron (MLP) network
with back propagation algorithm (BP) so as to avoid the uncertain results brought by uncertain
parameter initializations which often happened by only using the K nearest neighbors (KNN)
method. Compared with those clustering methods often used in image-based sedimentary
phase analysis, such as Self Organizing Map (SOM), Dynamic Clustering (DYN) and
Ascendant Hierarchical Clustering (AHC), etc., the AMRGC performs much better without
the prior knowledge of data structure. Eventually, the experimental results illustrate that the
proposed method also outperformed the original MRGC method on the task of clustering and
propagation prediction, with a higher efficiency and stability.
Keywords: MRGC, AMRGC, MLP, logging facies analysis

Introduction algorithms are not well suited to practical situations,

(i.e. log space is not equivalent to geological space, and
two points that are close to each other in log space may
Logging facies analysis is an important prerequisite not be similar geologically). At an earlier stage, Rogers
for identifying reservoir sedimentary characteristics, and proposed a BP neural network aiming at solving this
the automatic clustering method is usually required in problem point wisely, which requires few advanced
the analysis process. However, due to the high feature statistical knowledge or strong log interpretation skills
dimensions extracted on the dataset, many clustering (Rogers, 1992). However, the BP neural network has

Manuscript received by the Editor September 28, 2019; revised manuscript received February 07, 2020.
*This work was sponsored by the Science and Technology Project of CNPC (No. 2018D-5010-16 and 2019D- 3808)
1. Research Institute of Petroleum Exploration & Development, PetroChina, Beijing 100083, China.
2. School of Computer Science and Technology, North China University of Technology, Beijing 100144, China.
3. College of Software, Beihang University, Beijing 100191, China.
♦Corresponding Author: Wang Hua-Feng (Email: [email protected])
©2020 The Editorial Department of APPLIED GEOPHYSICS. All rights reserved.
13
 Adaptive multi-resolution graph-based clustering algorithm

not achieved good results in the practical applications networks, we proposes a graph-based adaptive multi-
because of slow learning speed and low generalization resolution clustering analysis method (AMRGC).
capacity. Since then, Ye (Ye, 2000) proposed a multi- Compared with the MRGC, we mainly make two
resolution graph-based clustering method (MRGC), aspects of improvement in new proposed method: (1) a
and this algorithm has been developed based on some light kernel representative index (LKRI) is used instead
lithofacies analyzing software (Sutadiwirya, 2008; of the KRI in MRGC algorithm; (2) Use the BP(MLP)
Pabakhsh, 2012; Khoshbakht and Mohammadnia, 2000; instead of KNN in the prediction of the distribution
Pabakhsh, 2012; Nouri‐Taleghani, 2015; Tian, 2016). of the new data stage (also named as the propagation
Compared with the traditional clustering algorithms stage). Consequently, the unstable results of the
(TCAs) such as the Self Organizing Map (SOM)( traditional K-nearest neighbor algorithm due to random
Kohonen, 1990; Tian, 2016), Dynamic Clustering (DYN) initialization parameters are effectively avoided.
(Diday, 1971; Mourot, 1993), Ascendant Hierachical
Clustering (AHC) (Lance and Williams, 1967; Lukasová,
1979) and Artificial Neutral Network (ANN) (Tang, Graph-based adaptive multi-resolution
2011), the MRGC (Ye, 2000) is proved to be a better
option, because those TCAs have been thought to have cluster analysis method
following limitations: (1) need to know the number of
clusters beforehand; (2) be sensitive to initial conditions
and variations of parameter values; (3) not robust to In general, the MRGC at least consists of five steps
the data variation in practical application (Mourot and (Ye, 2000): (1) Calculating the Neighbor Index(NI)
Bousghiri, 1993). which will help estimate the probability density function
For all clustering methods, determining the optimal (PDF); (2) Calculating the KNN attractions based on the
number of clusters is one of the most important tasks calculated NI values in the first step; (3) Calculating the
of a clustering algorithm. For instance, MRGC is Kernel Representative Index (KRI) which is necessary
characterized by the nonparametric K-nearest neighbor, for generating a proper range in determining the number
approach and a graph data representation. In practice, of clusters; (4) Merging those initial small groups to
the MRGC is a good tool which analyzes the structure of form final clusters by the Nearest Neighbors’ Attraction
the complex data and partition natural data groups into Power; (5) Predicting the distribution of the new data. As
different shapes, sizes, and densities. However, due to shown in Figure 1, the difference between the traditional
the large amount of calculation of the MRGC algorithm MRGC and AMRGC mainly lies in (i) calculating
itself, and the use of the KNN method as an important the Light-Kernel Representative Index (LKRI); and
step in the MRGC propagation prediction stage, the (ii) using LPA based on BP network to predict the
overall process has seriously reduced the calculation distribution of new data.
efficiency of the MRGC algorithm. As mentioned earlier,
MRGC needs to rely on initialized parameter values, Data normalization
which ultimately leads to unstable analysis results. In order to eliminate the inconsistency of the actual
In summary, the BP neural network method and the data dimension of the original log and the fluctuation
MRGC method each have advantages and disadvantages, of the value range, the acquired raw log data is usually
but there is a certain complementarity between the cleaned and being normalized prior to further processing
two. In this paper, inspired by MRGC and BP neural (Dodge, 2003). Practically, the raw log data is

MRGC MRGC
Calculating the Kernel Using KNN to predict the
Representative Index (KRI) distribution of new data
Data
Normalization AMRGC
Calculating the Light-Kernel Merging groups AMRGC
Representative Index (L-KRI) to form final Using MLP to predict the
Calculating the
Neighbor Index(NI) clusters distribution of new data
Form KNN
Attraction Sets

Experts divide lithofacies

Log Data AMRGC Algorithm Prediction

Fig.1 The pipeline of the AMRGC algorithm which also indicates the difference between the MRGC and AMRGC.
14
 Wu et al.

normalized to a range and converted into dimensionless vijth nearest neighbor of xi . According to the relationship
numbers. As for the AMRGC algorithm, we also adopted described above, finally, W[i][j]=V[U[i][j]][i].
a strategy of normalization on the original log data
obtained. Suppose dataset D has n objects, that is, D={x1,
x2, x3, ..., xn}, every object has the same dimension. every X4
 4NN
object xi can be seen as a vector, where xi ={xi1, xi2, xi3,
 W[1][1]=5 3NN
..., xim}, 1≤i≤n , and is the dimension for xi , xij is the j 2NN X2

th attribute value of xi . Prior to the normalization, the 2NN
 
mean and variance (  and  2 ) of given sample over all X1
5NN
1NN
X5
dimensions are calculated respectively. The mean  j 4NN
3NN 1NN
and variance  of the attribute, as well as the mean 
2
j U[1][1]=5 X6
and variance σ2 of all the samples, are calculated as U[1][5]=1
equations (1) to (7), which is the data normalization 5NN X3

process in Figure 1.
Fig.2 Illustration of the relationship among U, V, and W.

1
 j   i n 1 xij ( j  1,2,, m ), (1) As shown in Figure 2, 1NN of x1 is x5, so we can get
n
U[1][1]=5, which means 1NN of x1 is x5, V[1][5]=1,
1 n which means x5 is 1NN of x1, and x1 is x5’s 5NN, so
  2j 
n 1 i  1
( xij  j )2 ( j  1,2,, m ), (2) there is W[1][1]=5, which means that x1 is 5NN of its
1NN. Among them, W[i][j] can be calculated from the
 matrix U, V, and further we have W[i][j] according to
  (1 , 2 , , m ) ( j  1,2,, m ), (3) V[U[i][j]][i].


 2  ( 12 ,  22 , ,  m2 ) ( j  1,2,, m ), (4)
Calculation of the neighbor index(NI)
Let the measurement of a given point x be an element
 

where xi  represents the vector of xi after the norma- of a set of measurement points D = {x1, x2, x3, ..., xn}, and
lization, such that, let point y be measurement point x’s th Nearest Neighbor
(NN) in the set of measurement points S, n≤K. The
xij   j “limited rank” τn of measurement point x with respect to
xij  ( j  1,2,, m ), (5) its nth NN(y) is defined to be:
 2j

  c (
xi ) exp( m /  ), (8)
xi  ( xi1 , xi 2 , , xij ,, xim ) ( j  1,2,, m ), (6)
where α is greater than zero and c≤n-1. It is noted that
Thus, we have the Euclidean distance between any
α is insensitive to the size of the data set and may be
given two data points (xi, xj) as:
initialized once for all data set. In practice, α will be set
  to 10.
D ( xi , x j )  mj 1 ( xij  x jj ) 2 , (7) In fact, in this paper, the limited rank τi is defined for
calculating xi’s K nearest neighbors. The sum of the
Next, we define a matrix U=(uij)n×n, and U[i][j]=uij
limited ranks for each point x is expressed as,
which is the index value of the jth nearest neighbor(NN)
  
of measurement of vector xi based on the calculation s( xi )  in11  c ( xi ),
    (9)
of D( xi , x j ) , i.e. x j ’s j th nearest neighbor is xuij . For

convenience, xi is considered to be the n th nearest Secondly, the smallest value Smin and largest value
Smax of rank sums are calculated by
neighbor of itself. Once U is determined, a sorting
algorithm can be applied to identify the rank matrix  
Smin ( x ) Min{s(
 xi )},(i 1, 2,, n  1), (10)
W=(w ij) n×n in which w ij means W [i][j] and W indicates
  
that how much xi is relevant to its’ jth nearest neighbor. Smax ( x ) Max{s (
 xi )},(i 1, 2,, n  1), (11)
Meantime, a companion matrix V=(vij)n×n is initialized

with vj=j (j=1, 2, ..., n), which indicates that x j is the Thus, the NI calculation formula can be expressed as,
15
 Adaptive multi-resolution graph-based clustering algorithm
  
 s ( xi )  S min ( x ) means xi is most likely in the valley zone.
NI ( xi )    ,(i 1, 2, , n  1). (12)
S max ( x )  S min ( x ) Also, at very start, we need to determine all objects’
situations, then the attractions sets are formed. We have
the attraction sets S={Si}(i=1, 2, ..., N) (N is the number
The K nearest neighbors’ attraction calculation of kernel points) which is the initial group which can
method help cluster the delta.
After calculating the NI of all objects, we can use Calculating the light kernel representative
a multidimensional KNN point-to-point attraction
algorithm. And it will help us calculate adherence points’ index (LKRI)
attraction for every measurement ofpoint and determine In order to further determine the cluster core, we need
 
the small data group center. Attrxi ( x j ) is xi ’s Kth nearest to calculate the KRI of each kernel point xi (i=1, 2, ...,
   
neighbor x j ’s attraction to xi , which can be calculated n), we set 1-st nearest neighbor with x j , whose NI( x j )
   
by, > NI(xi ), and H(xi , x j ) = h , which means that x j is the
 
    hthNN of xi . Thus, the KRI of xi can be calculated as
Attrx ( x j ) NI ( x j )Vx ( x j )  NI ( xi ).
 (13)
i i follows,
It is acknowledged that a better data clustering results      
could be obtained even with a larger amount of data and KRI ( xi )  NI ( xi )a H ( xi , x j )b D( xi , x j )c , (16)
a higher data dimension when the K value is ranged from
4 to 12 (Hastie, 2009). Meantime, it is also realized that where a, b, c are used to weight each corresponding
      
the probability of misclassification with K value in this function NI( x j ) , H(xi , x j ), and H(xi , x j ), D(xi , x j ) under
interval could be minimized. In view of this reason, the condition that a = b = c = 1.
K value is configured by 4 for the following experiments. Compared to MRGC, AMRGC only needs to calculate
 the KRI value according to the equation (16) based on
The adherence function Vx ( x j ) is defined as ,
i
the previously obtained free attractor nodes. In contrast,
  the MRGC will have to do the calculation on all the
 1 (if x is one of K nearest neighbor points of x ),
Vx ( x j )   j i
data points. Therefore, we call the new algorithm LKRI,
i
0 otherwise.
(14) which means that AMRGC is with a less computational
load. In the experimental section, the benefits of
   replacing KRI with LKRI will be discussed in detail.
When xi belongs to the maximum of Attrxi ( x j ) , xi
  On the other hand, the reason why this algorithm
is attracted to the nearest neighbor x j . If all x j in the K can obtain suitable kernel point candidates is that by
 
nearest neighbor of xi do not attract xi , and this implies analyzing NI, as we know, the points with higher NI
 values are more likely to become the center of clustering.
that xi is not adhered to any other points. We calculate
 Actually, an NI value can be analogously attractive, and
the maximum of Attrxi ( x j ) as the Attrx , that is,
i
a higher NI value means that the selected point is much
 more attractive to other data points, the more likely it
Attrx  Max{ Attrx ( x j )}. (15) will be near the center of the cluster.
i i

For example, as shown in Figure 3a, the data point

It should be noted that all objects are classified based with an NI value 0.9, i.e. the center point of figure 3a,
on the three basic situations as follows, and the rest of the points nearby whose NI values are

(1) For j=1, 2,...,K(j≠i), the Attrxi ( x j ) always true, mostly less than 0.9 formed an attraction set. Since x1

which means xi is not attracted by any points, even is not attached to other data points, it is a free attractor.
 More likely, the same thing happened with the data point
the K nearest neighbor, xi is a “Free Attractor”, which
is similar to a kernel point of a local maximum of the whose NI value is 0.95, i.e. the center point of figure 3b.
probability distribution function (PDF). As a result, the distance of x2 from x1 is larger than that
 of most points in x1’s attraction dataset, that is, only x2
(2) If xi is directed to another point and at the same
 with an NI value of 0.95 satisfies the KRI calculation
time attracts
 other points, xi is a “Related Attractor”. In condition. Note, the method for calculating the KRI of x1
this case,xi is on the slope. point can refer to equation (16).
(3) If xi is directed
 to another point but does not attract For a given point with a high NI value, it will be a
any other point. xi is the “Pending Related” , which difficult task to find another data point with an even
16
 Wu et al.

higher NI in the vicinity. This also means a point whose cluster center points. Not only is its’ clustering result
NI value is larger than that of the known point tends to quite consistent with that of the traditional MRGC when
be a long way from the known point. That is, the free there are fewer center points needed to be determined,
attractor nodes chosen by thresholding NI values of but also it outperformed the traditional MRGC when
given points tend to have a high probability of being many center points occurred. In short, the new algorithm
kernel point candidates. Later experiments also show can greatly reduce the amount of process computation
that, compared with the results of previous kernel under the premise of ensuring that the correct cluster
selection algorithms, AMRGC is good at selecting the center is chosen.
(a) (b) 0.2
0.1
0.03
0.1
0.2
0.6
0.8
0.95
0.05 0.08
0.9
0.7 0.7
0.5 0.55
0.15 0.2

0.3 0.1

Fig.3 Attraction schematic diagram.

Finally, we arrange the LKRI in descending order, a relational complete graph model based on given
which will help us pick the final kernel points. And the samples, we need to construct a graph structure first.
merging stage will merge the obtained attraction sets and The nodes of each graph are those known data points
finally output the clustering result (Sutadiwirya, 2008). that contain the labeled data as well as the data to be
The details of the algorithm are described in Appendix. labeled, and the edges between the nodes represent the
similarities between them. In each graph, the labels
of nodes are transmitted to other nodes according to
LPA and MLP in AMRGC propagation stage their similarities. The greater the similarity of nodes,
(LPA+MLP) the easier the labels isare to spread. There are many
ways to build this map such as RBF and KNN and so
In order to take full advantage of the data labels
on. KNN graph was exploited in this study. In view of
obtained by the AMRGC algorithm, we no longer use
that, only will the K-neighbor weight of each node for
a simple KNN for propagation, but a KNN-based label
this KNN graph be kept, while the rest will be set to 0.
propagation algorithm (LPA).
Hence, such a KNN graph can be regarded as a sparse
Using LPA based KNN to prediction similarity matrix. Suppose that all these graphs were
The KNN is a classical supervised learning method, fully connected, then the weight between any given node
 
for a new object x p , we use k trainingobjects xi (i=1, 
2, xi and node xj could be calculated by equation (18), and α ,
..., k) closest in distance d (d ≥ 0) to x p , then classify x p an N×N probability transfer matrix P is defined, where
by using majority vote among the k neighbors. Usually, Pij denotes the probability of node xi transferring to node
after data normalization, the distance d is calculated by xj. According to Zhu’s definition (Zhu, 2002), we have,
two-norm based on different dimensions attributes, it || xi  x j ||2
can also be called “Euclidean distance”. As previously ij exp( 
 ), (18)
2
claimed, the distance d is defined as,
    2  ij
d D ( x p , x
 i ) || x i  x p ||
Pij  P(i  j )  . (19)
 K

k 1 ik
(17)
=  mj1 ( xij  x pj  ) 2 , 1, 2, , k .
i MLP to make new data prediction 
As shown in Figure 4a, for inputted object x = (x1,
Inspired by Zhu’s method (Zhu, 2002), the main x2, ..., xi), its each dimension attribute
 is an input unit in
idea of our proposed LPA algorithm is to utilize the input layer. The output vector is yL  ( y1 , y2 ,, yn ) (l is
tag information of tagged nodes to predict the tag the number of classes label which is predestinated for

information of unlabeled nodes. In order to build AMRGC). And t  (t1 , t2 ,, tl ) is the target vector value.
17
 Adaptive multi-resolution graph-based clustering algorithm

At each layer, we first compute the total input z for each equation (20), the backward propagation algorithm is
unit, which is a weighted sum of the outputs of the units used. In Figure 4b, the equations are used for computing
in the layer below. Then a non-linear function f(∙) is the backward propagation. The output error between
applied to z to get the output of the unit. For simplicity, output and target value can be describe as equation (21),
the bias terms isare omitted. The non-linear function which we also call it as “Loss function” and is shown as
used in neural networks is a rectified linear unit (ReLU) following:
f(z) = Max(0, z). The output yl is classes’ label vector 1   1 l
which is predestinated for AMRGC. Such that yl will be E  || t  yL ||22   (tl  f ( ml f (  jm f ( ij xi )))) 2 ,
2 2 l 1
expressed as
(m  H 2, j  H1 , i  input, l  output ). (21)
 ( zl ) f (  ml ym )
yl f
where the ωml, ωjm and ωij can be referred to Figure 4.
= f (  ml f ( zm ))  f (  ml f (   jm y j ))
Once the ∂E/∂zm is known, the error-derivative for the
weight ωjm on the connection from unit j in the layer
below is perceived as yj ∙ ∂E/∂zm. Then we can determine
(  ml f (   jm f ( z j ))) f (  ml f (   jm f (  ij xi ))),
f (20)
the relationship between E and input vectors by using
In order to confirm the values of ω ml, ω jm, ω ij in the chain rule of derivatives.

(a) (b)

Output
units

Hidden
units H2
Hidden
units H1
Input
units

Fig. 4 Propagation lgorithm workflow. (a) The four-layer forward propagation perceptron.
(b) The four-layer backward propagation perceptron.

When we have new objects to train, we can use parameter K in this experiment was set to 4, and the
gradient descent algorithm to determine the unknown merging parameter K’ was set 8. In the experiment, we
weight parameters and predict the test objects by randomly selected a group of acquired logging data,
corresponding weight parameters. In recent years, the which consists of 840 data points. After descending
neural network technology has made a breakthrough the LKRI, the result is shown in Figure 5. These points
progress (Hinton, 2006) and received great attention. are arranged in descending order according to the
As aspect of our contribution, the new proposed method LKRI value, and the free attraction points in the front
combines the neural network with the clustering method. more likely become kernel points, which are retained
Specifically, this method extracts features from the
training data by neural network, and then uses the trained 1.0

model to predict the propagation of the data without tags. 0.8

The specific network parameters selected in the process
0.6
will be described in detail in the experimental section in
LKRI

Section 3. 0.4

0.2

Experimental design and result 0

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Cluster number

Clustering results of AMRGC and MRGC in

comparison
As for the calculation process of attraction sets, the Fig. 5 Kernel points LKRI and cluster number choosing.

18
 Wu et al.

in the subsequent fusion process. The dark blue circle little effect on the clustering results and the clustering
area indicates the number of corresponding point sets results are relatively stable, which also indicates the
obtained by selecting different numbers of free attraction robustness of AMGRC. The effect of K value on
points. Wherein, each ellipse or circle represents a set of clustering results is shown in Figure 6. According to the
points formed by clustering. clustering results shown in Figure 6, it can be concluded
In this experiment, five-dimensional data such as that the initial value of the K value has little effect on
natural gamma-ray(GR), lithology density(DEN), the final clustering results. The colored balls in Figure
compensated neutron (CNL), sonic interval transit time 6 represent different clusters acquired after data points
(AC), and deep resistivity (RD) were used for clustering are processed by the clustering algorithm, and the colors
experiments. Finally, the two-dimensional data of GR represent different categories. The lithofacies label is
and CNL are drawn and displayed. In the figure, the represented by each color needs to be obtained after
abscissa axis is GR and the ordinate axis is CNL. It can further processing.
be seen from these experiments that the K value has
(a) K=5 (b) K=6 (c) K=7 (d) K=8
25 25 25 25

20 20 20 20

15 15 15 15
CNL (%)

CNL (%)

CNL (%)

CNL (%)
10 10 10 10

5 5 5 5

0 0 0 0
0 50 100 150 200 0 50 100 150 200 0 50 100 150 200 0 50 100 150 200
GR (API) GR (API) GR (API) GR (API)

Fig. 6 Effects of different K values on clustering results.

Through the LKRI value, we can obtain multiple Therefore, the AMRGC algorithm can greatly reduce
important indicators from those data points, which the amount of calculation. The experiment results depict
corresponds to the best cluster number at different that the AMRGC algorithm can still produce valid
resolutions. After obtaining the optimal number of results although the amount of calculations decreases
clustering centers, we can still manually adjust the dramatically . In other words, the AMRGC algorithm can
clusters. As for computing resource consumption, the obtain the same kernel points as the MRGC algorithm,
AMRGC algorithm only needs to calculate the KRI and can even obtain better clustering results than the
value of a free attractor node (in this case, about 40 MRGC algorithm. As shown in Figure 8, it illustrates
data points), while the traditional MRGC algorithm different clustering results with different numbers of
must calculate the KRI value of all data points (in this clusters: 3, 5, 7, and 10 (both the result of traditional
case Case, 840 data points). This indicates that the MRGC and our proposed AMRGC algorithm are listed).
computational cost of the new KRI calculation method
According to the results, we can observe that when the
(LKRI) is reduced by 1/16 compared to what MRGC
number of clusters is small, LKRI and KRI will choose
usually requires. In fact, as the amount of data increases,
the same kernel and produce the same clustering results.
the proportion of all data points and free attractor nodes
Only when the number of clusters is large (for example,
increases rapidly (see Figure 7).
we found that it is 10 in this experiment), LKRI will
4000
select different kernel points from that the KRI select,
3500 KRI
LKRI and certainly the clustering results will be different.
3000
Also, the experimental results show that the kernel
Operations

2500
points selected by the LKRI method can obtain the same
2000
1500
or even better clustering results.
1000
In general, log interpretation experts understand
500 the surface truth of lithofacies by referring to the
0 information provided by coring wells. Based on coring
0 500 1000 1500 2000 2500 3000 3500 4000
well data, the number of clusters was determined.
Data size (Sample number)
According to the recommendations of experts, in the
Fig. 7 The comparison of the calculation consumption following experiments, we used the cluster clustering
between AMRGC and MRGC.
19
 Adaptive multi-resolution graph-based clustering algorithm

model to classify the lithofacies. In addition, in order to DYN algorithms to 5 (all of these methods need to know
compare the clustering effect with other algorithms, we the number of clusters in advance). Figures 9 and 10
also set the number of clusters of the SOM, AHC, and show the final clustering performance of each method.

Clustering
AMRGC MRGC
number
25 25
20 20

15 15
CNL (%)

CNL (%)
3 3
10
N=3 10 2 2
5 1 5 1

0 0
0 50 100 150 200 0 100 200
GR (API) GR (API)

25 25
20 20
5 5
15 15
CNL (%)

CNL (%)
4 4
N=5 10 3 10 3

5 2 5 2
1 0 1
0
0 50 100 150 200 0 50 100 150 200
GR (API) GR (API)

25 25
7 7
20 20
6 6
15 15
CNL (%)

CNL (%)

5 5
N=7 10 4 10 4
5 3 5 3
2 2
0 0
0 50 100 150 200 1 0 50 100 150 200 1
GR (API) GR (API)

25 10 25
10
9
20 20 9
8
8
15 7 15 7
CNL (%)

CNL (%)

6 6
N=10 10 5
10
5
4 5 4
5
3 3
0
0 2 2
0 50 100 150 200
0 50 100 150 200 1 1
GR (API) GR (API)

Fig. 8 The clustering results by using AMRGC and MRGC.

(a) (b) (c) (d)

25 25 25 25
CNL (%)

CNL (%)

CNL (%)

CNL (%)

0 GR (API) 200 0 GR (API) 200 0 GR (API) 200 0 GR (API) 200

SMO AHC DYN AMRGC
Fig. 9 the illustrations of clustering results produced by different methods.

20
 Wu et al.

As shown in Figure 10, it intuitively proves that the each class in clustering result, one point is one target,
AMRGC algorithm performs better than other methods and the target has two dimensional attribute values of
(such as SOM, AHC, DYN, etc.) in well facies GR and CNL.
prediction. The legends in Figure 8 and Figure10 are In order to compare AMRGC quantitatively with
further illustrated in Figure 11. The point number column other unsupervised clustering algorithms, we divided
in Fig. 11 represents the number of sample points for each well section of the experiment into 10 small

5 5 5 5

3990

4000

4010

4020

4030

Fig. 10 The clustering results correspond to the special classification of lithofacies.

Classes id in GR(API) CNL(%) Nums of GR average CNL

Facies Type Lithology average(%)
AMRGC 0 150 -15 45 Points (API)

Dolomite
Facies 1 5 limestone
290 26.0 5.1

Argillaceous
Facies 2 4 dolomite
58 46.0 4.2

Containing
Facies 3 3 dolomite
130 63.6 6.4

Facies 4 2 Mudstone 132 103.0 9.0

Facise 5 1 Limy dolomite 231 25.7 3.0

Fig. 11 The correspondence of the facies types and the AMRGC clustering results.
21
 Adaptive multi-resolution graph-based clustering algorithm

sections to calculate the coincidence rate separately. The hyperparameters used in the experiment are shown in
coincidence rate is calculated by comparing the sum Table 2. The specific steps are as follows: first, randomly
of each actual length in the prediction result with the select 80% of the data points from 840 data points and
total length of the actual lithofacies. The final degree of apply the AMRGC algorithm to obtain the clustering
overall consistency is measured by an average. From result; then, use the obtained output as input for training
Figure 10 and Table 1, it can be seen that the AHC MLP; finally, apply the trained model to the remaining
algorithm is easy to confuse the two types of lithology, 20% of the data points are predicted. Experimental
which seriously affects the coincidence rate. For the results show that when the accuracy of the training data
AMRGC algorithm, it has a higher coincidence rate, so reaches 87.5%, MLP can obtain better and more stable
it has better discrimination ability for most lithologies. prediction ability than LPA. The results are shown in
Figure 13. Next, for comparison, we use 20% of the
Tab. 1 the conformity ratios given by methods
data set to train the MLP, and use the generated training
Algorithm Conformity Ratio
model to predict the labels for the remaining 80% of the
AHC 30.2%
DYN 77.5%
data points. However, because the amount of training
SOM 81.2% data at this time is much smaller than the amount of test
AMRGC 84.2% data, it cannot have a good generalization ability for
the test data. In this case, using semi-supervised LPA
Propagation prediction generally yields better results than using multi-layer
perceptual neural networks with supervised learning.
In the experiment of propagation prediction, a 4-layer
However, without loss of generality, that is, to ensure a
perceptual neural network was used. The specific
sufficient amount of data for training, it is not difficult

Table 2 The MLP Architecture and parameters

Parameters Value
Neuron in Input layers 2
Neuron in Output layers 5
Hidden layers 2
Neuron in each hidden layer 8,10
Iterations 500
Activation function in hidden and output layer Re-LU
Learning rate base 0.003
Learning rate decay 0.8
Decay step 5000
Regularization rate 0.0001
Number of samples selected for one training iteration 8
Probability of each network connection being kept 0.75

(a) (b)
12 12
10 Confound 10
8 8
1 1
CNL (%)

CNL (%)

6 2 6 2
3 3
4 4 4 4
5 5
2 Confound 2
0 0
0 50 100 150 0 50 100 150
GR (API) GR (API)
Fig. 12 Prediction results comparison based on Label propagation and MLP propagation.
(a) LPA based on KNN. (b) MLP based on BP.

22
 Wu et al.

to find that a multi-layered perceptual neural network Specifically, we compare the performance of AMRG
that has iterated 500 times has better propagation with other traditional clustering algorithms based
performance than LPA. The corresponding comparison on various indicators(as shown in Table 3). Specific
results are shown in Figure 14. indicators include whether prior knowledge is
In addition, we also compared the new method needed, processing efficiency in large-scale data, and
with the traditional unsupervised clustering method. characteristics of the propagation stage.

(a) (b)
25 25

20 20
1 1
15 15
CNL (%)

CNL (%)
2 2
10 3 10 3
4 4
5 5 5 5
Unbalanced 0 0
Confound
0 100 200 0 100 200
Unbalanced GR (API) GR (API)
Fig. 13 Results for a crossing validation. (a) LPA based on KNN. (b) MLP based on BP.

Tab. 3 AMRGC compared with other unsupervised clustering algorithms

Clustering Methods AMRGC MRGC SOM DYN AHC
Requires no prior knowledge of data structure   - - -
No Operator Bias   - - -
Faster against large datasets   -  -
More magical propagation stage  - - - -

Conclusions for cluster analysis ), While avoiding the shortcomings

of the MRGC method (such as computational efficiency).
The experimental results show that the newly proposed
This study addresses the challenges of traditional method performs well in the task of clustering high-
log facies cluster analysis algorithms that require dimensional data or data with complex structures. This
manually specifying the number of clustering kernels algorithm not only improves the operation efficiency
and the sensitivity of clustering results to initialization of the traditional MRGC algorithm, but also can better
values. The new proposed method (AMRGC) exposit predict the unlabeled data. In addition, without prior
the Light Kernel Representative Index (LKRI) to knowledge of the data, the new algorithm performs
determine the number of clustering kernels, and uses a significantly better than other clustering methods
combination of neural networks and clustering methods commonly used in the industry, such as self-organizing
to predict the data propagation. In general, it provides a mapping (SOM). Therefore, the new method will be
useful exploration of a combination of supervised and helpful for the clustering analysis of logging facies
unsupervised methods for logging facies cluster analysis with large-scale logging data as input. With the rapid
tasks. From the perspective of method implementation, development of the current deep learning field, this
this method is a multi-dimensional point pattern method still needs to explore new data propagation
recognition method based on non-parametric KNN methods in subsequent research work, including the use
and graph clustering, which takes into account the of deep learning technology, in order to better meet the
advantages of the MRGC algorithm (such as making full continuous data processing efficiency requirements.
use of high-dimensional information in geological data
23
 Adaptive multi-resolution graph-based clustering algorithm

References using a hybrid neural network and multi-resolution

graph-based clustering method case study from South
Pars Gasfield, Iran: Journal of Petroleum Geology,
Abrar, B., 2011, Integration of log data analysis
38(2), 177−191
and facies core to define electrofacies using multi
Pabakhsh, M., Ahmadi, K. and Riahi, M. A., and et al.,
resolution graph-based clustering: Iranlan Journal of
2012, Prediction of PEF and LITH logs using MRGC
Chemistry & Chemical Engineering- International
approach: Life Science Journal, 9(4), 974−982.
English Edition, 30(3),1−7.
Rogers, S. J., Fang, J. H., Karr, C. L., and et al., 1992,
Diday, E., 1971, Une nouvelle méthode en classification
Determination of lithology from well logs using a
automatique et reconnaissance des formes la méthode
neural network: AAPG Bulletin, 76(5), 731−739.
des nuées dynamiques: Revue de statistique appliquée,
Rosenblatt, F., 1958, The perceptron: a probabilistic
19(2), 19−33.
model for information storage and organization in the
Dodge, Y., and Commenges, D. 2006, The Oxford
brain: Psychological Review, 65(6), 386.
dictionary of statistical terms: Oxford University
Sutadiwirya, Y., Abrar, B., Henardi, D., and et al.,
Press, England, 119-370.
2008, Using MRGC (multi resolution graph-based
Hastie, T., Tibshirani, R., & Friedman, J., 2009, The
clustering) method to integrate log data analysis and
elements of statistical learning: data mining, inference,
core facies to define electrofacies, in the Benua Field.
and prediction: Springer Science & Business Media.
In Central Sumatera Basin, Indonesia: International
467−468.
Gas Union Research Conference(IGRC), Paris.
Hinton, G. E., Salakhutdinov, R. 2006, Reducing the
Currans Associates, Inc., 733−744.
dimensionality of data with neural networks: Science,
Tang, H., Meddaugh, W. S., and Toomey, N., 2011,
313(5786), 504−507.
Using an artificial-neural-network method to predict
Khoshbakht, F., Mohammadnia, M., 2010, Assessment
carbonate well log facies successfully: SPE Reservoir
of clustering methods for predicting permeability
Evaluation & Engineering, 14(01), 35−44.
in a heterogeneous carbonate reservoir: Journal of
Tian, Y., Xu, H. Zhang, X. Y., et al., 2016, Multi-
Petroleum ence & Technology, 2(2), 50−57.
resolution graph-based clustering analysis for
K o h o n e n , T. , 1 9 9 0 , T h e s e l f - o rg a n i z i n g m a p :
lithofacies identification from well log data: case study
Proceedings of the IEEE, 78(9), 1464−1480.
of intraplatform bank gas fields, Amu Darya Basin:
Lance, G. N., and Williams, W. T., 1967, A general
Applied Geophysics, 13(4), 598−607.
theory of classificatory sorting strategies Hierarchical
Tian, Y., Zhang, X. Y., Zhu, G. W., et al., Controlling
systems: Computer Journal, 9(4), 373−380.
effects of paleogeomorphology on intraplatform
Lukasová, A., 1979, Hierarchical agglomerative
shoal reservoirs distribution and gas reservoirs
clustering procedure: Pattern Recognition, 11(5−6),
characteristics: Taking intraplatform shoal gasfields
365−381.
of the Amu Darya basin as examples: Natural Gas
Mourot, G. 1993, Contribution au diagnostic des
Geoscience, 27(2), 320−329.
systèmes industriels par reconnaissance des formes
Ye, S. J., and Rabiller, P., 2000, A new tool for
(Doctoral dissertation): Doctorat del’Institut National
electro-facies analysis: multi-resolution graph-
Polytechnique de Lorraine, France.
based clustering: SPWLA 41st Annual Logging
Mourot, G., Bousghiri, S., and Ragot, J., 1993, Pattern
Symposium, Dallas,Texas, USA,Jun 4−7.
recognition for diagnosis of technological systems: a
Zhu, X. J., and Ghahramani, Z. B., 2002, Learning from
review: International Conference on Systems IEEE,
labeled and unlabeled data with label propagation:
275−281.
Carnegie Mellon University CALD tech report CMU-
Nouri‐Taleghani, M., Kadkhodaie‐llkhchi, A., & Karimi‐
CALD-02-107.
Khaledi, M., 2015, Determining hydraulic flow units

24
 Wu et al.

Appendix

Algorithm 1 The algorithm of getting the passage between attraction sets

1: Input: attraction sets (S), boundary points sets (BPS) , K’
2: Output: passage list (L)
3: for EACH i in [1, N] do
4: for EACH m in [1, len(Si)] do
5: minDis←Double_MAX
6: passage←0
7: if Pm ∈Si and Pm ∈DPS then
8: for Each j in [1, N] do
9: for EACH n in [1, len(Si)] do
10: if Pn ∈Si and Pn ∈DPS then
11: if Pm in the K’NN of Pn or Pn in the K’NN of Pm then
12: if D(Pm, Pn) < minDis then
13: minDis←D(Pm, Pn)
14: passage←M in (NI(Pm), NI(Pn))
15: end if
16: end if
17: end if
18: end for
19: end for
20: end if

Algorithm 2 The algorithm of merging groups

1: Input: value list L, selected cluster kernel set(CKS)
2: Output: merge results
3: L.sort (P1, P2, passage)by passage in decreasing order
4: for items in L do
4: if set S1 corresponding to point P1 contains a point in CKS then
5: if set S1 corresponding to point P1 contain a point in CKS then
6: continue
7: end if
8: end if
9: merge set S1 and S2
10: end for

Wu Hongliang, senior engineer, received a Ph.D in

Earth exploration and information
technology from Research Institute of
Petroleum Exploration & Development
in 2013. He is currently an enterprise
expert of China Petroleum Exploration
and Development Research Institute,
devoted to the research on well logging
processing and interpretation methods.
Email: [email protected]

25
View publication stats