You are on page 1of 21

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

net/publication/282471018

Developing Fracture Measure as an Index of Fracture Impact


on Well-Logs

Article  in  Energy Exploration & Exploitation · August 2015


DOI: 10.1260/0144-5987.33.4.555

CITATIONS READS

14 28,460

4 authors:

atie mazaheri Hossein Memarian


University of Tehran University of Tehran
5 PUBLICATIONS   22 CITATIONS    279 PUBLICATIONS   1,097 CITATIONS   

SEE PROFILE SEE PROFILE

Behzad Tokhmechi Babak Nadjar Araabi


University of North Dakota University of Tehran
108 PUBLICATIONS   614 CITATIONS    350 PUBLICATIONS   4,471 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

3D Reconstruction of Porous Media Based on 2D Microstructures and Investigating the Petrophysical Trends from Micro Scale to
Laboratory Scale View project

Decision Making under Uncertainty (Theoretical approach) View project

All content following this page was uploaded by Hossein Memarian on 19 October 2015.

The user has requested enhancement of the downloaded file.


ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 pp. 555–574 555

Developing fracture measure as an index of fracture


impact on well-logs

Atie Mazaheri1, Hossein Memarian1,2*, Behzad Tokhmechi1,3


and Babak Nadjar Araabi4,5
1
School of Mining Engineering, Faculty of Engineering, University of Tehran, Tehran, Iran
2
Department of Earth & Environmental Sciences, University of Waterloo, Ontario, Canada
3
School of Mining, Geophysics and Petroleum Engineering, University of Shahrood,
Shahrood, Iran
4
School of Electrical and Computer Engineering, Control and Intelligent Processing Center of
Excellence, Faculty of Engineering, University of Tehran, Tehran, Iran
5
School of Cognitive Sciences, Institute for Studies in Theoretical Physics and Mathematics,
Tehran, Iran
*Author for corresponding. E-mail: atie_mazaheri@ut.ac.ir (A. Mazaheri),
memarian@ut.ac.ir (H. Memarian),

(Received 11 March 2015; accepted 16 June 2015)

Abstract
Due to the three-dimensional nature of fractures, it is difficult to characterize
them completely and accurately. In this paper, a novel fractured zone detection
criterion, Fracture Measure (FM), is proposed. FM is a parameter calculated by
aperture, fracture type, azimuth and apparent distance. These factors have not
been incorporated in previous studies to detect fractured zones. This study
attempts to estimate FM by Artificial Neural Network to see if there is any
relation between FM and conventional logs and to check the generalization
ability of FM. Two datasets were used for the investigation: a real carbonate
reservoir of an oil field in Iran and a synthetic heterogeneous reservoir, here
incalled SYN. Comparing outputs of heterogeneous and homogeneous
conditions showed that the Classification Correctness Rate (CCR) of the model
in the homogeneous state was approximately 97%, and in the heterogeneous
condition, it was between 74% and approximately 92%. Generalization ability
in the homogeneous state varied from 91% to 94%, and in the heterogeneous
condition, varied from 52% to 86%. In the real dataset, ANN was able to
estimate FM with an average accuracy of approximately 80%and Classification
Correctness Rate (CCR) of approximately 100%, which shows that FM could be
modeled through well-logs. It is noteworthy that FM is capable of providing a
fuzzy measure for fracture study.

Keywords: Fuzzy fracture measure, Synthetic fractured reservoir, Fracture


azimuth, Fracture aperture
556 Developing fracture measure as an index of fracture impact on well-logs

1. INTRODUCTION
Natural fractures show complex behavior and have an important influence on porosity
and permeability and there by enhance production of oil in fractured carbonate reservoirs.
A complete description of fractures is difficult because of their three-dimensional nature.
Therefore, in fracture modeling, understanding fracture properties is significant (Sisler,
1971; Roehl and Choquette, 1985). Fractures can be identified by some techniques,
including core analysis, study in gseismic sections, well testing, drilling mud loss history,
Petrophysical Logs (PLs), etc. (Thompson, 2000; Nelson, 2001).
Earlier efforts to detect natural fractures consist of the use of sonic waves (Hsu et
al., 1987), wavelet transforms (Daiguji et al., 1997), core data (Song et al., 1998),
seismic data (Behrens et al., 1998), and also using velocity deviation logs (Flavio and
Gregor, 1999). Surjaatmadja et al. (2002) used a frequency decomposition method to
analyze down-hole data to identify natural fractures. Extrapolating fracture intensity
indicators to distant well locations contains high uncertainty. A new methodology,
based on neural network and fuzzy logic is devised for fracture modeling, using three-
dimensional geological and petrophysical attributes containing an error bar for
uncertainty assessment (Wong, 2003). Tran (2004) used a classification algorithm to
study and detect natural fractures, whereas Shen and Li (2004) proposed a combined
method for characterization of naturally fractured reservoirs. Application of the shear-
wave method to recognize faulted zones is studied in Dutta et al. (2007).The factor
analysis method to detect open fractures near the faults (Ozkaya and Siyabi, 2008) and
applying probabilistic decision trees to detect fracture corridors (Ozkaya, 2008) are
other techniques used by other researchers to detect and characterize natural fractures.
A method to model fracture porosity was presented by Yan et al. (2009). This study
asserted that identifying fracture locations and some of their properties is possible.
There are some important studies in which petrophysical logs are used to detect
fracture planes. In 2001, a wavelet transform was applied to porosity log data, and it
was proposed that high frequency variations indicate the existence of fractures (Sahimi
and Hashemi, 2001). To validate the model, permeability data were used in which
large enhancements inpermeability are interpreted to be compatible with the existence
of fractures. Martinez-Torres (2002) applied a fuzzy logic approach to different
petrophysical logs. In other studies, wavelet transforms were applied to petrophysical
logs (Mohebbi et al., 2007);
During his PhD dissertation, Tokhmechi extended novel approaches to detect
fractured zones using petrophysical logs. Wavelet, classification and data fusion
techniques have been used in his proposed methodologies. The generalization
capability of novel methodologies was confirmed with an accuracy between72% and
80% (Tokhmechi et al., 2009a; 2009b).
Tokhmechi et al. (2010), using energy of petrophysical logs, proposed a new
method for fracture density estimation in fractured zones. They demonstrated a
powerful correlation between fracture density and the caliper energy, sonic (DT),
density (RHOB) and photoelectric (PEF) logs within each well.
Toestimate fracture density by conventional well-logs, Ja’fari et al. (2012)
proposed a model using an adaptive neuro-fuzzy inference system. The results showed
a correlation coefficient of approximately 98% between the measured and neuro-
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 557

fuzzy-estimated fracture density. Additionally, for making a quantitative correlation


between fracture densities obtained from image logs and conventional well-log data,
Ja’fari et al. (2013), by integrating different artificial intelligent systems, proposed
amodified method.
Masoudi et al. (2014) fulfilled a research work on feature selection for reservoir
characterization. Correlation coefficients and Bayesian Network were used in this
study to find the most relevant data. They concluded that the most important well-logs
in fracture prediction are DT, SGR, RHOB, PEF and CGR, in order of importance.
Zerrouki et al. (2014) investigated the prediction of natural fracture porosity from
well-log data by means of fuzzy ranking and an artificial neural network. The
correlation coefficient () between the naturally fracture-related porosity obtained from
ANN and log data was equal to 87.8%.
Finally, in 2014, Xue et al. (2014) suggested a new approach for fracture detection
in terms of deep-shallow laterolog curves and micro-electrode logging curves, in
addition to their amplitude distinction. This is done by analyzing the relevance
between observed fractures in the cores and well-log datain-depth. Due to the good
stability of the prediction with an oilfield’s actual development performance, the
reliability of the new method is demonstrated (Xue et al., 2014).
Each of these methods is subjected to some limitations that prevent their full use in
different related engineering applications. Presently, image logs are the main
specialized tools for fracture detection; however, the response of this tool is not
limited beyond the wellbore. In addition, unfortunately, no image log is available for
many wells drilled prior to the availability of these new technologies. In most of the
mentioned fracture modeling methods, an inadequate volume of data can lead to false
reasoning interpretations (Tokhmechi et al., 2010; Ja’fari et al., 2013).
The authors of this current work believe that the difficulty of fracture interpretation
is multiplied by a binary (0/1) approach in studying or modeling fractures. Therefore,
using another source of available information to supplement our knowledge of fracture
properties is necessary. In this work, we propose an index called Fracture Measure
(FM) to show the fracture effectiveness in each horizon fuzzily. FM is aparameter
calculated by aperture, fracture type, azimuth and apparent distance. Finally, this study
attempts to estimate FM by ANN to see if there is any relation between FM and
conventional logs and checks the generalization ability of FM.

2. DATASETS
2.1. Real Data
The data set of this research belongs to one of the giant Iranian oil fields, located in
Zagros. More than 300 wells are drilled on this field, mostly for production purposes.
Conventional logs including caliper, density, neutron, sonic, resistivity and natural
gamma ray spectrometry are common in most of the drilled wells. Petrophysical logs
and interpreted image logs of the four wells were used to generate an estimator for
finding fractured zones within the Asmari reservoir (Table 1).
The Asmari reservoir is the target of the study within this giant field. A UGC map
of the top of Asmari and the well locations is shown in Figure 1.
558 Developing fracture measure as an index of fracture impact on well-logs

Table 1. Available well-logs for four wells in the study.


Well WR1 WR2 WR3 WR4
Caliper
DT
CGR
RHOB
PEF
SGR
NPHI
GR
POTA
URAN
THOR
PHIE
SWE
CALCITE
DOLOMITE
ANHYDRITE
SHALE
ILM
ILD
MSFL
LLS
LLD
VOL_CALCITE
VOL_SHALE

Figure 1. Pattern of fractures at the four studied wells, based on rose diagrams and
trends on stereonets.

2.2. Synthetic Data


In this study, for evaluation of the generalization ability of the proposed method,
first, we investigated the pattern of all open fractures at each of these wells based
on the rose diagram and trends presented on stereonets (Fig. 1). According to this
figure, there is no similarity between the fracture patterns of these wells,
potentially disrupting the testing of the generalization ability of the proposed
algorithm from one well to another. We also calculated a priori knowledge
 number of fractures 
  ( at each well (Table 2). Differences between the
 number of petrophysical data 
values of this factor for wells are illustrated. There is no data balance between them.
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 559

Nevertheless, we checked the generalization ability, but ANN and naïve Base were
unable to show an accurate generalization ability of FM.

Table 2. Calculation of a priori knowledge for each well.

Well A priori knowledge (%)


WR1 70.34
WR2 36.37
WR3 1.82
WR4 23.27

Although the generalization check on the real data did not show accurate outputs,
we were not disappointed about FM. Thus, we decided to generate synthetic data for a
conceptual reservoir, named SYN, shown in Figure 2. SYN stands for synthetic. Next,
we applied the methodology, considering two cases: a homogeneous SYN reservoir
and a heterogeneous SYN reservoir. Figure 2 shows the heterogeneous SYN reservoir,
which has a sandy interlayer. For the homogeneous SYN reservoir, the sandy layer is

Fi 2 S h i ll i SYN i
Figure 2. Synthetic wells in SYN reservoir.
560 Developing fracture measure as an index of fracture impact on well-logs

removed to investigate the effect of heterogeneity on the outputs. For this dataset, we
considered 2 wells (WS1 and WS2) with petrophysical logs: CGR, RHOB, DT and PEF.
Properties of each horizon are expressed in Table 3. Then, we examined the
generalization ability of the proposed method while one of these two wells was chosen
as an unobserved well.

Table 3. Properties of synthetic well-logs at depth intervals of wells. PH stands


for Property of Host rock, PF stands for Property of Fracture, and DI stands
for Depth of Investigation of tool.
WS1
Depth RHOB PEF CGR DT
Interval (m) DI=40 cm DI=10 cm DI=45 cm DI=10 cm
2000-2060 PH= 2.71, PF= 0.85 PH= 5.084, PF= 0.119 PH= 10, PF= 8.5 PH= 48, PF= 25.84
2060-2080 PH= 2.308, PF= 0.85 PH= 1.754, PF= 0.119 PH= 15, PF= 12.75 PH= 85, PF= 58.847
2080-2090 PH= 2.71, PF= 0.85 PH= 5.084, PF= 0.119 PH= 10, PF= 8.5 PH= 48, PF= 25.84
2090-2100 PH= 2.71, PF= 1.086 PH= 5.084, PF= 0.807 PH= 10, PF= 8.5 PH= 48, PF= 155.604
WS2
2030-2070 PH= 2.71, PF= 0.85 PH= 5.084, PF= 0.119 PH= 10, PF= 8.5 PH= 48, PF= 25.84
2070-2090 PH= 2.308, PF= 0.85 PH= 1.754, PF= 0.119 PH= 15, PF= 12.75 PH= 85, PF= 58.847
2090-2100 PH= 2.71, PF= 0.85 PH= 5.084, PF= 0.119 PH= 10, PF= 8.5 PH= 48, PF= 25.84
2100-2200 PH= 2.71, PF= 1.086 PH= 5.084, PF= 0.807 PH= 10, PF= 8.5 PH= 48, PF= 155.604

3. METHODOLOGY
In this paper, a novel fractured zone detection criterion, FM, is proposed. This measure
is a quantity that is related directly to fracture type, its aperture and the inverse of
distance.At first, a three-dimensional grid with cell dimensions of 30 cm in four
azimuths (0, 90, 180 and 270) is defined. The cell dimension was selected due to the
average depth of investigation of logging tools (Fig. 3). Thus, a 3-Dreservoirgrid, called
a prototype, consisting of four azimuthal sectors, was designed around the wellbore.

Figure 3. Depth of investigation and resolution of different petrophysical logs,


redrawn from Darling (2005).
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 561

Next, we updated values of the grid proto type using information obtained from
manual interpretation of image logs, such as fracture type, dips and its azimuth. The
formula of FM is presented below:

 
FM ( x ) = log ∑ AP × Type × cosα ×
 1 (1)
 D 

Where:

VDip .VCell
cosα = (2)
VDip VCell

D = d cell − d fracture (3)

where AP is aperture; Type is a weight corresponding to fracture type: continuous (1),


major (1), minor (0.5) and possible (0.25); is a measure of whether the fracture plane
passes through the center of the cell or not; is a unit vector, parallel to fracture dip; is
a unit vectortoward the cell center; and isa criterion for the apparent measurable
distance between the cell center and the fracture trace (Fig. 4).

Figure 4. Possible modes in encountering fracture with grid in 2D view.


562 Developing fracture measure as an index of fracture impact on well-logs

According to relation (1), we have calculated a fuzzy fracture measure index as the
summation of the effect of one or more fractures within or near each cell. Fractures up
to three cells up and down in each cell are considered in this summation.
It should be noted that these factors (aperture, fracture type, fracture azimuth,
vector of fracture gradient, vector of cell azimuth and criterion for apparent distance)
have not been incorporated in previous studies of fractured zone detection.
Undoubtedly, these parameters are related to the nature of the fracture and
differentiate the proposed method from conventional methods.

3.1. Neural network


An Artificial Neural Network (ANN) is a strong self-adaptive and self-learning tool
suitable for mapping highly nonlinear variables. Recently, the most commonly used
method to train ANN is the Back Propagation (BP) neural network (Xue et al., 2014).
In this work, ANN is utilized to estimate FM to see if ANN is able to establish any
non-linear relationship between well-logs and FM or not. If we are able to discover
any relation between logs and FM, it will show that FM is predictable by logs;
therefore, there is a logical relation. Based on this logical relation, one can estimate
FM by logs, there by studying the behavior of near well fractures.
In this section, we used the below algorithm for the estimation of FM with ANN:
a. Input selection
b. Dividing dataset: the whole dataset is randomly divided into three parts. For
real data: 70% for training the network, 15% for testing the trained network
and avoiding over-fitting issues. The remaining 15% is stored for validation
check. For synthetic data: 60% for training the network, 20% for testing, and
20% for validation check.
c. Normalizing input and output values due to training dataset.
d. Defining geometry of ANN (layers, neurons, functions and training method):
Trained ANNs, here, have only one mid-layer, i.e., in total, there are three
layers in the whole structure. In this work, the input layer consists of five
neurons, i.e., the caliper, corrected gamma ray (CGR), sonic (DT),
photoelectric (PEF) and density (RHOB) (Fig. 5). The Levenberg–Marquardt

Figure 5. Schematic structure of ANN.


ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 563

method is utilized for training the network, which is one of the reputable
supervised methods of optimization. The number of neurons in the hidden
layer is checked from 5 to 200 for Real Data and 5 to 100 for Synthetic Data
with steps of 5, but the most optimal structure is selected due to MSE. Every
structure is run 5 times for real data and 100 times for synthetic. The
activation function for input and mid-layers is a tangent sigmoid and for the
output layer is pure-line.
e. Training the networks and selecting the most precise network; after that,
simulating and plotting the results.

4. INPUT SELECTION
Caliper, corrected gamma ray (CGR), sonic (DT), photoelectric (PEF) and density
(RHOB) logs were selected as input features. To detect fractured zones, first,
webuilt3-dimensional gridded models for each well. Despite previous fractured
prediction methodologies that have a crisp approach, i.e., 0 for non-fractured intervals
and 1 for fractured horizons, we used a fuzzy measure (FM). FM is an indicator of how
a fracture or fractures are affecting a specific horizon in different azimuths. Then,
using thein formation obtained from interpretations of image logs and considering
factors associated with a 3D fracture model, we defined FM parameters.
According to Table 4, there are some related studies for selection of effective well-
logs. In this study, after evaluating available well-logs as an input parameter in
estimating the Fracture Measure index based on an iterative pruning algorithm, it was
realized that 5 common well-logs (CALI, DT, CGR, RHOB and PEF) in the studied
wells led to the best result based on the Mean Square Error (MSE) Criterion (Fig. 6).

Table 4. Input parameters for predicting fractures in the literature.


Parameters used for Source
identification
Fracture Water saturation, GR (Tokhmechi et al., 2009b)
Caliper, DT, RHOB, PEF (Tokhmchi et al., 2010)
DT, RHOB, NPHI, resistivity (Ja’fariet al., 2012)
Deep Resistivity Laterolog (RLLD), Shallow Resistivity Laterolog
(RLLS),the amplitude difference of deep and shallow resistivity (Xue et al., 2014)
laterolog (∆DS), mini-log (VDX-W, micro inverse (VDX-T), the
amplitude difference of mini-log and micro-inverse(∆WT)
GR, NPHI, RHOB, saturation of oil (Tahar Aïfa et al., 2014)
DT, SGR (Masoudi et al., 2014)
Fracture log data (RHOB, NPHI and transit time) and the characteristics of the
Porosity mud (fluid density, transit time of the saturating (Zerrouki et al., 2014)
fluid)
564 Developing fracture measure as an index of fracture impact on well-logs

Figure 6. Relation between Mean Square Error (MSE) and inputs in order to find
optimal inputs for WR2of Asmari reservoir.

5. RESULTS
5.1. Real Data
In this study, we presented a fuzzy measure to model fracture intensity by
conventional well-logs within oil wells. Modeling was conducted by training ANN
using real and synthetic data. Five selected logs for real field and four produced logs
for a SYN reservoir were used as the input layer and FM as the output neuron.
Two strategies were tested in estimating FM: 1-estimating FM in all four azimuths
of each horizon (around the well); and 2-estimating the summation of FM of four
azimuths in each horizon (single value). In the first strategy, because of the increasing
volume of data (four times, compared to the second strategy), a better training of
neural network was obtained, improving the estimation accuracy.
According to Table 5, the best results were obtained for the all azimuth strategy,
based on the criteria of mean square error, correlation coefficient, confusion matrix
and classification correctness rate. Figure 7 shows cross plots of validation datasets
after estimation by ANN in each of the four real wells. By means of FM, trained ANN
is not only able to distinguish fractured and non-fractured intervals but can also
prioritize the fracture effectiveness in each horizon due to the FM quantity. It is
noteworthy that FM of -4 denotes “no fracture” in the model. In fact, the proposed
index is a criterion for determination of the impact of fractures on well-logs.
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 565

Table 5. Mean Squared Error (MSE), Correlation Coefficient (R), Confusion


Matrix and Classification Correctness Rate (CCR) of two cases, namely
summed and all azimuths. The correlation coefficients presented only belong to
fractured zones, i.e., FM≠-4, to avoid overestimation of R. In the confusion
matrix, R stands for real and C stands for classified.
Mean Squared Error
(MSE), Correlation
Coefficient (R) and Summed All azimuths
Classification
Correctness Rate (CCR
MSE R CCR MSE R CCR
WR1 33.03 55% 0.46 72%

CCR=0.73 CCR=0.94

WR2 2.33 74% 0.003 95%

CCR=1.00 CCR=1.00

WR3 1.54 62% 0.03 78%

CCR=0.58 CCR=0.99

WR4 3.07 26% 0.04 73%

Figure 8 compares trends of calculated and estimated FM to real fractures in WR1.


It is shown that there is a good compatibility between FM, estimated FM and fractures.
Fuzzy inherit of FM is the other advantage of this measure for presenting the
effectiveness of fractures on well-logs.
566 Developing fracture measure as an index of fracture impact on well-logs

a) Validation set of WR1 b) Validation set of WR2

c) Validation set of WR3 d) Validation set of WR4

Figure 7. Correlation between estimated FM by the proposed method (Target) and


neural network (Output) on validation datasets of a) WR1, b)WR2, c) WR3 and d)WR4.
The correlation coefficients presented only belong to fractured zones, i.e., FM≠-4, to
avoid overestimation of R.
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 567

a) interval: 2370-2470meters b) interval: 2470-2570 meters c) interval: 2570-2690 meters

Figure 8. Continued
568 Developing fracture measure as an index of fracture impact on well-logs

d)interval: 2400-2450 meters

Figure 8. Comparing calculated and estimated FMs within whole WR1 (a, b, c) and
between depths of 2400-2450 meters of WR1 (d).

5.2. Synthetic Data


Finally, the generalization of the proposed method was investigated with two synthetic
wells of the SYN reservoir considering two cases: a homogeneous state (sandy part is
removed) and a hetero generous state (sandy part is not removed). There is a vast
difference in accuracy in these two states. The results show the necessity of the
homogeneous state in modeling FM using well-log data (Table 6).
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 569

Table 6. Mean Squared Error (MSE), Correlation Coefficient (R), Confusion


Matrix and Classification Correctness Rate (CCR)for training and
generalization stages (summed strategy, in two states: homogeneous and
heterogeneous).The Correlation Coefficients presented only belong to fractured
zones, i.e., FM≠-16,to avoid overestimation of R. In the confusion matrix, R
stands for real and C stands for classified.
Consider homogeneity
Mean Squared Error
(MSE), Correlation
Coefficient (R) and
Classification Correctness
Rate (CCR) Test wells
WS1 WS2
Training Wells MSE R CCR MSE R CCR
WS1 3.47 82% 8.49 54%

CCR=0.97 CCR=0.94

WS2 22.95 53% 3.02 78%

CCR=0.91 CCR=0.97

Consider heterogeneity
Mean Squared Error
(MSE), Correlation
Coefficient (R) and
Classification Correctness
Rate (CCR) Test wells
WS1 WS2
Training Wells MSE R CCR MSE R CCR
WS1 21.55 39.60

CCR=0.74 CCR=0.52

WS2 61.94 10.11

CCR=0.86 CCR=0.92

Figure 9 shows the correlation between the calculated FM and its estimation by neural
network for synthetic wells for each well and generalization stage. In this figure, the
fractured and non-fractured zones are illustrated by blue cadres, and the uncertainty
interval is shown by the dashed red lines. Calculated correlation coefficients (R) only
570 Developing fracture measure as an index of fracture impact on well-logs

belong to data limited to the fractured zone. R in training wells is 78% to 82% for
validation data and 53% to 54% for generalization.

a) Validation set of WS1 b) Validation set of WS2

c) Estimating FM in WS2, while WS1 is the training dataset

Figure 9. Correlation plots for estimating FM, for checking MSE in training (a, b)
and generalization (c, d) states. The correlation coefficients presented only belong to
fractured zones, i.e., FM≠-16, to avoid overestimation of R.
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 571

6. CONCLUSIONS
In this study, we proposed an index called FM to show the effectiveness of fractures
in each horizon fuzzily. FM could be calculated with well-log data, aperture, fracture
type, etc. A major advantage of FM is in presenting a fuzzy measure instead of a crisp
(0/1) output for a fracture study. In the next stage, we used an Artificial Neural
Network to check whether we can model FM by logs or not. Investigations were
conducted using two datasets. One data set was obtained from a real carbonate
reservoir, whereas the other was from a synthetic heterogeneous reservoir, called
SYN. Comparing results in the heterogeneous and homogeneous conditions showed
that the Classification Correctness Rate (CCR) of the model in the homogeneous state
is approximately 97%, and in the heterogeneous condition, it is between 74% and
approximately 92%. Additionally, the generalization ability in the homogeneous state
varies from 91% to 94%, and in the heterogeneous condition, it varies from 52% to
86%. In the real dataset, ANN was able to estimate FM with an average accuracy of
approximately 80% and a Classification Correctness Rate (CCR) of approximately
100%, which shows that FM is determinable through well-logs. Correlation plots and
comparison of logs with real fractures showed compatibility of modeled FM with
observed fractures on the wellbore.

ACKNOWLEDGEMENTS
The authors would like to declare their sincere thanks to the National Iranian South Oil
Company (NISOC) and Mrs. Mohammadian and Mr. Kordavani of the NISOC for
their support in providing the information and their technical input regarding this
work.

Nomenclature
ANN Artificial Neural Network
AP aperture
CCR Classification Correctness Rate
CGR Corrected Gamma Ray
d depth
D distance
DT sonic transferring time
DI Depth of Investigation of tool
FM Fracture Measure
GR Gamma Ray
ILD InductionLog, Deep
ILM Induction Log, Medium
LLD Laterolog, Deep
LLS Laterolog, Shallow
MSE Mean Square Error
MSFL Micro Spherically Focused Log
NPHI Neutron Porosity
PEF Photo-Electric Factor
PF Property of Fracture
572 Developing fracture measure as an index of fracture impact on well-logs

PHIE effective porosity


POTA potassium
PH Property of Host rock
R Correlation Coefficient
RHOB bulk density
SGR Summation of Gamma Rays
SWE Water Saturation
SYN Synthetic
THOR thorium
URAN uranium
V vector

REFERENCES
Aïfa T., Zerrouki A.A, Baddari K. and Géraud Y., 2014. Magnetic susceptibility and
its relation with fractures and petrophysical parameters in the tight sand oil
reservoir of Hamraquartzites, southwest of the Hassi Messaoud oil field ,
Algeria. Journal of Petroleum Science and Engineering 123, 120-137.
Daiguji M., Kudo O. and Wada T., 1997. Application of wavelet analysis to fault
detection in oil refinery. Computers and Chemical Engineering 21, S1117-
S1122.
Darling T., 2005. Well Logging and Formation Evaluation. Elsevier, pp. 346.
Dutta P., Mardi C., Singh S.K., Al-Genai J., Akhtar A. and Akbar M., 2007. A novel
approach to fracture characterization utilizing borehole seismic data. 15th SPE
Middle East Oil and Gas Show and Conference. Manama, Bahrain, March, 2007.
Society of Petroleum Engineers, pp. 6.
Flavio S.A. and Gregor P.E., 1999. The velocity-deviation log a tool to predict pore
type and permeability trends in carbonate drill holes from sonic and porosity or
density logs. AAPG bulletion 83(3), 450-466.
Hsu K., Brie A. and Plumb R.A., 1987. A new method for fracture identification using
array sonic tools. Journal of petroleum technology 39(06), 677-683.
Ja’fari A., Kadkhodaie A., Sharghi Y. and Ghanavati K., 2012. Fracture density
estimation from petrophysical log data using the adaptive neuro-fuzzy inference
system. Journal of Geophysics and Engineering 9(1), 105-114.
Ja’fari A., Kadkhodaie A., Sharghi Y. and Ghaedi M., 2013. Integration of Adaptive
Neuro-Fuzzy Inference System, Neural Networks and Geostatistical Methods for
Fracture Density Modeling. Oil and Gas Science and Technology-Revue d’IFP
(Impact Factor: 1.11), 69(7), 1-15.
Martinez-Torres L.P., 2002. Characterization of naturally fractured reservoirs from
conventional well logs: Doctoral dissertation, University of Oklahoma USA,
Oklahoma, USA, pp. 178.
Masoudi P., Asgarinezhad Y. and Tokhmechi B., 2014. Feature selection for reservoir
characterization by Bayesian network, Arabian Journalof Geosciences 8(5), 1-
13.
ENERGY EXPLORATION & EXPLOITATION · Volume 33 · Number 4 · 2015 573

Mohebbi A.R., Haghighi M. and Sahimi M., 2007. Using conventional logs for
fracture detection and characterization in one of Iranian field. International
Petroleum Technology Conference. UAE, Dubai, December 4-6, 2007,
Australian School of Petroleum publications, pp. 4-6.
Nelson R.A., 2001.Geologic Analysis of Naturally Fractured Reservoirs, 2nd Edition.
Gulf Professional Publishing, Oxford, pp. 332.
Ozkaya S.I. and Siyabi S., 2008. Detection of fracture corridors from dynamic data by
factor analysis. In SPE Saudi Arabia Section Technical Symposium, Alkhobar,
May 10-12, 2008, Society of Petroleum Engineers, pp. 10.
Ozkaya S.I., 2008. Using probabilistic decision trees to detect fracture corridors from
dynamic data in mature oil fields. SPE-105015-PA Reservoir Evaluation and
Engineering 11(06), 1061–1070.
Roehl P.O. and Choduette P.W., 1985. Carbonate Petroleum Reservoirs. Spring-
Verlag, New York, pp. 622.
Sahimi M. and Hashemi M., 2001. Wavelet identification of the spatial distribution of
fractures. Geophysical research letters, 28(4), 611-614.
Shen F. and Li S.Q., 2004. A combined geological, geophysical and rock mechanics
approach to naturally fractured reservoir characterization and its applications. In
Annual Technique Conference and Exhibition, Houston, Texas, U.S.A.,
September 26-29, 2004. Society of Petroleum Engineers Inc, pp. 1-8.
Sisler J.J., 1971. Study of Asmari Core Material from Wells Parsi 15 and 16 Parsi
Field Iran. OSCO Report, pp. 19-45.
Song X.M, Zhu Y.X, Liu Q., Chen J., Ren D.X, Li Y., Wang B. and Liao M.S, 1998.
Identification and distribution of natural fractures. Sixth International Oil and
Gas Conference and Exhibition in China., China , Beijing, November 2-6, 1998,
SPE, Richardson TX, ETATS-UNIS.
Surjaatmadja J.B., Stephenson S., Bhaumik C., Thompson S. and Cheng A., 2002.
Analysis of generated and reflected pressure waves during fracturing reveals
fracture behavior. In SPE Annual Technical Conference and Exhibition, San
Antonio, Texas, 29 September - 2 October, Society of Petroleum Engineers, pp.
19.
Tran N.H., 2004. Characterization and Modeling of Naturally Fractured Reservoirs:
Ph.D. Thesis, University of New South Wales Australia, Australia, pp. 182.
Tokhmechi B., Memarian H., Noubari H.A. and Moshiri B., 2009.A novel approach
proposed for fractured zone detection using petrophysical logs. Journal of
Geophysics and Engineering 6(4), 365.
Tokhmechi B., Memarian H., Rasouli V., Noubari H.A. and Moshiri B., 2009.
Fracture detection from water saturation log data using a Fourier-wavelet
approach. Journal of Petroleum Science and Engineering 69(1), 129-138.
Tokhmechi B., Memarian H. and Rezaee M.R., 2010. Estimation of the fracture
density in fractured zones using petrophysical logs. Journal of Petroleum
Science and Engineering 72(1), 206-213.
574 Developing fracture measure as an index of fracture impact on well-logs

Thompson L.B., 2000. Fractured reservoirs: integration is the key to optimization.


Journal of petroleum technology 52(02), 52-54.
Wong P.M., 2003. A novel technique for modeling fracture intensity: A case study
from the Pinedale anticline in Wyoming. AAPG bulletin 87(11), 1717-1727.
Xue Y., Cheng L., Mou J. and Zhao W., 2014. A new fracture prediction method by
combining genetic algorithm with neural network in low-permeability
reservoirs. Journal of Petroleum Science and Engineering 121,159-166.
Yan J., Lu L.W., Lubbe R. and Payne S., 2009. Petrophysical fracture identification
for rock physics studies. In 71st EAGE Conference and Exhibition, The
Netherlands, Amsterdam, June 8-11, 2009. EAGE Publications, pp. 1-5.
Zerrouki A.A., Aïfa T. and Baddari K., 2014. Prediction of natural fracture porosity
from well log data by means of fuzzy ranking and an artificial neural network in
Hassi Messaoud oil field, Algeria. Journal of Petroleum Science and
Engineering 115, 78-89.

View publication stats

You might also like