You are on page 1of 14

SPE-197396-MS

Application of Machine Learning Approach for Intelligent Prediction of Pipe

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Sticking

Ahmed K. Abbas, Iraqi Drilling Company; Haidar Almubarak, King Saud University; Hayder Abbas, Missan Oil
Company; Jawad Dawood, Basra Oil Company

Copyright 2019, Society of Petroleum Engineers

This paper was prepared for presentation at the Abu Dhabi International Petroleum Exhibition & Conference held in Abu Dhabi, UAE, 11-14 November 2019.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
Stuck pipe has been recognized as one of the serious problems in drilling operations that has a significant
impact on drilling efficiency and well costs. The events related to the stuck pipe can be responsible for
losses of hundreds of millions of dollars each year in the drilling industry. This paper presents a study on the
application of machine learning methodologies to predict the stuck pipe occurrence which can be utilized
to modify drilling variables to minimize the likelihood of sticking. The new models were developed to
predict the stuck pipe incidence for vertical and deviated wells using artificial neural networks (ANNs) and
a support vector machine (SVM). The proposed models were examined using a few examples of real stick
pipe cases from the field. The results of the analysis have revealed that both ANNs and SVM approaches
can be of great use, with the SVM results being more promising.
The present analysis supplies knowledge that can be used during well pre-planning and developmental
phases to make informed decisions that will avoid pipe sticking problems and essentially optimize drilling
performance. The risk of pipe sticking can then be minimized and the costs associated with its occurrence
will be reduced.

Introduction
Drilling operation is one of the most expensive and necessary part of petroleum and gas exploration (Abbas
et al., 2018a). One of the most prominent troubles occurring while drilling operations is stuck pipe incidents.
These incidents generally result in the significant amount of non-productive time in terms of loss of rig days
either due to the stopping of drilling operations or an attempt to free the stuck pipe (Rostami and Khaksar
Manshad, 2014). Pipe sticking problems are estimated to cost the industry about 250 millions of dollars
per year (Bradley et al., 1991).
The stuck pipe is identified as the resistance of drilling mud flow in the annular space and the drill string
cannot be rotated or moved either in the upward or downward direction. In a complete stuck pipe situation,
neither circulation nor drill string movement is possible (Siruvuri et al., 2006). Generally, stuck pipe
mechanisms are divided into two major categories: differential sticking and mechanical sticking (Shivers
and Domangue, 1993; Santos, 2000). Differential sticking occurs when the drill string or tool is stationary
2 SPE-197396-MS

(or sometimes when it is moving very slowly) across the permeable zone. A major cause of differential
stuck pipe is due to differential pressure forces from an overbalanced drilling mud column acting on the drill
string against a filter cake deposited on porous and permeable formations such as sandstone and limestone
(Krishan et al., 2000). The portion of the area of the drill string, casing, or logging tools that is embedded
into the mud cake has a pressure equal to the formation pressure acting on it, while the hydrostatic pressure
of the drilling mud column acts on the other side. When the hydrostatic pressure in the wellbore exceeds
the formation pressure, there will be a net force pushing the drill string towards the borehole wall (Brandon

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


et al., 1993). The resultant force of the overbalance acting on an area of drill string is the force that sticks
the string (Aadnøy et al., 2003). Commonly, this type of sticking does not occur in very low permeability
formations such as shales, where mud filter cakes normally do not form (Shadizadeh et al., 2010). As the
second category of stuck pipe, mechanical sticking is caused by physical obstruction or restriction (Reid
et al., 2000). This type of sticking usually occurs when the drill string is moving. Mechanical sticking can
be classified into two main subcategories: (1) hole pack-off and bridges; this refers to stuck pipes which
are related to wellbore instability or settled cuttings, and (2) wellbore geometry interferences; stuck pipes
which are related to the condition of wellbore geometry such as key seats or an under-gage hole. Major
causes of this type of sticking are wellbore instability and improper borehole cleaning. Wellbore instability
problems mainly caused shale instability by shale/fluid interaction or shear failure (borehole breakouts) due
to excessively low wellbore hydrostatic pressure (Abbas et al., 2018b). This includes unconsolidated rock,
swelling shale, flowing formations such as salt and plastic shale, and geo-pressured formations (Mohammed
et al., 2018). On the other hand, the drilled cuttings in the borehole, if not being appropriately removed, can
settle around the drill string. Such an issue, if not treated on time, can result in the drill string to become
stuck.
Investigating the various causes and symptoms that lead to stuck pipe makes it possible to early
identification of the most likely cause of this problem's occurrence. Prediction of stuck pipe can be
considered as the preventive treatments through which the risk of getting stuck can be minimized by
modifying drilling variables for the condition of the high risk of sticking. The difficulty of detecting
upcoming stuck pipe comes from its dependence on a wide range of drilling parameters and its complex
interrelationships between these different factors. In the past, several researchers tried to utilize multivariate
statistical analysis techniques to find prevailing differences in drilling parameters between wells which were
non-stuck and wells which were stuck. These methods use the drilling parameters linearly or exponentially
modified by constants as inputs. Therefore, there are many cases where the values of these constants cannot
be explained, or are against the nature of the physical phenomenon itself.
Because of this complexity, and because of this restriction of the statistical models that were developed
earlier, machine learning techniques look like an attractive alternative to model this complicated physical
process (Abbas et al., 2019a). An important characteristic regarding machine learning techniques are their
ability to generalize a result based only on the knowledge contained in a dataset. Because of that, they have
been used to treat system identification and function approximation problems such as the one focused herein.
Usually, artificial neural networks (ANNs) and support vector machine (SVM) are used to identify complex
relationships when sufficient data exists. Many research papers succeeded in modeling many of the oil and
gas industry processes using ANNs and SVM (Salehi et al., 2009; Amer et al., 2017; Alkamil et al., 2018).
The scope of this paper is to investigate the various conditions that lead to the pipe sticking problems using
drilling data of previously drilled wells in Southern Iraq. This investigation makes it possible to develop
some preventive treatments to avoid this problem's occurrence. In this study, ANNs and SVM are employed
to analyze drilling data to predict the risk of lost circulation while drilling and tripping operations.
SPE-197396-MS 3

Data Pre-processing
A total of 1010 datasets were collected from daily drilling reports, daily mud reports, final well reports, and
master logs of drilled wells in Southern Iraq. After analyzing recent stuck pipe incidents thoroughly and
investigating the root causes, the mechanical (pack off/bridging) sticking was the major cause of stuck pipe.
Therefore, the focus of this study is on the mechanical pipe sticking prediction. According to the primary
operational modes (drill string status), the datasets can be divided into three categories: stuck pipe while

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


drilling, stuck pipe during tripping in, and stuck pipe while tripping out. This classification is crucial because
each situation has an effect on pipe sticking occurrence. For these categories, each set has a target that was
identified by a stuck index as pipe sticking (PS) or non- pipe sticking (NPS). The number of datasets and
relative percentage of each stuck category of datasets are presented in Table 1.

Table 1—Classification of the used datasets.

Dataset Stuck Index Dataset Number Percentage (%)

Pipe Sticking (PS) 87 33.46


Stuck Pipe While Drilling
Non- Pipe Sticking (NPS) 173 66.54
Pipe Sticking (PS) 169 45.68
Stuck Pipe During Trip In
Non- Pipe Sticking (NPS) 201 54.32
Pipe Sticking (PS) 203 53.42
Stuck Pipe During Trip Out
Non- Pipe Sticking (NPS) 177 46.58

As shown in Table 2 and Table 3, numbering classes are applied to convert the lithology type and stuck
index into a numeric form, respectively (Abbas et al., 2019b). The collected data were normalized into a
form that can be used with the ANNs or the SVM. Further details of normalization data preprocess were
comprehensively described in the studies conducted by Abbas et al. (2019b).

Table 2—Generated code for lithology type.

Lithology Code

Limestone 0
Shale 1
Sandstone 2
Dolomite 3
Dolomite Limestone 4
Anhydrite 5
Gypsum 6
Chalky Limestone 7
Marly Limestone 8
Argillaceous Limestone 9
Shaly Limestone 10

Table 3—Generated code for stuck index.

Condition Code

Pipe Sticking (PS) 0


Non- Pipe Sticking (NPS) 1
4 SPE-197396-MS

Understanding the effect of the input parameters is considered the primary concern when developing
machine learning models. In this study, the fscaret package of R environment was applied to justify the worth
of the input parameters in predicting the pipe sticking occurrence (Szlek, and Mendyk, 2018). This method
is a simple package providing fast and automated feature ranking based on the caret package (Kazemi et
al., 2016). It uses a group of models (i.e., 103 various models) to determine the significance of variables.
The importance of features is averaged through all models to select the best relevant predictors for building
a model. In the end, the result is scaled from 0 to 100. A total number of 25 parameters were selected

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


as significant factors in the stuck pipe happening. The constant or nearly constant parameters had to be
excluded, with the remaining parameters considered to be the input parameters to predict the stuck pipe
incidence. Figures 1, 2, and 3 show the results of the variable ranking and the importance achieved by
using fscaret for stuck pipe (1) while drilling, (2) during trip in, and (3) while tripping out, respectively.
The final selected input parameters consisted of 20 variables for predicting stuck pipe while drilling and
19 variables for predicting stuck pipe during tripping (in or out) operations, as displayed in Tables 4, 5,
and 6, respectively.

Figure 1—Ranking of variables for stuck pipe while drilling dataset.

Figure 2—Ranking of variables for stuck pipe while tripping in dataset.


SPE-197396-MS 5

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Figure 3—Ranking of variables for stuck pipe while tripping out dataset.

Table 4—Range of input parameters for stuck pipe while drilling dataset.

Parameter Minimum Maximum

Inclination (degree) 0.25 78


Rate of penetration (m/hr) 1 32
Flow rate (l/min) 900 2850
WOB (ton) 1 30
Rotation per minute (rev/min) 40 265
Torque (klb/ft) 1 22
Lithology 0 10
BHA length (m) 100 295
Hole size (in) 8.5 16
Mud weight (g/cc) 1.02 1.4
Yield point (g/100 cm ) 2 6 40
Plastic viscosity (cp) 8 36
Marsh funnel viscosity (Sec.) 35 79
Gel 10″ (g/100 cm ) 2 2 17
Gel 10′ (g/100 cm2) 3 29
Circulating pressure (psi) 600 3100
Fluid loss (cm3/min) 1 22
Solids content (vol. %) 1 27
Azimuth (degree) 0 360
Measured depth (m) 500 4250
6 SPE-197396-MS

Table 5—Range of input parameters for stuck pipe while tripping in dataset.

Parameter Minimum Maximum

Inclination (degree) 0.25 83


Lithology 0 10
Drag force (ton) 1 25
BHA length (m) 120 295

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Hole size (in) 8.5 16
Fluid loss (cm /min)
3 1 24
Solids content (vol. %) 1 28
Mud weight (g/cc) 1.01 1.39
Rotation per minute (rev/min) 0 150
Torque (klb./ft) 0 20
Yield point (g/100 cm2) 5 51
Plastic viscosity (cp) 6 40
Gel 10″ (g/100 cm ) 2 1 16
Gel 10′ (g/100 cm ) 2 1 33
Marsh funnel viscosity (Sec.) 40 75
Flow rate (l/min) 0 2667
Circulating pressure (psi) 0 2700
Azimuth (degree) 0 360
Measured depth (m) 598 4050

Table 6—Range of input parameters for stuck pipe while tripping out dataset.

Parameter Minimum Maximum

Inclination (degree) 0.25 80


BHA length (m) 115 290
Hole size (in) 8.5 16
Lithology 0 10
Pull force (ton) 1 25
Fluid loss (cm /min)
3 1 20
Solids content (vol. %) 2 32
Mud weight (g/cc) 1.01 1.43
Yield point (g/100 cm2) 7 57
Plastic viscosity (cp) 5 39
Marsh funnel viscosity (sec) 38 79
Rotation per minute (rev/min) 0 130
Flow rate (l/min) 0 2700
Torque and drag (klb./ft) 0 23
Gel 10″ (g/100 cm ) 2 1 23
Gel 10′ (g/100 cm ) 2 1 34
Azimuth (degree) 0 360
Measured depth (m) 855 4231
Circulating pressure (psi) 0 2750
SPE-197396-MS 7

Model Performance Evaluation


Commonly, to examine the generalization ability of the models, the normalized dataset is randomly divided
into two parts: a training dataset is applied to adjust and optimize the control parameters of the model, and
a testing dataset is applied to examine the final performance of the developed model (Anemangely et al.,
2019). However, because of the limited number of datasets, this is considered a poor evaluation method
for this model's performance (Cawley and Talbot, 2003). In low- and medium-sized datasets, achieving

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


an accurate model by chance is possible. One of the techniques to assess a comprehensive evaluation of
model performance is a 10-fold cross validation (10CV) approach (Bahrami et al., 2016). In this technique,
the dataset is split into10 parts where; at first step, the 9 subsets are all chosen as the training subset,
and the remaining one of the 10 subsets is considered as the test subset (Almubarak et al., 2017). This
technique is repeated 10 times such that at each step one of the parts is used once as the testing subset.
The average of the accuracy, sensitivity, and specificity is then calculated over all 10 rounds to provide an
estimate of the model's predictive performance. The 10-fold cross validation (10CV) method has a lower
variance compared to other methods, and it always confirms that the model is generalized and not over-
fitted (Eskandarian et al., 2017). Figure 4 illustrates the 10CV procedure.

Figure 4—Schematic of the 10-fold cross validation (10CV) method.

To demonstrate the robustness of the developed models, accuracy, sensitivity, and specificity were used.
These statistical measures of the performance were conducted to evaluate the model's goodness of fit for a
binary classification test. To discriminate between the actual class and the predicted class, some definitions
such as true positive (TP), false positive (FP), true negative (TN), and false negative (FN) were used for
the class predictions generated by a classifier (Jahanbakhshi and Keshavarzi, 2016). There are two possible
consequences of the test outcome: (1) positive (PS occurrence) or negative (non-PS) according to this
scenario:
➢ True positive (TP): Pipe sticking (PS) cases are correctly identified as pipe sticking (PS).
➢ False positive (FP): Non-pipe sticking (NPS) cases are incorrectly identified as pipe sticking (PS).
➢ True negative (TN): Non-pipe sticking (NPS) cases are correctly identified as non-pipe sticking
(NPS).
➢ False negative (FN): Pipe sticking (PS) cases are incorrectly identified as non-pipe sticking (NPS).
Based on the previous four scenarios, the accuracy, sensitivity, and specificity can be obtained by use
of the following equations:
(1)
(2)
(3)
8 SPE-197396-MS

Artificial Neural Networks (ANNs)


Neural networks are powerful tools for the approximation of unknown nonlinear functions and have
gained wide applications in a variety of fields (Khamehchi et al., 2014). ANNs are information processing
systems that are a rough approximation and simplified simulation of a biological learning process and have
performance characteristics similar to those of biological neural networks. The network structures of ANNs
are made up of several neurons which are distributed in layers based on their different functions. Generally,

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


a complete neural network consists of three different types on layers, namely, an input layer, one or more
hidden layers, and an output layer in which each layer includes a preset number of neurons (Fig. 5) (Shi et
al., 2016). Hence, a challenging step in ANNs modeling is tuning of these parameters, which is mostly done
by try and error technique. In ANNs, arriving signals (inputs) to each neuron in hidden layers, are multiplied
by the adjusted weight elements (w1j … wnj), combined, and then flow forward through a transfer function
to produce the outputs for neurons. The most popular types of ANN are feed-forward networks trained with
a back-propagation algorithm (FFBP) which are applied in this study. The term back propagation refers to
the mechanism of adjusting and the connection weights and besides of the network by applying a two-phase
propagate-adapt cycle (Abbas et al., 2019c).

Figure 5—ANNs structure with one hidden layer.

In developing the networks, the TANSIG and LOGSIG transfer functions were examined for one, two,
and three hidden layers for stuck pipe while drilling, stuck pipe during a trip in, and stuck pipe while tripping
out, as shown in Figures 6, 7, and 8, respectively. Also, the optimal number of neurons in hidden layers
was determined by increasing the numbers incrementally. As can be seen in these figures, the accuracy
between the predicted results and the measured values increased as the number of neurons in the hidden
layers gradually increased, except at the individual fluctuant points. However, the accuracy values decreased
when the number of hidden neurons included more than 40 nodes. This indicates that the number of neurons
in hidden layers with more than 40 nodes can negatively affect the final accuracy of ANNs. For the TANSIG
transfer function, the accuracy reached its highest values at 0.83 with 40 nodes and two hidden layers for
stuck pipe while drilling, 0.87 with 35 nodes and two hidden layers for stuck pipe during trip in, and 0.87
with 35 nodes and three hidden layers for stuck pipe while tripping out (Figs. 6a, 7a, and 8a, respectively).
In contrast, the accuracy attained its highest points for the LOGSIG transfer function at 0.81 with 40 nodes
and two hidden layers for stuck pipe while drilling, 0.83 with 45 nodes and two hidden layers for stuck
pipe during trip in, and 0.85 with 40 nodes and three hidden layers for stuck pipe while tripping out (Figs.
6b, 7b, and 8b, respectively).
SPE-197396-MS 9

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Figure 6—Different structures of the FFBP for stuck pipe while drilling dataset:
(a) using the TANSIG transfer function, (b) using the LOGSIG transfer function.

Figure 7—Different structures of the FFBP for stuck pipe while tripping in dataset:
(a) using the TANSIG transfer function, (b) using the LOGSIG transfer function.

Figure 8—Different structures of the FFBP for stuck pipe while tripping out dataset:
(a) using the TANSIG transfer function, (b) using the LOGSIG transfer function.

Support Vector Machine (SVM)


Support Vector Machine is a supervised learning method with an associated learning algorithm that analyzes
data and recognizes patterns of input/output data (Vapnik, 1995). It is a tool used for classification and
regression tasks. The SVM technique builds the input prototypes in a space with greater dimensions by
employing a nonlinear mapping method. The SVM finds specific linear models between two different
classes and orientates them in a way that maximizes the margin of separator hyperplane. The nearest training
data points, used to define the margin, are called support vectors (Mollajan et al., 2013). In the case of
linearly separable data, for a given training data set [xi, yi] where xi є Rn and yi ∈{-1,1}, i =1,…,n, the SVM
finds two hyper separator planes with maximum distance from marginal points in a way that they separate
the data with equations described as follow (Fig. 9) (Akande et al., 2015):
(4)
(5)
10 SPE-197396-MS

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Figure 9—Hyperplanes and the classified datasets.

These can be presented as:


(6)
for all 1 ≤ i ≤ n
Moreover, the distance between the two hyperplanes is 2/||w||. The task is to minimize w to maximize
the margin in order to get the optimal classifier. Based on the above, we can get the optimization problem
(Helaleh and Alizadeh, 2016):
(7)
Subject to (for any i=1,…, n)
(8)
This is a quadratic programming problem. Using Lagrange multipliers method in the dual space and
solving a set of linear equations, we can get the function for classification as follows:

(9)

Here αn is Lagrange multiplier for solving the optimization problem. For regression, we just use a function
to approach the result instead of classification, which is defined as follows:

(10)

In short, in the case of nonlinear separated training data, to maximize the margin of hyperplanes, the
kernel trick can be applied and the resulting solution has the form:

(11)

where K (xn,x) is kernel transformation which maps nonlinear data into a higher-dimensional space and it can
supersede a dot product wherever it is required. Now there are several kinds of kernel function frequently
used including the linear kernel, polynomial kernel, exponential basis function (RBF) kernel, and sigmoid
function, etc. The popular ones are polynomial and Gaussian kernel (Table 7) (Adib et al., 2013). Optimal
settings for SVM is done by the user, carefully selecting regularization factor c, the type of kernel function
and its specific parameters.
SPE-197396-MS 11

Table 7—Different type of kernel functions.

Type Equation

Polynomial

Gaussian

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


In this study, the SVM model was developed primarily using Gaussian and polynomial kernel functions,
which are two of the most efficient. The SVM user-controlled parameters were optimized by trial-and-error
for the entire available dataset. These parameters include the regularization factor (c), the variance of the
Gaussian function (σ), and the degree of the polynomial function (d), which are varied to tune the SVM
to acquire the best possible generalization performance (Figs. 10, 11, and 12). The optimal combination of
parameters giving the highest accuracy were chosen. The results revealed that the SVM with the Gaussian
kernel function reached the highest accuracy values of 0.96, with c equal 1, and σ equal 0.5, for stuck pipe
while drilling; 0.94, with c equal 10, and σ equal 0.75, for stuck pipe during trip in; and 0.93, with c equal
1, and σ equal 1, for stuck pipe during trip out (Figs. 10a, 11a, and 12a, respectively). In contrast, the best
performance of the accuracy values for the polynomial kernel function were achieved at 0.94, with c equal 1,
and d equal 4, for stuck pipe while drilling; 0.91 with c equal 10, and d equal 3, for stuck pipe during trip in;
and 0.86, with c equal 50, and d equal 5, for stuck pipe during trip out (Figs. 10b, 11b, and 12b, respectively).

Figure 10—Different structures of the SVM for stuck pipe while drilling dataset:
(a) using the Gaussian kernel function, (b) using the polynomial kernel function.

Figure 11—Different structures of the SVM for stuck pipe while tripping in dataset:
(a) using the Gaussian kernel function, (b) using the polynomial kernel function.
12 SPE-197396-MS

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Figure 12—Different structures of the SVM for stuck pipe while tripping out dataset:
(a) using the Gaussian kernel function, (b) using the polynomial kernel function.

To show the robustness for all the developed ANNs and SVM models, the comparison of the performance
criteria including the accuracy, sensitivity, and specificity were calculated for all the presented datasets. As
seen in Table 8, the predicted performance for the SVM with the Gaussian kernel function yielded the best
efficiency.

Table 8—Comparison of designed models.

Dataset Model Accuracy (%) Sensitivity (%) Specificity (%)

Stuck Pipe While Drilling FFBP with TANSIG transfer function* 0.83 0.85 0.81
FFBP with LOGSIG transfer function 0.81 0.84 0.78
SVM with Gaussian kernel function** 0.96 0.97 0.95
SVM with polynomial kernel function 0.94 0.96 0.92
Stuck Pipe During Trip In FFBP with TANSIG transfer function* 0.87 0.89 0.85
FFBP with LOGSIG transfer function 0.83 0.85 0.81
SVM with Gaussian kernel function** 0.94 0.95 0.93
SVM with polynomial kernel function 0.91 0.94 0.88
Stuck Pipe During Trip Out FFBP with TANSIG transfer function* 0.87 0.89 0.85
FFBP with LOGSIG transfer function 0.85 0.88 0.82
SVM with Gaussian kernel function** 0.93 0.94 0.92
SVM with polynomial kernel function 0.86 0.90 0.82

*Best result for ANNs.


**Best result for SVM.

Conclusions
An accurate and early prediction of stuck pipe has been of great importance in order to avoid the risks
associated with this problem's occurrence. Substantial technical and economic benefits can be accomplished
if the correct and identification of the most likely cause of stuck pipe is considered. This will lead to
optimization and cost reduction because the prevention of stuck pipe is far more economical than even the
best of free procedures. The results of this study lead to the following conclusions:

• The primary operational modes (drill string status) affects pipe sticking occurrence. Therefore,
the training datasets should be divided into three categories: stuck pipe while drilling, stuck pipe
during tripping in, and stuck pipe while tripping out.
• An appropriate selection of the learning datasets yields realistic results.

• Wellbore inclination, rate of penetration, lithology type, and BHA length have a more significant
effect on stuck pipe occurrence.
SPE-197396-MS 13

• The results showed that SVM are more accurate in stuck pipe prediction than ANNs. Besides, it
can be found that SVM are more convenient than ANNs since they need fewer parameters to be
optimized.
• With the available real-time data feed of various operational parameters, predicting potential stuck
condition can be very effective during drilling to reduce the well cost.
• In addition, the present analysis allows the drilling engineer to assess the risk of stuck pipe
occurrence during the well planning procedure.

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Acknowledgments
The authors would like to gratefully acknowledge Basra Oil Company (BOC), Iraqi Drilling Company
(IDC), and Missan Oil Company (MOC) in Iraq for providing technical data and their permission to publish
the results.

References
Aadnøy, B. S., Larsen, K., and Berg, P. C., 2003. Analysis of stuck pipe in deviated boreholes. J. Pet. Sci. Eng. 37 (3–4),
pp.195–212. http://dx.doi.org/10.1016/s0920-4105(02)00353-4.
Abbas, A. K., Alameedy, U., Alsaba, M., and Rushdi, S., 2018a. Wellbore Trajectory Optimization Using Rate of
Penetration and Wellbore Stability Analysis. Presented at the SPE International Heavy Oil Conference and Exhibition,
Kuwait City, Kuwait, December 10-12, Paper No. SPE-193755-MS. http://dx.doi.org/10.2118/193755-ms.
Abbas, A. K., Alameedy, U., Alsaba, M., and Rushdi, S., 2018b. Wellbore Trajectory Optimization Using Rate of
Penetration and Wellbore Stability Analysis. Presented at the SPE International Heavy Oil Conference and Exhibition,
Kuwait City, Kuwait, December 10-12, Paper No. SPE-193755-MS. http://dx.doi.org/10.2118/193755-ms.
Abbas, A. K., Bashikh, A. A., Abbas, H., and Mohammed, H. Q. 2019a. Intelligent Decisions to Stop or Mitigate Lost
Circulation Based on Machine Learning. Energy 183: 1104 – 1113. https://doi.org/10.1016/j.energy.2019.07.020.
Abbas, A. K., Al-Haideri, N. A., and Bashikh, A. A. 2019b. Implementing Artificial Neural Networks and Support Vector
Machines to Predict Lost Circulation. Egyptian Journal of Petroleum. https://doi.org/10.1016/j.ejpe.2019.06.006.
Abbas, A. K., Rushdi, S., Alsaba, M., and Al Dushaishi, M. F. 2019c. Drilling Rate of Penetration Prediction of
High-Angled Wells Using Artificial Neural Networks. ASME J. Energy Resour. Technol. 141(11), 112904. https://
doi.org/10.1115/1.4043699.
Adib, H., Sharifi, F., Mehranbod, N., Kazerooni, N. M. and Koolivand, M. 2013. Support Vector Machine Based
Modeling of an Industrial Natural Gas Sweetening Plant. J. Nal. Gas Sci. Eng. 14: 121–131. http://dx.doi.org./10.1016/
j.jngse.2013.06.004.
Akande, K. O., Owolabi, T. O. and Olatunji, S. O. 2015. Investigating the Effect of Correlation-Based Feature Selection
on the Performance of Support Vector Machines in Reservoir Characterization. J. Nal. Gas Sci. Eng. 22: 515–522.
http://dx.doi.org./10.1016/j.jngse.2015.01.007.
Alkamil, E. H., Abbas, A. K., Flori, R., Silva, L. E., Wunsch, D. C., and Chumkratoke, C., 2018. Learning from experience:
real-time h2s monitoring system using fuzzy art unsupervised learning. Presented at the IADC/SPE Asia Pacific
Drilling Technology Conference and Exhibition, Bangkok, Thailand, August 27-29, Paper No. SPE-191097-MS.
http://dx.doi.org/10.2118/191097-ms.
Almubarak, H. A., Stanley, R. J., Long, R., Antani, S., Thoma, G., Zuna, R., and Frazier, S. R., 2017. Convolutional Neural
Network Based Localized Classification of Uterine Cervical Cancer Digital Histology Images. Procedia Computer
Science, 114, 281–287. https://doi.org/10.1016/j.procs.2017.09.044.
Amer, M. M., Dahab, A. S., and El-Sayed, A. H., 2017. An ROP Predictive Model in Nile Delta Area Using Artificial
Neural Networks. Presented at the SPE Kingdom of Saudi Arabia Annual Technical Symposium and Exhibition,
Dammam, Saudi Arabia, 24-27 April, Paper No. SPE-18796-MS. http://dx.doi.org/10.2118/187969-ms.
Anemangely, M., Ramezanzadeh, A., Amiri, H., and Hoseinpour, S., 2019. Machine learning technique for the prediction
of shear wave velocity using petrophysical logs. J. Pet. Sci. Eng. 174, pp.306–327. http://dx.doi.org./10.1016/
j.petrol.2018.11.032.
Bahrami, P., Kazemi, P., Mahdavi, S., and Ghobadi, H., 2016. A novel approach for modeling and optimization of
surfactant/polymer flooding based on Genetic Programming evolutionary algorithm. Fuel, 179, 289–298. http://
dx.doi.org./10.1016/j.fuel.2016.03.095.
Bradley, W., Jarman, D., Plott, R., Wood, R., Schofield, T., Auflick, R., and Cocking, D., 1991. A task force approach to
reducing stuck pipe costs. Presented at the SPE/IADC Drilling Conference, Amsterdam, Netherlands, March 11–14,
Paper No. SPE-21999-MS. http://dx.doi.org./10.2118/21999-ms.
14 SPE-197396-MS

Brandon, N., Panesar, S., Bonanos, N., Fogarty, P., and Mahmood, M., 1993. The effect of cathodic currents
on friction and stuck pipe release in aqueous drilling muds. J. Pet. Sci. Eng. 10 (2), pp.75–82. http://
dx.doi.org./10.1016/0920-4105(93)90032-a.
Cawley, G. C. and Talbot, N. L., 2003. Efficient Leave-One-Out Cross-Validation of Kernel Fisher Discriminant
Classifiers. Pattern Recognit. 36 (11): pp.2585–2592. http://dx.doi.org./10.1016/s0031-3203(03)00136-5.
Eskandarian, S., Bahrami, P., and Kazemi, P., 2017. A comprehensive data mining approach to estimate the rate of
penetration: Application of neural network, rule based models and feature ranking. J. Pet. Sci. Eng. 156, pp.605–615.
https://doi.org/10.1016/j.petrol.2017.06.039.

Downloaded from http://onepetro.org/SPEADIP/proceedings-pdf/19ADIP/2-19ADIP/D022S175R001/1122174/spe-197396-ms.pdf/1 by U. of Alberta Library user on 15 September 2023


Helaleh, A., and Alizadeh, M., 2016. Performance prediction model of miscible surfactant-CO2 displacement in porous
media using support vector machine regression with parameters selected by ant colony optimization. J. Nal. Gas Sci.
Eng., 30, 388–404. http://dx.doi.org./10.1016/j.jngse.2016.02.019.
Jahanbakhshi, R., and Keshavarzi, R., 2016. Intelligent classifier approach for prediction and sensitivity analysis
of differential pipe sticking: a comparative study. ASME J. Energy Resour. Technol. 138 (5), p. 052904. http://
dx.doi.org./10.1115/1.4032831.
Kazemi, P., Khalid, M. H., Szlek, J., Mirtič, A., Reynolds, G. K., Jachowicz, R., and Mendyk, A., 2016. Computational
intelligence modeling of granule size distribution for oscillating milling. Powder Technol. 301, pp.1252–1258. https://
doi.org/10.1016/j.powtec.2016.07.046.
Khamehchi, E., Rahimzadeh Kivi, I., and Akbari, M., 2014. A novel approach to sand production prediction using artificial
intelligence. J. Pet. Sci. Eng., 123, 147–154. http://dx.doi.org./10.1016/j.petrol.2014.07.033.
Krishan, P., Mangla, V., Joshi, N., Tewari, P., and Bose, U., 2000. A new approach solves differential sticking problems
in Geleki and Ahmedabad fields. Presented at the SPE Annual Technical Conference and Exhibition, Dallas, Texas,
October 1–4, Paper No. SPE-63055-MS. http://dx.doi.org/10.2523/63055-ms.
Mohammed, H. Q., Abbas, A. K., and Dahm, H. H., 2018. Wellbore instability analysis for Nahr Umr Formation in
southern Iraq. Presented at the 52nd US Rock Mechanics/Geomechanics Symposium (ARMA), Seattle, Washington,
June 17–20, Paper No. ARMA 18–916.
Mollajan, A., Memarian, H., and Jalali, M. R., 2013. Prediction of reservoir water saturation using support vector
regression in an Iranian carbonate reservoir. Presented at the 47th US Rock Mechanics/Geomechanics Symposium,
San Francisco.
Reid, P., Meeten, G., Way, P., Clark, P., Chambers, B., Gilmour, A., Sanders, M., 2000. Differential-sticking mechanisms
and a simple wellsite test for monitoring and optimizing drilling mud properties. SPE Drill. & Compl. 15(02),
pp.97–104. http://dx.doi.org/10.2118/64114-pa.
Rostami, H., and Khaksar Manshad, A., 2014. A New Support Vector Machine and Artificial Neural Networks for
Prediction of Stuck Pipe in Drilling of Oil Fields. ASME J. Energy Resour. Technol. 136(2), p. 024502. http://
dx.doi.org/10.1115/1.4026917.
Salehi, S., Hareland, G., Dehkordi, K. K., Ganji, M., and Abdollahi, M., 2009. Casing collapse risk assessment and depth
prediction with a neural network system approach. J. Pet. Sci. Eng. 69 (1-2), pp.156–162. http://dx.doi.org/10.1016/
j.petrol.2009.08.011.
Santos, H., 2000. Differentially stuck pipe: early diagnostic and solution. Presented at the IADC/SPE Drilling Conference,
New Orleans, Louisiana, February 23–25, Paper No. SPE-59127-MS. http://dx.doi.org/10.2118/59127-MS.
Shadizadeh, S. R., Karimi, F., and Zoveidavianpoor, M., 2010. Drilling stuck pipe prediction in Iranian oil
fields: an artificial neural network approach. Iran J. Chem. Eng. 7 (4), pp.29–41. http://www.ijche.com/
article_10326_1700.html.
Shi, X., Liu, G., Gong, X., Zhang, J., Wang, J., and Zhang, H., 2016. An Efficient Approach for Real-Time
Prediction of Rate of Penetration in Offshore Drilling. Mathematical Problems in Engineering, 1–13. http://
dx.doi.org/10.1155/2016/3575380.
Shivers, R. and Domangue, R., 1993. Operational decision making for stuck pipe incidents in the Gulf of Mexico: A risk
economics approach. SPE Drill & Compl. 8(02), pp.125–130. http://dx.doi.org/10.2118/21998-pa.
Siruvuri, C., Nagarakanti, S., and Samuel, R., 2006. Stuck pipe prediction and avoidance: a convolution neural network
approach. Presented at the IADC/SPE Drilling Conference, Miami, Florida, February 21–23, Paper No. SPE-98378-
MS. https://doi.org/10.2118/98378-MS.
Szlek, J., and Mendyk, A., 2018. Fscaret: Automated feature selection from caret. https://cran.rproject.org/web/packages/
fscaret/index.html (accessed 3 November 2018).
Vapnik, V. 1995. The Nature of Statistical Learning Theory. New York: Springer-Verlag.

You might also like