SPE-194980-MS

AccuPipePred: A Framework for the Accurate and Early Detection of Stuck

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

Pipe for Real-Time Drilling Operations

Arturo Magana-Mora, Salem Gharbi, Abrar Alshaikh, and Abdullah Al-Yami, Drilling Technology Team, EXPEC
ARC, Saudi Aramco

Copyright 2019, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Middle East Oil and Gas Show and Conference held in Manama, Bahrain, 18-21 March 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
Thorough preplanning and best drilling practices are effective in reducing stuck pipe incidents, data analytics
offer additional insight into further reducing the significant non-productive time (NTP) that results from
this unplanned event. The severity of the stuck pipe problem may stop the drilling operations for a short
time, or in more difficult cases, the drill string has to be cut and the borehole is sidetracked or plugged and
abandoned. Consequently, detecting the early signs of this problem, in order to take the right actions, may
considerably or entirely reduce the risk of a stuck pipe.
Although computational models have been proposed for the early detection of the stuck pipe incidents,
the models are derived from a reduced set of wells with stuck pipe incidents, which may result in under-
trained models that predict a large number of false positive alarms. A sufficient amount of data or wells
that statistically represent the parameters surrounding stuck pipe incidents under different circumstances
is required in order to derive a generalizable and accurate prediction model. For this, we first derived a
framework to automatically and systematically extract relevant data from the historical data. As such, our
framework searches through the historical data and localizes the surface drilling and rheology parameters
surrounding the stuck pipe incidents. Moreover, we performed feature selection by selecting the top-ranked
parameters from the analysis of variance, which measures the capability of the drilling and rheology
parameters to discriminate between stuck pipe incidents and normal drilling conditions, such as, weight on
bit, revolutions per minute, among others.
Using the relevant features selected by the analysis of variance, we derived a robust and fast classification
model based on random forests that is able to accurately detect stuck pipe incidents. The implemented
framework, which includes the automated data extraction module, the analysis of variance for feature
selection, and prediction, is designed to be implemented in the real-time drilling portal as an aid to the
drilling engineers and the rig crew in order to minimize or avoid the NTP due to a stuck pipe.

Introduction
Stuck pipe incidents occur routinely during drilling operations, and are mainly due to one of three
mechanisms, which are: differential sticking, packoff, and wellbore geometry. Sometimes, stuck pipe
2 SPE-194980-MS

incidents occur due to a combined effect of differential pressure and mechanical restrictions. In general,
however, mechanical sticking, due to either packoff or wellbore geometry, accounts for more that 70% of
stuck pipe incidents according to a 2012 study by Muqeem et al (Muqeem et al. 2012).
Despite all the effort to optimize drilling practices to avoid stuck incidents or to free the string with the
minimum lost time (LT) possible, stuck pipe incidents remain as one of the major causes of non-productive
time during drilling operations. Several studies have been published since 1987 about the use of discriminant
analysis, artificial neural networks and other techniques in order to build stuck pipe classifiers, on other

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

words, autonomous tools for identifying the mechanism of stuck pipe incidents after the fact. Other studies
published in 2013 and 2015 propose another approach which combines the automation aspect of the machine
learning techniques with the human knowledge in mitigating drilling problems. These models use pattern
recognition for risk assessment incorporated in an automated decision support tool to propose the mitigation
for the identified drilling trouble based upon a case-based reasoning in real-time (Biegler and Kuhn 1994,
Howard and Glover 1994, Miri et al. 2007, Murillo et al. 2009, Shoraka et al. 2011, Jahanbakhshi et al. 2012,
Chamkalani et al. 2013, Naraghi et al. 2013, Sadlier et al. 2013, Ferreira et al. 2015, Salminen et al. 2017).
Most of the computational models for the prediction of stuck pipe incidents published in the literature,
were derived from a reduced set of wells with stuck pipes, which may result in models with poor
generalization capabilities for classifying events in unseen wells. This is because machine-learning models
assume that the data used for model development are representative of the population and expect the new
data to have a similar distribution. Therefore, a sufficient amount of data or wells that statistically represent
the parameters surrounding stuck pipe incidents under different circumstances are required in order to derive
generalizable and accurate prediction models.
Moreover, it is also important to properly formulate the classification problem so that the model is capable
of learning the trends of the drilling parameters that lead to a stuck pipe incident. For instance, if we consider
the surface drilling parameters at a given time step and assign a label (stuck or normal operation), the model
would assume that samples are independent of each other and would not learn the trends or changes of the
drilling parameters that characterize a stuck pipe. Consequently, we formulated the problem as a time series
where the drilling parameters of a set of time steps surrounding a stuck incident correspond to a sample.
In this study, we present the process of data extraction, normalization, and time series formulation.
Moreover, we ranked the surface drilling and rheology parameters based on the F-value from the analysis
of variance (ANOVA) to show the most relevant parameters to detect stuck incidents. The resulting set of
relevant parameters was then used to derive a random forest model for the prediction of stuck incidents.

Methods
Datasets
In this study, we identified a statistically representative number of stuck pipe incidents and extracted the
surface drilling parameters, rheology parameters, and the inclination and dogleg severity from the survey
data. Table 1 shows the list of the selected parameters and their units. From the drilling parameters, we
considered WOB, flow-in rate, and RPM as these are the variables controlled by the driller and which
have a significant effect on the drilling performance. We have considered HKL, SPP, ROP, and torque to
capture the drilling performance. Concerning the drilling fluid, we considered the fluid density and viscosity
as they are the determinant factors in maintaining the pressure overbalance between the wellbore and the
formation and for maintaining proper hole cleaning throughout the operation, which can lead to stuck pipe
incidents. Finally, we have also included the inclination and dogleg severity from the survey data as these
are indicatives of the well profile complexity and wellbore misconfiguration, which are critical in predicting
mechanical stuck pipe incidents, mainly due to wellbore geometry, i.e. a conflict between the wellbore shape
and the BHA configuration (Mitchell 2002).
SPE-194980-MS 3

Table 1—Selected parameters.

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

For each of the identified wells, we extracted the considered parameters for the time steps occurring 30
minutes before the stuck pipe incidents. As such, each stuck incident is defined by a 2-dimensional (2D)
matrix where rows represent the time steps and the columns the surface drilling parameters and survey data.
In order to capture the normal drilling operations, we extracted the parameters for three 30-minute intervals
not overlapping the stuck segment for each of the considered wells. Fig. 1a shows the process for extracting
stuck incidents and normal operations for a well. Real-time surface drilling parameters and real-time survey
data are referenced to time and are recorded with a frequency of ~0.2 hertz (every five seconds). This
results in approximately 360 time steps in the 30-minute window surrounding stuck incidents and normal
operations. The large number of drilling parameters in these ~360 time steps (~ 360 × 11 ≈ 3,960 parameters)
considerably increases the complexity of a machine learning model. Therefore, we averaged the real-time
drilling parameters for every 30 seconds to reduce the number of parameters, which also smoothens and
reduces the measurements outliers due to possible transmission or sensor errors. As a result, stuck incidents
and normal operations are defined by a 60 × 11 matrix where rows indicate the averaged time steps and
columns contains the drilling parameters and survey data. Finally, we converted the 2D matrix containing
the parameters referenced to time into a vector to which we added the rheology parameters. Fig. 1b shows
the final encoding of each stuck/normal operation sample into a vector of 663 numerical values.
4 SPE-194980-MS

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

Figure 1—a) Illustration of the extraction of stuck pipe incidents and two normal operations from real-time drilling
parameters. b) Encoding of the drilling parameters, rheology parameters, and survey data surrounding an event into a vector.

Data Normalization and Feature Selection

The ranges and means of the parameters defining stuck incidents or normal operations may considerably
differ from well to well. This is because different parameters are required depending on the well profiles,
formations drilled, rig capacities, among others. These differences often hamper the learning of most of
the machine-learning algorithms and tend to assign a higher priority to the parameters with higher values.
Therefore, normalization or standardization of the parameters is an important step to give all parameters an
equal weight and to speed up the learning of the models. In this study, parameter values were normalized
by using the z-score defined as follows

where μ and σ refer to the mean and standard deviation for the i-th parameter from Table 1 and x is the
parameter value that will be normalized. The z-score converts the original inputs into unitless variables and
indicates the signed number of standard deviations by which the value of an observation is above (positive
number) or below (negative number) the mean.
SPE-194980-MS 5

We computed the ANOVA F-value using Scikit-learn in Python (Pedregosa et al. 2011) for each of the
663 considered parameters to measure the significance of each of these parameters to discriminate between
stuck pipe incidents and normal operations. The F-value is the ratio of the variances from the classes and
assigns higher scores to relevant parameters. It is important to note that the F-value measures the relevance
of each parameter alone and does not consider combinations of parameters. However, it allows us to identify
the relevant parameters for detecting stuck pipe incidents and to remove those that do not contribute to the
model. In this study, we show the results obtained by deriving a model with the entire parameter set (663

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

parameters) and models derived by using 66% of the original parameter set (437 parameters).

Machine-learning Model and Cross-validation

The performance of the classification models derived by different machine-learning algorithms greatly
depends on the complexity and size of the classification problem defined by the data. We can arguably
conclude that simple classification problems are better solved by simpler and faster linear models, while
more complex problems are better fitted by more sophisticated non-linear models (Magana-Mora and Bajic
2017). For this, we tested the performance of different models, i.e., random forests (RF), decision trees
(DT), artificial neural networks (ANN), and support vector machines (SVM) from the Scikit-learn machine-
learning package (Pedregosa et al. 2011) in Python. The justification for the selection of such models
was done based on the extensive empirical study that showed that RF and ANN produced the best results
on average over the UCI machine-learning repository (Fernandez-Delgado et al. 2014), while DT was
considered due to their simplicity and fast training.
We used the k-fold cross-validation technique to assess the performance of the considered machine-
learning algorithms. The k-fold cross-validation consists in partitioning the original data defining the
classification problem into k (approximately) equal-sized subsets. For each of the k folds, k-1 subsets are
used to train the classification model while the remaining subset is used for testing the model. The final
result is the average of the results obtained from the k folds on the test subset which were not considered
during the training of the model. We selected k = 10 as the number of samples may be considered to be
relatively small (944 samples).
Moreover, since each of the considered classification models have different parameters that can be tuned,
we used nested 5-fold cross-validation to assess the performance of the different parameter combination
obtained from only the training data. The use of outer cross-validation to determine the model performance
and inner cross-validation for model parameter tuning based on the performance from the validation set
is referred to as nested cross-validations. It is important to notice that the test subset of the data (used to
compute the final performance) is never available to the model during the parameter tuning in the inner
cross-validation, therefore, biases are not introduced. Fig. 2 shows the considered nested cross-validation
setup.
6 SPE-194980-MS

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

Figure 2—Nested cross-validation for model validation and tuning of model's parameters.

Statistical Measures for Model Performance Evaluation

We considered different statistical measures to evaluate the performance of the models derived by the
considered classification algorithms, namely, accuracy, precision, sensitivity, and specificity, defined in
Table 2. The accuracy measure indicates the capability of the model to accurately identify stuck incidents
and normal operations. Nevertheless, accuracy may not be representative when the number of samples from
both conditions (stuck incidents/normal operations) is uneven or unbalanced. For instance, in a hypothetical
classification problem with 1,000 samples from class 1 and 100 samples from class 2, a model that predicts
all samples as class 1 would achieve a 90% accuracy. To account for this, other statistical measures are often
used to represent different capabilities of the model. Among these are sensitivity and specificity measures,
which separately evaluate the model capability to accurately detect stuck incidents and normal operations,
respectively. In the hypothetical classification problem with 1,100 samples, the model would achieve a
sensitivity of 100% and specificity of 0%. Moreover, we considered precision, which specifies the number
of stuck incidents correctly identified by the model and the false positives (false alarms), this is of particular
importance for applications where it is important to reduce false positives as much as possible even at the
expense of increasing the false negative rate. Lastly, we show the results in terms of the receiver operating
characteristics (ROC) curve, a plot which shows the true positive rate on the Y-axis and the false positive
rates on the X-axis that enables easily visualize the tradeoff of the model between true and false positive
rates.
SPE-194980-MS 7

Table 2—Statistical measures used to evaluate the performance of machine-learning models. TP and TN represent the number
of true positive predictions (true stuck incidents), true negative predictions (true normal operations), respectively. FP and FN
refer to the incorrectly predicted samples by the model, i.e., false positive predictions (normal operations incorrectly predicted
as stuck incidents), and false negative predictions (stuck incidents incorrectly predicted as normal operations), respectively.

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

Results and Discussion
The main contribution of this study is the analysis of real-time surface drilling parameters, rheology
parameters, and real-time survey data to predict stuck pipe incidents from a statistically representative
number of wells. The large-scale analysis of the large number of wells is critical for avoiding model
undertraining or lack of model generalization, i.e., ability to perform as expected once the model is deployed
for predicting conditions in unseen wells. Moreover, we formulated the classification problem as a time
series problem in order to consider parameter dependencies and trends to accurately discriminate between
normal operations and conditions that may lead to stuck pipe incidents. Finally, we performed feature
selection to reduce the number of parameters used to define the stuck pipe incidents and normal operations
and tested the performance of models derived by using different machine learning algorithms.
Occam's razor principle states that simpler models/solutions are more likely to be correct than more
complex (Sober 1994) and may result in more generalizable models. Therefore, feature selection is an
essential process when deriving machine-learning models in order to remove the input parameters that
are irrelevant to discriminate between the classes (stuck pipe incidents and normal operations) and that
unnecessarily complicate the classification model. For this, we used the ANOVA F-value to determine
the relevance of each of the 663 parameters for detecting stuck pipe incidents and filter those that do not
contribute to the performance of the model. Table 3 shows the considered parameters ranked based on the
averaged F-value.
From Table 3 we see that RPM and SPP the top ranked parameters for detecting stuck pipe incidents.
Clearly, RPM is an indicative of stuck pipe as the drill bit is unable to rotate in the cases of packoff and
undergauge sticking. The SPP parameter is the second most relevant and is an indicative of fluid circulation.
This is because a circulation restriction, and thus SPP abnormality, is a critical characteristic of packoff.
Conversely, the rheology parameters where the least contributing and this may be due to the fact that
these parameters are not obtained in real-time, therefore, these values are not related to the other real-time
parameters.
8 SPE-194980-MS

Table 3—Feature ranking based on the ANOVA F-value. The table shows the considered
parameters based on the averaged F-value for all time steps to avoid listing the 663 parameters.

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

In order to select a robust machine-learning model that best suits our problem, we derived models by
using RF, ANN, DT, and SVM. Moreover, we compared their performance when using all 663 parameters
or a subset based on the feature selection process. Table 4 shows the performance metrics using 10-
fold cross-validation (see Methods). In accordance to the previous studies comparing the performance of
machine-learning models and the effects on the model after feature selection (Fernandez-Delgado et al.
2014, Magana-Mora and Bajic 2017, Alshahrani et al. 2017, Soufan et al. 2015), RF achieved the best
results with an accuracy of 82.75% when using 66% of the original set of parameters. On the other extreme,
the SVM model was unable to learn from the data and predicted everything as being normal operations. A
predictive model for the early detection of stuck pipe incidents is required to reduce the number of false
alarms in order to not interfere unnecessarily with the drilling operation, which may result in NTP. This is
precisely captured by the precision measure where 100% indicates that the model does not predict any false
alarms. However, precision does not specify the proportion of stuck incidents detected by the model. In
other words, a model could achieve a high precision while only detecting few stuck pipe incidents. For this,
have to account for the precision score and the sensitivity at the same time. In our model, a precision score
of 79.32% indicates that out of 100 predicted stuck incidents, ~11 will be false alarms, while a sensitivity
of 60% implies that the model will be able to detect 60 stuck incidents of out 100 real stuck incidents.

Table 4—Results obtained from 10-fold cross-validation. The ‘subset’ column shows the results
of the model derived by considering the top 66% parameters ranked based on the F-value.

From Table 4 we can observe the tradeoff of the model between precision and sensitivity. In other words,
if the model is set to detect most of the stuck pipe incidents (higher sensitivity), the number of false alarms
SPE-194980-MS 9

would also increase (lower precision). For this, we show the results in terms of the ROC curve to better
understand this tradeoff in Fig. 3. In this ROC plot, the top left corner represents an ideal model that is able
to achieve a false positive rate of zero and a true positive rate of one. In our case, however, we can observe
that if we set the model to predict 0 false alarms, then the model would simply predict everything as normal
operations (with a true positive rate of 0%). A middle point would be to set the model to achieve a true
positive rate of 80% while predicting on average 20% false alarms. It is important to note that since the
stuck pipe incidents are obtained from few hundred completely different wells, the results achieved in this

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

study represent the expected results of a deployed model with no biases introduced.

Figure 3—Averaged ROC from an RF using10-fold cross-validation. ‘Chance’

indicates the curve for a hypothetical model with random predictions.

Conclusions and Steps Forward

In this study, we presented a large-scale analysis of a statistically representative number of wells with stuck
pipe incidents. From these wells, we extracted the real-time surface drilling parameters, survey data, and
rheology parameters for 30 minutes before the stuck pipe incident took place. This formulates the problem as
a time-series where we no longer assume that the time steps are independent. Moreover, by considering the
time steps surrounding the stuck incident, the model is able to recognize the trends of the drilling parameters,
e.g., slower RPM, increase in torque, among others. However, some of the parameters captured in the 30-
minute window may be irrelevant for detecting stuck pipe incidents and may unnecessarily complicate
the training of the machine-learning models. To remove these parameters and to derive simpler and faster
models, performed feature selection based on the ranked the parameters from the individual ANOVA F-
value. Finally, we tested the performance of different machine-learning models to determine the best suited
for this particular problem. As a result, RF achieved the best performance with a precision score of ~80%
and sensitivity of ~60 from a 10-fold cross-validation.
Although we only extracted the real-time parameters from 30-minute window before stuck incidents, we
believe that it will be interesting to test the performance of the model by considering the surface real-time
parameters from a 30-minute window obtained 30 minutes, one hour, two hours, etc., before the stuck pipe
10 SPE-194980-MS

incident. Clearly, the farther the model gets from the stuck incident the harder it will be for the model to
detect the stuck pipe incident, as the parameters will resemble normal operations.

Acknowledgment
The authors would like to thank Eric Moellendick, Chief Technologist of the Drilling Technology Team,
for his continuous support for the team's research.

Downloaded from http://onepetro.org/SPEMEOS/proceedings-pdf/19MEOS/3-19MEOS/D032S077R003/1147197/spe-194980-ms.pdf by U. of Alberta Library user on 15 September 2023

References
Alshahrani, Mona, Soufan, Othman, Magana-Mora, Arturo et al 2017. DANNP: an efficient artificial neural network
pruning tool. PeerJ Computer Science 3 (e137).
Biegler, M. W. and Kuhn, G. R. 1994. Advances in Prediction of Stuck Pipe Using Multivariate Statistical Analysis.
Presented at the SPE/IADC Drilling Conference, Dallas, Texas, USA. 15–18 February. SPE-27529-MS. https://
doi.org/10.2118/27529-MS.
Chamkalani, A., Shahri, M. P., and Poordad, S. 2013. Support Vector Machine Model: A New Methodology for Stuck
Pipe Prediction. Presented at the SPE Unconventional Gas Conference and Exhibition, Muscat, Oman. 28–30 January.
SPE-164003-MS. https://doi.org/10.2118/164003-MS.
Fernandez-Delgado, Manuel, Cernadas, Eva, and Barro, Senen. 2014. Do we Need Hundreds of Classifiers to Solve Real
World Classification Problems? Journal of Machine Learning Research 15 (1): 3133–3781.
Ferreira, A. P. L. A., Carvalho, D. J. L., Rodrigues, R. M. et al 2015. Automated Decision Support and Expert Collaboration
Avoid Stuck Pipe and Improve Drilling Operations in Offshore Brazil Subsalt Well. Presented at the Offshore
Technology Conference, Houston, Texas, USA. 4–7 May. OTC-25838. https://doi.org/10.4043/25838-MS.
Howard, J. A. and Glover, S. B. 1994. Tracking Stuck Pipe Probability While Drilling. Presented at the APE/IADC Drilling
Conference, Dallas, Texas, USA. 15–18 February. SPE-27528-MS. https://doi.org/10.2118/27528-MS.
Jahanbakhshi, R., Keshavarzi, R., and Jahanbakhshi, R. 2012. Intelligent Prediction of Wellbore Stability in Oil and Gas
Wells: An Artificial Neural Network Approach. Presented at the 46th US Rock Mechanics/Geomechanics Symposium,
Chicago, Illinois, USA. 24–27 June. ARMA-2012-243.
Magana-Mora, Arturo and Bajic, Vladimir B. 2017. OmniGA: Optimized Omnivariate Decision Trees for Generalizable
Classication Models. Scientific Reports 7 (1).
Miri, R., Sampaio, J. H. B., Afshar, M. et al 2007. Development of Artificial Neural Networks to Predict Differential
Pipe Sticking in Iranian Offshore Oil Fields. Presented at the International Oil Conference and Exhibition in Mexico,
Veracruz, Mexico. 27-30 June. SPE-108500-MS. https://doi.org/10.2118/108500-MS.
Mitchell, John. 2002. Trouble-Free Drilling: Stuck Pipe Prevention, 3rd edition, Vol. 1: Drilbert Engineering.
Muqeem, M. A., Weekse, A. E., and Al-Hajji, A. A.. 2012. Stuck Pipe Best Practices - A Challenging Approach to
Reducing Stuck Pipe Costs. Presented at the SPE Saudi Arabia Section Technical Symposium and Exhibition, Al-
Khobar, Saudi Arabia. 8–11 April. SPE-160845-MS. https://dx.doi.org/10.2118/160845-MS.
Murillo, A., Neuman, J., and Samuel, R. 2009. Pipe Sticking Prediction and Avoidance Using Adaptive Fuzzy Logic
Modeling, Oklahoma City, Oklahoma, USA. 4–8 April. SPE-120128-MS. https://doi.org/10.2118/120128-MS.
Naraghi, M. E., Ezzatyar, P., and Jamshidi, S. 2013. Prediction of Drilling Pipe Sticking by Active Learning Method
(ALM). Journal of Petroleum and Gas Engineering 4 (7): 173–183. https://doi.org/10.5897/JPGE2013.0166.
Pedregosa, Fabian, Varoquaux, Gal, Gramfort, Alexandre et al 2011. Scikit-learn: Machine Learning in Python. J Mach
Learn Res 12: 2825–2830.
Sadlier, A., Says, I., and Hanson, R. 2013. Automated Decision Support to Enhance While-Drilling Decision Making:
Where Does It Fit Within Drilling Automation? Presented at the SPE/IADC Drilling Conference and Exhibiton,
Amesterdam, The Netherlands. 5-7 March. SPE-163430-MS. https://doi.org/10.2118/163430-MS.
Salminen, K., Cheatham, C., Smith, M. et al 2017. Stuck-Pipe Prediction by Use of Automated Real-Time Modeling and
Data Analysis. SPE Drilling & Completion 32 (3): 184–193. https://doi.org/10.2118/178888-PA.
Shoraka, S. A. R., Shadizadeh, S. R., and Shahri, M. P. 2011. Prediction of Stuck Pipe in Iranian South Oil Fields Using
Multivariate Statistical Analysis. Presented at the Nigeria Annual International Conference and Exhibition, Abuja,
Nigeria. 30 July-3 August. SPE-151076-MS. https://doi.org/10.2118/151076-MS.
Sober, Elliott 1994. Let's Razor Occam's Razor. In In Knowles, Dudley Explanation and Its Limits, 73–93. Cambridge
University Press.
Soufan, O., Kleftogiannis, D., Kalnis, P. et al 2015. DWFS: a wrapper feature selection tool based on a parallel genetic
algorithm. PLoS One 10 (2): e0117988. http://www.ncbi.nlm.nih.gov/pubmed/25719748.