You are on page 1of 12

SPE-189284-STU

Identifying the Optimum Zone for Reducing Drill String Vibrations

Etaje Darlington Christian, University of Calgary

Copyright 2017, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE International Student Paper Contest at the SPE Annual Technical Conference and Exhibition held in San Antonio,
Texas, USA, 9-11 October 2017.

This paper was selected for presentation by merit of placement in a regional student paper contest held in the program year preceding the International Student Paper
Contest. Contents of the paper, as presented, have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The
material, as presented, 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
Drilling operators witness significant lost time due to early failure of bottomhole equipment resulting from
vibration and shock. This eventually leads to cumulative losses of millions of dollars for the industry.
Increase in applied weight on bit (WOB) at low angular velocity (RPM) can trigger instabilities leading
to stick slip. Potentially, compression and stretch that occurs along the BHA during drilling operations
could lead to whirling and buckling. The complexity of the whirling motion causes lateral shifts, shocks
and friction against the borehole walls.
The driller has limited options. If stick-slip is identified, the driller decreases weight-on-bit (WOB) but
whirling may occur from increasing revolutions per minute (RPM). Since the overall goal is to optimize
drilling then reducing both WOB and ROB would not be an option since that would results in decrease in
rate of penetration (ROP). This puts the driller in a tough situation where both severe vibrations and low
ROP could occur simultaneously during drilling operations.
There is an optimum zone where drilling parameters – RPM and WOB -- improve BHA/bit stability. A
machine learning methodology is described which is able to (a) identify the zone of stability through the
use of supervised and unsupervised learning and (b) anticipate an upcoming optimum for safe drilling by
merging historic data with real time analysis through the use of online learning. A comparison is presented
which compares supervised and unsupervised machine learning in identifying and updating the optimum
zone. From this zone, a parameter set of permissible combinations of WOB and RPM can be estimated. The
methodology described is then applied to data derived from several hours of drilling in a highly tortuous
zone with persistent vibration problems.

Introduction
To achieve improved drilling efficiency and better productivity of the driller, there is need for real-time
optimization of drilling parameters during drilling operations through each formation in order to optimize
weight on bit and bit rotation speed to increase drilling rate as well as reduce the drilling cost. The driller
only sees the surface data but there is usually a deviation in the downhole drilling parameters. The driller
needs to make better decisions as he manipulates the parameters to improve drilling and deal with various
issues that may arise during drilling.
2 SPE-189284-STU

The parameters collected during drilling are weight on bit (WOB), rotation rate (RPM), pump parameters
(SPM), depth, inclination, azimuth and rate of penetration (ROP). These parameters have a significant
impact on the entire optimization process. The success of drilling optimization is closely related with the
quality of the recorded drilling parameters. However, the driller has to make those important decisions in
real time when drilling problems arise.
Several methods have been used to optimize the drilling parameters. In 1975, Tansev explained how
to improve drilling performance. His method involves the interaction of raw data, regression and an
optimization technique in order to predict ROP and the life of the bit (Tansev 1975). Karlsson et al in
1985, observed the use of a BHA design that included a navigation sub. They noticed that the tool allowed
the driller to always know the direction of the well and make required trajectory changes while drilling
(Karlsson et al 1985). In 1997, Kamata et al explained a drill-bit seismic technique which provides important
subsurface structure information by using acoustic energy radiated during drilling operations. Sensors,
placed at the top of drill string, were used to record the information. They achieved drilling optimization
from the information gathered thereby improving safety records and saving cost (Kamata et al 1997). Paes et
al in 2005, focused on the use of sensors for pressure-while-drilling (PWD) and vibration sensors to reduce
the drilling cost, non-productive time (NPT), and improve drilling effectiveness without adding more cost
to the cost of the routine measurement while drilling cost (Paes et al 2005). Elshafei et al in 2015 determined
the right combination of drilling parameters to reduce drilling time and minimize deviation from planned
drilling path by inputting control commands on angular velocity and torque for quad bit drilling system
(Elshafei et al 2015). In 2017, Torres-Cabrera et al observed the difficulty in predicting BHA behaviour
which leads to low ROP, unnecessary tripping, and occasionally lost pipe in hole. They addressed the issues
through a series of drilling improvements based on real-time and post-well analyses (Torres-Cabrera et al
2017).
Another method that can be applied to optimize drilling parameters is "machine learning." Machine
learning isn't new; it has been around at least since the 1970s, when the first related algorithms appeared.
The general idea behind most machine learning is that a computer learns to perform a task by studying a
training set of examples. The computer (or system of distributed or embedded computers and controllers)
then performs the same task with data it hasn't encountered before (Louridas et al 2016). This method is
primary focus of this paper. Machine learning has been applied to other aspects in the oil industry. Zhang
et al in 1991, applied machine learning to rock mechanics and observed that all of the factors governing the
rock mass behaviors could be considered as input variables to predict the varying rock behaviors. They made
these observations without limiting the amount of input variables that could be used (Zhang et al 1991).
Alvarado et al in 2002 used machine learning in their aim to adapt EOR/IOR technologies to rejuvenate a
large number of the mature fields in Venezuela. They used machine learning algorithms to draw rules for
screening (Alvarado et al 2002). In 2016, Cao et al used machine learning algorithms to predict production
for several wells using pressure and production data, geological maps, and constraints during operations.
They used a well-known machine learning method – Artificial Neural Network (ANN). Without assuming
a prearranged model, ANN learns from large volume of data points and can change based on the flexibility
of the data available (Cao et al 2016). In 2017, Bangert proposed the use of machine learning in order
to conduct smart condition monitoring. He realized that his proposed method was more successful than
standard condition monitoring thus preventing false alarms and always alarming unhealthy states of plants or
equipment (Bangert et al 2017). This paper aims to use machine learning principles to help the driller make
quality decisions in terms of drilling parameter adjustments during vibration problems thereby optimizing
the drilling process.
SPE-189284-STU 3

Vibration Issues Facing the Driller during Drilling Operations


Frequent vibrations of the drill string may lead to poor drilling performance and non-productive time. The
concerns arising from drilling vibration are: wasted energy input, low ROP, lengthy drilling time, spoilt
bit, damage to motor steerable leading to unintended trips, damaged MWD/LWD tools causing lost data,
increased fatigue in the drill string, higher caving due to borehole wall damage, discrepancy in data due
to meddling with downhole tool telemetry during vibrations, increased cost of rig equipment repairs and
increase downtime.
Two kinds of vibration are of significant concern. First is Stick-Slip. In this case, the bit periodically stops
rotating in a torque up moment then spins freely, this goes on through a non-uniform rotation of the drill
string. During stick slip, the downhole RPM is can be 3x to 15x average surface RPM. The consequences of
Stick-slip are PDC bit damage, lower ROP, connection over-torque, back-off and drill string twist-offs. Stick
slip occurrence also leads to wear on bit gauge and stabilizer as well as interruption in mud pulse telemetry.
The second vibration type is drill string whirling. The bulk of drill string whirling happens in the
bottomhole assembly (BHA). During whirling, parts of the BHA face lateral displacements which generate
bending stresses and lateral shocks when the BHA contact the borehole wall (JPT Staff 1998). The act of
moving around the wellbore and not its centreline is the whirling phenomenon. Three types of whirling
can occur; forward whirling is a scenario where the drill string is rotating around the wellbore in the same
direction with its rotation around its own centreline; backward whirling is a situation where the drill string
is rotating around the wellbore in a direction opposite the direction of its rotation around its own centreline.
Chaotic whirling occurs where the bits moves in a zig-zag manner with no consistent direction. Whirling
creates an over gauge hole reinforcing the tendency for the bit and BHA to whirl.
The driller has to constantly manipulate available parameters to mitigate vibration problems. A driller’s
dilemma emerges when increasing (WOB) induces stick-slip whereas increasing (RPM) induces whirl.
Keeping both WOB and RPM low reduces vibration levels but it negatively affects ROP. As a result, the
drilling operation either suffers low ROP or experiences higher ROP but with severe vibrations (Wu et al
2010).

The Concept of Machine Learning


Machine learning gives computers the ability to optimize performance criterion based on sample data or past
knowledge (Ethem 2009). The goal of machine learning is to identify and reveal potential veiled patterns
linked with the data being analysed. The world today is circled with applications of machine learning. A
perfect example is the use of Google search which learns to display the best results. Another example is the
anti-spam software which filters email messages (Ben-david 2014).
As seen in Figure 1, there are two major types of machine learning. First is supervised (predictive)
learning where the aim is to learn a pattern from inputs x to outputs y, given a labeled set of input-output
pairs.R = {(Xi, Yi)}iN = 1. R is called the training set, and N is the number of training examples (Murphy
2012).
4 SPE-189284-STU

Figure 1—Machine Learning Methods (Louridas 2016)

Unsupervised (descriptive) learning is the second major type of machine learning. In this case, R = {Xi}iN
= 1, and the aim is to search the data for unique patterns. Unlike predictive learning, where prediction of
y for a given x can be compared to the observed value, unsupervised is not really defined since there are
no clear error metric to use (Murphy 2012).

How Machine Learning is utilized for Vibration Problems


WOB and RPM causing whirling and stick slip can be predetermined if the total drilling conditions are
known (Wu et al 2010). A boundary condition for stable drilling can be obtained in a plot with WOB on the Y
axis and RPM on the X axis (Figure 2). This means if the driller keeps his drilling parameters to keep the bit
in the optimum zone, then drilling will be stable depending on the bit and mechanical properties of the rock.

Figure 2—Optimum Zone Chart (Wu et al 2010)

The boundaries of the optimum zone helps determine the best combination of WOB and RPM for
optimum ROP. The hard question to answer is if the stick slip and whirling zone is predicted accurately.
SPE-189284-STU 5

Thus the need for machine learning. In order to identify the optimum zone effectively, a process, as shown
in Figure 3, is adopted to ensure all drilling parameters have an impact on the optimum zone. The process
uses available real-time data from the drilling rig, performs a variable transformation and reduction, and
then utilizes machine learning algorithms to identify the optimum zones. The zones are then transformed
back into the original variables and displayed to the driller. Principal Component Analysis (PCA) is used
to achieve the variable transformation and reduction. K-means clustering and decision tree classification
are then used to identify the optimum zones; after which the boundaries are represented based on WOB
and RPM.

Figure 3—Using Machine Learning to Identify Drilling Optimum Zones

Compressing Data Using PCA


PCA can be used for identifying hidden patterns in high dimension data; it is commonly applied in image
compression. PCA is used to find differences and similarities in data. Part of the PCA advantage is dimension
reduction without losing vital information. PCA is an orthogonal transformation of a correlated data matrix
of n objects by p variables into uncorrelated axes called principal components that are linear combinations
of the original p variables. Consider the data to be matrix M where the transpose of the matrix is MT. A scalar
λ is called an eigenvalue of the n × p matrix M. If Mx = λx, then x is called an eigenvector corresponding
to the eigenvalue λ.
The PCA procedure involves obtaining the eigenvectors for either MMT or MTM. Applying this to the
initial data, the principal eigenvector axis is the one with its points well spread which is where the data has
its variance maximized. These points lie along this axis with very little deviation away from the axis. The
second axis is the eigenvector with the second highest eigenvalue in which the variance of distances from
the first axis is greatest, and so on (Leskovec et al 2014).
In other words, PCA is a technique which generates new features which are linear combination of the
original feature. For instance, given a dataset in a "d dimensional space" is transformed to a "z dimensional
subspace" where z < d. The new dimensions in z are called principal components and they are directed in
the path of maximum variance without repeating variance in preceding components. This means the first
component has the highest variance while each component that follows it have a corresponding smaller
variance value (Vasan et al 2016). The Principal Components can be represented as the following
pci = a1X1 + a2X2+…+adXd
pci — Principal Component ‘i’; Xj — original feature ‘j’; aj — numerical coefficient for Xj.
6 SPE-189284-STU

Clustering and Classification


Clustering is a process forming groups whose objects are somewhat similar. A cluster is grouping of objects
which are alike and different from objects in other clusters. K-means clustering is type used in this study.
Widely used in data mining, K-means algorithm is a type of clustering analysis based on partitioning. The
centre of each cluster represents the cluster as the algorithm ensures convergence towards stable centroids
of clusters. The centroid is the centre or mean point, of the cluster. K is the number of clusters. After
initialisation, there are 3 steps in the K-means process.
Initialisation: set seed points (randomly)

• Step 1: Each object is placed in a cluster of the nearest seed point (centroid) measured with a
specific distance metric (Euclidean distance)
• Step 2: Estimate new centroid for each cluster in the current partitioning

• Step 3: Repeat Step 1; continue iterating until there are no more changes in membership in each
cluster.
Classification is a kind of arrangement where like data are classed together and separated from unlike
data; the main reasons behind classification is to (a) put knowledge in shape and storage, (b) do structural
analysis of the data being stored; and (c) figure out the relationship existing among different parts of the
structure found (Mirkin 1996).
Decision tree classification is used in this research. Decision trees are based on algorithms which split
data into branches. Unlike a tree where the root is at the bottom, a decision tree has its root node at the apex
of the tree (Ville et al 2013). The basis for building the decision tree is echoed in this root node: the name
of the field of data and the arrangement of the values that are contained in that field.
There are 3 types of nodes in a decision tree:

• Decision nodes – the squares

• Chance nodes – the circles

• Leaf or terminal or end nodes – the triangles (Bloomsbury Publishing 2013).

The algorithm for decision tree is probably the simplest among other classification algorithms. The
algorithm uses tree representation to create the classification. Each internal node of the tree reflects certain
characteristics of the system, and each leaf node represents a class label. There are 3 steps to contrasting
the decision tree:

• Step 1: At the root of the tree, place the most defining feature of the dataset

• Step 2: The training set is then split into subsets with values corresponding to their respective
attributes.
• Step 3: Redo step 1 and step 2 on each subset till there are terminal nodes in all the branches of
the tree.

Case Study
Surface and downhole data from a well in continental US was used in this study to evaluate the potential of
machine learning to identifying optimum zones while drilling. Six drilling parameters were of concerning
to ensure optimum drilling operations. These parameters formed the original data input into the machine
learning process. Table 1 shows the significance of each of these parameter to the drilling process. The data
for this well form a 1000 by 6 matrix which will now follow the process shown in Figure 3 in order to
inform the driller what the optimum zone for drilling is per formation drilled.
SPE-189284-STU 7

Table 1—Drilling Parameters and their significance

Results
The PCA results show that 99.9535 percent of the input variables was explained by the first principal
component as shown in Figure 4 hence achieving dimensionality reduction.
8 SPE-189284-STU

Figure 4—Dimensionality Reduction due to PCA Application

The plot shown in Figure 5 displays the principal components subspace alongside the coordinates of the
original data points. It can be observed that most of the points lie along the MSE coordinates hsowing that
MSE contains an aggregrate of the information contained in the other five variables.

Figure 5—Biplot of PC Coefficients, Scores and Original Data Points

The result of Kmeans done after PCA shows that for this data, 2 clusters are the most stable indicating
stick slip and whirling present as drilling took place as shown in Figure 6. The number of stable clusters was
picked based on the peak mean silhouette value which is 2 for this data. This value was confirmed when
Kmeans clustering was also carried out on the original data set. The peak mean silhouette value was also
2 as shown in Figure 6. The major benefit of running the Kmeans after PCA is done is to reduce the noise
in the data that often disturb the validity of the results.
SPE-189284-STU 9

Figure 6—Plot to Determine Most Stable Number of Cluster for Selection

Figure 7 compares clusters formed by applying Kmeans on just the original data and Kmeans on PCA
produced data. It is clearly seen that the distinction of the two clusters is higher after the PCA has been
done. This gives the driller a distinct idea of where the optimum zone is.

Figure 7—Kmeans Clusters for WOB and RPM plot

For more distinct clarity on where the optimum zone is, a decision tree is designed. The classes are chosen
based on conventional drilling solutions for stick slip and whirling as described in figure 8. In order to form
the basis for decision, the process was reversed to give the worst case scenario where stick slip and whirling
will definitely occur. The results in certain boundaries as shown in Figure 9. First, the truncated mean for
WOB and RPM in the input data is obtained and then four equation can be derived. The stick slip line is
the third reverse of 5 percent decrease in WOB, the low ROP line is the third reverse of 10 percent increase
in RPM, the backward whirling line is the third reverse of 10 percent decrease in RPM while the forward
whilring line is the third reverse of 10 percent increase in WOB. Based on these conditions, new data can be
classified into 5 classes. Class 1 is the "potential stick clip zone," class 2 is the "potential optimum zone,"
class 3 is the "potential backward whirling zone," class 4 is the "potential low ROP zone," and class 5 is
the "potential forward whirling zone."
10 SPE-189284-STU

Figure 8—Conventional Cures while Drilling (Schlumberger 2010).

Figure 9—Basis for Making Decisions to Determine Optimum Zone

Conclusions
Since the driller has to constantly manipulate available drilling parameters to mitigate vibration problems,
there is need to aid his decision making process. The process described in this paper practiced the use of
supervised (decision tree classification) and unsupervised (Kmeans clustering) machine learning to help
the driller visualize what appropriate value of WOB and RPM should be used when change is needed.
Based on literature, an optimum zone is formed which shows the driller how to avoid stick slip and whirling
problems that occur duringdrill string vibrations. The boundaries of the optimum zone helps determine the
best combination of WOB and RPM for optimum ROP. Running PCA on the data before clustering gives
the driller a more distinct view of where the optimum zone is. The zone for each data point can be determed
using decision tree classification.

Acknowledgements
The author wishes to thank his supervisor, Dr. Roman Shor, for his immense support and guidance
throughout the research that formed this paper. Tha author also thanks the University of Calgary for
providing necessary resources to facilitate the research.
SPE-189284-STU 11

Nomenclature
ANN Artificial Neural Network
BHA Bottomo Hole Assembly
LWD Logging While Drilling
MSE Mechanical Specific Energy
MWD Measurement-While-Drilling
NPT Non-productive Time
PCA Principal Component Analysis
PDC Polycrystalline Diamond Compact
PWD Pressure-While-Drilling
ROP Rate of Penetration
RPM Revolutions per Minute
WOB Weight on Bit

References
Alvarado, V., Ranson, A., Hernandez, K., Manrique, E., Matheus, J., Liscano, T., and Prosperi, N. 2002. Selection of EOR/
IOR Opportunities Based on Machine Learning. European Petroleum Conference, 29-31 October, Aberdeen, United
Kingdom. https://doi.org/10.2118/78332-MS
Belaskie J.P., Dunn M.D., Choo D.K. 1993. Distinct Applications of MWD, Weight on Bit, and Torque. SPE Drilling &
Completion. https://doi.org/10.2118/19968-PA
Ben-david, S. 2014. Understanding Machine Learning : From Theory to Algorithms. Cambridge University Press.
Bloomsbury Publishing. 2013. QFINANCE: The Ultimate Resource. Bloomsbury Information Ltd, Fourth edition.
ISBN-13: 978-1849300629.
Boonsri, K. 2014. Torque Simulation in the Well Planning Process. IADC/SPE Asia Pacific Drilling Technology
Conference, 25-27 August, Bangkok, Thailand. https://doi.org/10.2118/170500-MS
Cao, Q., Banerjee, R., Gupta, S., Li J., Zhou, W., and Jeyachandra, B. 2016. Data Driven Production Forecasting Using
Machine Learning. SPE Argentina Exploration and Production of Unconventional Resources Symposium, 1-3 June,
Buenos Aires, Argentina. https://doi.org/10.2118/180984-MS
Dupriest, F.E. 2006. Comprehensive Drill Rate Management Process to Maximize ROP. SPE Annual Technical
Conference and Exhibition, 24-27 September, San Antonio, Texas, USA. https://doi.org/10.2118/102210-MS
Dupriest, F. E., Witt, J. W., and Remmert, S. M. 2005. Maximizing ROP with Real-Time Analysis of Digital Data and MSE.
International Petroleum Technology Conference, 21-23 November, Doha, Qatar. https://doi.org/10.2523/IPTC-10607-
MS
Elshafei, M., Baig, M.M., Mysorewala, M.F., and Al-Majed, A. A. 2015. Control and Optimization of Directional Drilling
System. SPE Middle East Intelligent Oil and Gas Conference and Exhibition, 15-16 September, Abu Dhabi, UAE.
https://doi.org/10.2118/176759-MS
Ernst, S., Pastusek, P.E., and Lutes, P.J. 2007. Effects of RPM and ROP on PDC Bit Steerability. SPE/IADC Drilling
Conference, 20-22 February, Amsterdam, The Netherlands. https://doi.org/10.2118/105594-MS
Ethem, A. 2009. Introduction to Machine Learning. MIT Press. ISBN: 9780262012430.
Hammoutene, C. 2012. FEA Modeled MSE/UCS Values Optimize PDC Design for Entire Hole Section. North Africa
Technical Conference and Exhibition, 20-22 February, Cairo, Egypt. https://doi.org/10.2118/149372-MS
JPT Staff. 1998. Detecting Whirling Behavior of the Drill string. Journal of Petroleum Technology, May. https://
doi.org/10.2118/0598-0064-JPT
Kamata, M., Underhill, W., Meehan, R., and Nutt, L. 1997. Drill-Bit Seismic A Service For Drilling Optimization. SPWLA
38th Annual Logging Symposium, 15-18 June, Houston, Texas. SPWLA-1997-DD
Karlsson, H., and Brassfield, T. 1985. Performance Drilling Optimization. SPE/IADC Drilling Conference, 5-8 March,
New Orleans, Louisiana. https://doi.org/10.2118/13474-MS
Leskovec, J., Rajaraman, A., Ullman J. 2014. Mining of Massive Datasets, second edition. Cambridge University Press.
Louridas, P., and Ebert, C. 2016. Machine Learning. IEEE Software, Volume: 33, Issue: 5, Sept.-Oct. 2016, 10.1109/
MS.2016.114
Mirkin B. 1996. Mathematical Classification and Clustering. Kluwer Academic Publishers. ISBN 978-1-4613-0457-9
Murphy, K. P. 2012. Machine Learning, a Probabilistic Perspective. MIT Press ISBN: 9780262018029.
12 SPE-189284-STU

Paes, P., Aragao A.F.L., Felicissimo R.S., and Chen D.C-K. 2005. Cost-Effective Drilling Optimization Technologies in
Campos Basin. SPE Latin American and Caribbean Petroleum Engineering Conference, 20-23 June, Rio de Janeiro,
Brazil. https://doi.org/10.2118/94785-MS
Patrick, B. 2017. Smart Condition Monitoring Using Machine Learning. SPE Intelligent Oil and Gas Symposium. https://
doi.org/10.2118/187936-MS
Rabia, H. 2002. Well Engineering & Construction. Entrac Consulting. ISBN-10: 0954108701.
Rajnauth, J.J. 2003. Reduce Torsional Vibration and Improve Drilling Operations. SPE Latin American and Caribbean
Petroleum Engineering Conference, 27-30 April, Port-of-Spain, Trinidad and Tobago. https://doi.org/10.2118/81174-
MS
Schlumberger. (2010). Drilling Dynamics Sensors and Optimization. 10-DR-0169 https://www.slb.com/~/media/Files/
drilling/brochures/mwd/drilling_dynamics_sensors_opt_br.ashx
Solano, Y. P., Uribe, R., Frydman, M., Saavedra, N.F., and Calderon, Z.H. 2007. A Modified Approach to Predict Pore
Pressure Using the D Exponent Method : An Example from the Carbonera. CT&F - Ciencia, Tecnología y Futuro -
Volume 3, Númber 3.
Tansev E. 1975. A Heuristic Approach to Drilling Optimization. Fall Meeting of the Society of Petroleum Engineers of
AIME, 28 September-1 October, Dallas, Texas. https://doi.org/10.2118/5546-MS
Torres-Cabrera N., Pozo M., and Finessi A.A. 2017. Drilling Optimization of Argentina Horizontal Tight Sand Wells.
SPE Latin America and Caribbean Petroleum Engineering Conference, 18-19 May, Buenos Aires, Argentina. https://
doi.org/10.2118/185466-MS
Vasan, K.K., and Surendiran, B. 2016. Dimensionality reduction using Principal Component Analysis for network
intrusion detection. Perspectives in Science Volume 8: 510-512. Elsevier. https://doi.org/10.1016/j.pisc.2016.05.010
Ville, B., and Neville, P. 2013. Decision Trees for Analytics: Using SAS Enterprise Miner. SAS Institute. ISBN
978-1-61290-252-4.
Wu, S. X., Paez, L., Partin, U., and Agnihotri, M. 2010. Decoupling Stick-slip and Whirl to Achieve Breakthrough in
Drilling Performance. IADC/SPE Drilling Conference and Exhibition. https://doi.org/10.2118/128767-MS
Zhang, Q., Jiarong, S., 1991. The Application of Machine Learning to Rock Mechanics. 7th ISRM Congress, 16-20
September, Aachen, Germany. ISRM-7CONGRESS-1991-167

You might also like