SPE/IADC 166793

Formation Prediction Model based on Drill String Vibration Measurements

Using Laboratory Scale Rig
Abdolali Esmaeili, TDE Thonhauser Data Engineering, Behzad Elahifar, ADS Advanced Drilling Solutions, Rudolf
K. Fruhwirth, NGS Neuro Genetic Solutions, Gerhard Thonhauser, University of Leoben

Copyright 2013, SPE/IADC Middle East Drilling Technology Conference and Exhibition

This paper was prepared for presentation at the SPE/IADC Middle East Drilling Technology Conference and Exhibition held in Dubai, UAE, 7–9 October 2013.

This paper was selected for presentation by an SPE/IADC 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 or the International Association of Drilling Contractors and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers or the International Association of Drilling Contractors, 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 or the International Association of Drilling Contractors 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/IADC copyright.

Abstract
Characteristics of formations usually cause changes in the drilling parameters such as weight on bit and drill string
rotational speed as well as drill string dynamic behavior. When the drill bit breaks up formations, some of the energy dedicated
to drilling is transmitted to the drill string as axial, lateral and torsional vibrations. Although these vibrations are affected by
the drilling parameters, it has been experienced that vibration measurements can be used to evaluate formations during drilling
a well. For this purpose a laboratory scale drilling rig was used to perform numerous experiments. During the drilling
operations, several drilling parameters as well as vibrations with three-component accelerometers attached to the drill string
were recorded continuously. Numerous uniform concrete and rock samples were drilled by a double roller cone bit. The
uniaxial compressive strength of all concretes and rocks were measured prior to the experiments. The tests were executed
using different combinations of weight on bit and rotary speed.
For a subsequent evaluation statistical features were extracted in a first step from the recorded data in both, the time
domain and the frequency domain. Those features formed the base for the formation classification by linear and non-linear
models. Although linear models like Bayesian inference are not suited for that challenge, it turned out that non-linear models
like neural networks provide excellent results in both, the training and the test data-set. Finally, the application of parameter
selection methods showed that mechanical specific energy and rate of penetration in combination with higher order frequency
moments of all recorded vibration types play the major role in the formation classification.
Introduction
Knowing the characteristics of formations can enhance the quality of drilling operations. The driller optimizes drilling
operation by adjusting drilling parameters (weight on bit and rotary speed) to apply enough energy to the rocks and crash them
by overcoming their compressive strength (Teale). Whilst drilling a wellbore usually different types of rocks, with different
characteristics, are drilled by the drill bit. Therefore a wide range of combinations of weight on bit and rotary speed are used to
optimize drilling operations and compensate formation changes. As the vibrations (mostly axial vibration) are a consequent of
the drill bit-formation interaction, they also vary when the formation changes.
In the other words, it is possible to evaluate and predict the type of formation during drilling operations if the drilling
parameters and drill string vibrations are monitored in real time. The results of experimental tests have been revealed that there
are no linear relations between characteristics of formations and drilling parameters and drill string vibrations. Therefore non-
linear models are a good choice to establish a relation between formation characteristics, drilling parameters and drill string
vibration data. Artificial Neural Networks (ANN) are one of the most popular non-linear models to recognize complex patterns
and relationship between different parameters (Esmaeili et al., 2012).
Artificial neural networks have been used in the petroleum industry to ease the process of recognizing the
relations between different parameters without any assumptions about the nature and distribution of the data.
Dashevskiy et al. used neural networks to simulate and predict the dynamic behavior of a complex multi-parameter drilling
system under specific drilling conditions. Control parameters (hook load, rotary speed, mud flow rate and mud properties),
plant characteristics (parameters related directly to drilling downhole equipment such as BHA and bit design) and media
parameters (lithology, wellbore geometry and well profile) have been fed into the neural network as input channels. Rate of
penetration, drill string vibration, downhole weight on bit and rotary speed have been selected as output; the parameters were
likely to be controlled. A multilayer feedforward neural network (MFNN) has been used to build the models. The authors have
2 SPE/IADC 166793

published positive results of using neural networks as an accurate simulator of the non-linear drilling system through the
observation of its dynamic behavior.
Mohaghegh et al. used neural networks to predict permeability of the formations based on log and core data of five wells.
They have presented that the trained neural network was able to estimate the permeability of the Granny Creek Field in Clay
and Roan Counties, West Virginia.
Guo et al. used back propagation multilayer perceptrons to extract and recognize the structural lineaments and lithological
information from geology data.
In this paper a formation prediction model is presented to predict the type of formation drilled based on measured drilling
parameters and drill string vibration data in laboratory scale setup. The model was built using artificial neural networks.
Sequential forward selection was used to rank the parameters with respect to the impact on the model.
Artificial Neural Networks
Artificial neural networks (ANN) represent a class of heuristic model builders inspired by the human brain and nervous
system. Similar to biological brains, neural networks have the capacity to learn, memorize and create their own relationships
among data with the advantage that no formal description of the underlying physics is required. They provide more intuitive
solutions with good predictive accuracy to complicated problems.
Artificial neural networks have the ability to recognize complex patterns quickly with a high degree of accuracy without
any assumptions about the nature and distribution of the data. McCulloch and Pitts proposed the first systematic neural
networks in 1943. Major researches have been done to develop the idea of neural network until the 1970's. However, the most
significant steps in developing the more robust theoretical aspects of neural network have been conducted since last two
decades.
Artificial neural networks are not as complex as human neural system; however, they are similar in two principles. First,
creating units of both networks are simple and are highly interconnected. Second, the function of the network is defined by
connections between neurons (Hagan et al.). An artificial neuron as shown in Figure 1 is the basic element of a neural network.
It consists of four basic components:

Input data; Input data can be any scalar values to be fed into models.

Weights; the weights determine the strength of every input channel. The weights make some inputs more important
than other, so they have a great impact on the summation function of the processing elements. The inputs are multiplied
by a set of weights (w1 ... wn).

Summation function; the weighted inputs are summed up resulting in the so-called local receptive field. But the
function inside such an artificial neuron can be more complicated than building just the simple sum of the weighted
inputs. Such a function can select the minimum, maximum, majority or apply several normalizing algorithms.

Activation function; the local receptive field is fed through an activation function introducing non-linearity into the
system. There are different sophisticated activation functions such as hard limit, symmetrical hard limit, linear,
saturating linear, symmetric saturating linear, log-sigmoid, hyperbolic tangent sigmoid, positive linear and competitive.

Output values; the result of the activation function of a neuron is provided as output. The output can be the final result
or used as another input into subsequent neurons. (Hagan et al.; Zilouchian and Jamshidi; Anderson and McNeill;
Esmaeili et al., 2012).

Figure 1. Basic elements of an artificial neuron (Hagan et al.; Anderson and McNeill)

Multi-layer perceptron networks, trained with a back-propagation algorithm, are one of the most popular networks in use
today. However, finding an optimal choice for the number of hidden layers in multilayer perceptron networks is time
consuming and there is no analytical solution available.
The completely connected perceptron (CCP) (Figure 2) has in opposite to the multilayer perceptron the advantage, that a
network growing process is straight forward. In the first network generation, a network cluster containing networks without
any hidden units is trained. From all the networks in the cluster, the network with the lowest domain error is selected as the
best network in that generation; it will be compared with the best network of the next generation. In the second generation, one
hidden unit is added to the network; now it is a CCP with one hidden unit. In this generation again, the best network out of the
SPE/IADC 166793 3

networks in that cluster is selected. If it performs better than the previous network, it is kept for comparison with the best
network of the next generation. This process is repeated until a stable solution or a maximal number of hidden neurons are
obtained. Figure 3 illustrates the process of generating five layer networks using improved completely connected perceptron
architecture.

Figure 2. Improved Completely Connected Perceptron (iCCP) (cVision software manual)

Figure 3. Growing improved completely connected perceptron architecture (cVision software manual)

The size of a network is characterized by the number of hidden neurons and their interrelationship defines the complexity
of the network. If the network is too small, it will not completely solve a problem or just parts of it. On the other hand a
network with a too large number of hidden layers has the tendency to memorize the data (Figure 4). As a consequence too
much concentration on training data the network tends towards modeling the noise. These cases are called under-fitting or
over-fitting, respectively.

Figure 4. Oversized vs. regular model (cVision software manual)

Local adaptive learning has been used to overcome the challenges in manually tuning the learning rate. In local adaptive
learning, each weight has its “private” learning rate, and that learning rates are in addition optimized during the learning
process. Therefore it leads to fast learning speed without the necessity of human interaction (cVision Software Manual).
It is always a challenge to determine which of the input parameters have the most impact on the model. Sequential forward
selection (SFS) is a systematic procedure which tackles that challenge. The procedure begins by considering each of the inputs
individually and selecting the one which gives the best value for the selection criterion. At each successive stage of the
algorithm, one additional input (feature) is added to the set. The feature performing the best at that stage is selected as
permanent input to the model, the process is repeated until enough input channels are selected (Bishop).
Experimental Setup
To collect real data for the formation prediction model, different type of rocks (sandstone, limestone and dolomite) as well
as different type of concrete in terms of water to cement ratio (0.4, 0.5 and 0.6) were drilled with different combinations of
4 SPE/IADC 166793

weight on bit and rotary speed. A laboratory scale drilling rig (CDC miniRig) in combination with a vibration sensor sub
attached to the drill string was used to perform experimental tests and measure drilling parameters and vibrations.
The CDC miniRig is capable of drilling small diameter holes using 2 3/8 inch double cone drill bit in different types of
rock and concrete. Figure 5a shows the components of the CDC miniRig. Block position, hole depth, hook load, weight on bit,
rotary speed of the drill string and torque on bit were recorded in real-time with a sampling frequency of 1 Hz. The rate of
penetration and the mechanical specific energy (MSE) were calculated based on the measured data (Esmaeili et al., 2013).

Figure 5. CDC miniRig (a) and Vibration Sensor Sub (b) (Esmaeili et al., 2013)

The vibration sensor sub (Figure 5b) was used to record the drill string vibrations. It consists of a housing, an
accelerometer sensor with a measurement range of ±16 g attached to the drill string, a wireless communication system and
batteries. The three-component accelerometer measures vibrations in three orthogonal dimensions x, y, and z. Vibration data
were sampled with a sampling frequency of 3200 Hz. First and second order frequency moments of all vibrations were
calculated in addition (Esmaeili et al., 2013).
Samples Properties
To provide data for building the model, sandstone, limestone and dolomite as well as different types of concrete were
drilled. Although identical cement and aggregates were used, bedding effects during preparing the concretes were unavoidable.
Therefore in concretes WC05 and WC06 the zone with weak strength (soft part) was occurred when the cement was only
mixed with water and small size of aggregates (WC05-S and WC06-S). However in the hard part of concretes (WC05-H and
WC06-H), the cement was mixed with water and bigger size of aggregates. The mechanical and petrophysical properties of the
samples used for the experiments are listed in Table 1 and Table 2, respectively.
Table 1. Mechanical properties of the rock and concrete samples
Uniaxial Young’s Dynamic Young’s Dynamic Dynamic Young’s
Deformation Poisson’s Dynamic Poisson’s
Samples Compressive Modulus, Modulus (dry), Poisson’s Modulus
Modulus, GPa Ratio Ratio (Saturated)
Strength, MPa GPa GPa Ratio (dry) (Saturated), GPa
Sandstone 42.52 11.59 63.83 0.20 10.22 0.32 9.79 0.38
Limestone 60.70 74.23 68.5 0.21 61.75 0.26 55.6 0.34
Dolomite 115.52 45.44 72.79 0.24 32.88 0.13 38.71 0.34
Concrete WC04 51.89 34.03 30.99 0.07 33.31 0.25 26.77 0.34
Concrete WC05-
41.58 26.86 22.79 0.09 31.33 0.13 34.30 0.37
Hard Part
Concrete WC05-
39.67 19.14 18.5 0.1 26.39 0.33 18.79 0.4
Soft Part
Concrete WC06-
35.90 25.37 21.64 0.1 29.96 0.13 22.16 0.33
Hard Part
Concrete WC06-
37.84 25.96 23.86 0.1 16.87 0.3 13.87 0.39
Soft Part
SPE/IADC 166793 5

Table 2. Petrophysical properties of all the samples

Grain Density, Porosity Effective Average
Samples Density, g/cc
g/cc (%) Porosity (%) Permeability, mD
Sandstone 1.99 2.65 22.30 18.78 127.25
Limestone 2.69 2.71 0.6 0.24 0.86
Dolomite 2.63 2.83 6.78 2.77 15.63
Concrete WC04 2.29 2.58 11.25 10.48 0.14
Concrete WC05-Hard Part 2.16 2.66 18.81 15.79 0.8
Concrete WC05-Soft Part 2.10 NA NA 16.54 0.48
Concrete WC06-Hard Part 2.21 2.58 14.34 17.63 0.33
Concrete WC06-Soft Part 1.81 NA NA 25.45 0.48

Figure 6. Recorded and calculated drilling parameters and vibration data, sandstone cube,

WOB=800 N, different RPMs, 2 3/8” bit, ±16 g accelerometer sensor

6 SPE/IADC 166793

Experimental Tests
To have a sufficient data set for building the model, all samples were drilled with different combinations of weight on bit
(400, 600 and 800 N) and rotary speed (150, 120, 100, 80, 60 and 40 rpm). Drilling parameters and drill string vibrations were
recorded during all operations. Figure 6 shows recorded as well as calculated data versus time of the sandstone sample drilled
with constant weight on bit of 800 N and different rotary speed of 150, 120, 100, 80, 60 and 40 rpm.
These experimental tests provided 24 real data sets with measured drilling parameters such as block position (blue line in
the first chart in Figure 6), rotary speed of drill string (orange line in the second chart), weight on bit (red line in the second
chart) torque (yellow line in the first chart) and calculated drilling parameters such as hole depth (red line in the first chart),
instantaneous rate of penetration (gray in the first chart) and mechanical specific energy (gray line in the second chart) as well
as corresponding x, y and z component of vibrations. Based on the vibration measurements, first and second order frequency
moments (light and dark green lines in the last three charts) were calculated using the following equations:










𝐹𝐹𝐹𝐹
=



Equ. 1







𝐴𝐴 𝜔𝜔 =   𝐹𝐹(𝜔𝜔) Equ. 2
In above equations FMn denotes the n order frequency moment of the vibration data and ω the radian frequency (s-1).
th

F(ω) stands for the Fourier transform of a vibration signal and A(ω) denotes the corresponding amplitude spectrum.
Table 3. Drilling parameters used in formation prediction model
Parameters Unit Comment
wob:av N Average of weight on bit
wob:std N Standard deviation of weight on bit
wob:skew N Skewness of weight on bit
wob:kur N Kurtosis of weight on bit
wob:med N Median of weight on bit
wob:max N Maximum of weight on bit
wob:min N Minimum of weight on bit
rpm:av rpm Average of rotary speed
rpm:std rpm Standard deviation of rotary speed
rpm:skew rpm Skewness of rotary speed
rpm:kur rpm Kurtosis of rotary speed
rpm:med rpm Median of rotary speed
rpm:max rpm Maximum of rotary speed
rpm:min rpm Minimum of rotary speed
rta:av N.m Average of rotary torque
rta:std N.m Standard deviation of rotary torque
rta:skew N.m Skewness of rotary torque
Drilling parameters rta:kur N.m Kurtosis of rotary torque
rta:med N.m Median of rotary torque
rta:max N.m Maximum of rotary torque
rta:min N.m Minimum of rotary torque
mse:av kWh/m3 Average of mechanical specific energy
mse:std kWh/m3 Standard deviation of mechanical specific energy
mse:skew kWh/m3 Skewness of mechanical specific energy
mse:kur kWh/m3 Kurtosis of mechanical specific energy
mse:med kWh/m3 Median of mechanical specific energy
mse:max kWh/m3 Maximum of mechanical specific energy
mse:min kWh/m3 Minimum of mechanical specific energy
rop:av mm/min Average of rate of penetration
rop:std mm/min Standard deviation of rate of penetration
rop:skew mm/min Skewness of rate of penetration
rop:kur mm/min Kurtosis of rate of penetration
rop:med mm/min Median of rate of penetration
rop:max mm/min Maximum of rate of penetration
rop:min mm/min Minimum of rate of penetration
mdh:av mm Hole Depth

Formation Prediction Model and Results

To create the prediction model, real data obtained from experimental tests were used. To reduce the volume of the data to
be fed into model, both the drilling parameters and vibration data were sampled in a moving time window of 20 seconds of
length. For training the neural networks, the data are usually separated into three sub-sets; learning, validation and testing sub-
set. The learning sub-set is used for learning the neural network model and the validation sub-set is used for making decisions.
The testing sub-set, however, is not being used by the network during learning and validation phases and it is therefore
statistically independent. Thus, the error calculated using the testing sub-set is a measure for the quality of the created model.
SPE/IADC 166793 7

Statistical features like average, standard deviation, skewness, kurtosis, median, minimum and maximum values were
calculated from all 6 main drilling parameters which are listed in Table 3. For the vibration data, above features as well as
interdecile range, interquartile range and span range were calculated (Table 4). Thus a total of 71 input channels out of the
drilling parameters and drill string vibrations were available as input to the neural networks. The desired output of the network
was a categorical channel representing the formation type. The whole dataset was separated prior to training into three sub-
sets: 60% of the data were used for leaning, 20% for validation and 20% for testing. For building the model, the improved
completely connected perceptron (iCCP) architecture was used.
Table 4. Drill string vibration data used in formation prediction model
Parameters Unit Comment
accx:av g Average of x-component of vibration
accx:fm1 Hz 1st order frequency moment of x-component of vibration
accx:fm2 Hz 2nd order frequency moment of x-component of vibration
accx:idr g Interdecile range of x-component of vibration
accx:iqr g Interquartile range of x-component of vibration
accx:kurt g Kurtosis of x-component of vibration
accx:skew g Skewness of x-component of vibration
accx:med g Median value of x-component of vibration
accx:max g Maximum value of x-component of vibration
accx:min g Minimum value of x-component of vibration
accx:span g Span range of x-component of vibration
accx:std g Standard deviation of x-component of vibration
accy:av g Average of y-component of vibration
accy:fm1 Hz 1st order frequency moment of y-component of vibration
accy:fm2 Hz 2nd order frequency moment of y-component of vibration
accy:idr g Interdecile range of y-component of vibration
accy:iqr g Interquartile range of y-component of vibration
Drill String Vibrations accy:kurt g Kurtosis of y-component of vibration
accy:skew g Skewness of y-component of vibration
accy:med g Median value of y-component of vibration
accy:max g Maximum value of y-component of vibration
accy:min g Minimum value of y-component of vibration
accy:span g Span range of y-component of vibration
accy:std g Standard deviation of y-component of vibration
accz:av g Average of z-component of vibration
accz:fm1 Hz 1st order frequency moment of z-component of vibration
accz:fm2 Hz 2nd order frequency moment of z-component of vibration
accz:idr g Interdecile range of z-component of vibration
accz:iqr g Interquartile range of z-component of vibration
accz:kurt g Kurtosis of z-component of vibration
accz:skew g Skewness of z-component of vibration
accz:med g Median value of z-component of vibration
accz:max g Maximum value of z-component of vibration
accz:min g Minimum value of z-component of vibration
accz:span g Span range of z-component of vibration
accz:std g Standard deviation of z-component of vibration

The confusion matrices for the learning (Table 5), validation (Table 6) and testing sub-set (Table 7) indicate that the model
shows very good accuracy for predicting and recognizing the formation type. For creating that model the first 17 iInput
parameters identified by sequential forward selection as shown in Figure 7 have been used. Therefore our conclusion is that it
is possible to predict the formation type during drilling in real time if the model is fed with appropriate drilling parameters and
drill string vibration data.
Table 5. Confusion matrix - learning sub-set
formtype:Learning  Correct  Classification  Rate 99.7%
concwc04 concwc05-­‐hard concwc05-­‐soft concwc06-­‐hard concwc06-­‐soft dolomite limest sandst
98.7% 99.1% 100.0% 99.6% 100.0% 100.0% 100.0% 100.0%
concwc04 220 2 0 0 0 0 0 0
concwc05-­‐hard 3 213 0 1 0 0 0 0
concwc05-­‐soft 0 0 223 0 0 0 0 0
concwc06-­‐hard 0 0 0 223 0 0 0 0
concwc06-­‐soft 0 0 0 0 233 0 0 0
dolomite 0 0 0 0 0 218 0 0
limest 0 0 0 0 0 0 229 0
sandst 0 0 0 0 0 0 0 217
False  Classified  Samples 3 2 0 1 0 0 0 0
Total  Samples 223 215 223 224 233 218 229 217
8 SPE/IADC 166793

Table 6. Confusion matrix - validation sub-set

formtype:Validation  Correct  Classification  Rate 98.8%
concwc04 concwc05-­‐hard concwc05-­‐soft concwc06-­‐hard concwc06-­‐soft dolomite limest sandst
95.8% 98.6% 98.8% 97.6% 100.0% 100.0% 100.0% 100.0%
concwc04 69 0 0 1 0 0 0 0
concwc05-­‐hard 0 73 0 1 0 0 0 0
concwc05-­‐soft 0 0 79 0 0 0 0 0
concwc06-­‐hard 3 1 0 80 0 0 0 0
concwc06-­‐soft 0 0 1 0 70 0 0 0
dolomite 0 0 0 0 0 50 0 0
limest 0 0 0 0 0 0 74 0
sandst 0 0 0 0 0 0 0 92
False  Classified  Samples 3 1 1 2 0 0 0 0
Total  Samples 72 74 80 82 70 50 74 92

Table 7. Confusion matrix - test sub-set

formtype:Testing  Correct  Classification  Rate 98.0%
concwc04 concwc05-­‐hard concwc05-­‐soft concwc06-­‐hard concwc06-­‐soft dolomite limest sandst
95.6% 94.2% 97.3% 98.6% 100.0% 98.6% 98.8% 100.0%
concwc04 65 3 0 0 0 1 0 0
concwc05-­‐hard 3 65 0 1 0 0 0 0
concwc05-­‐soft 0 0 73 0 0 0 0 0
concwc06-­‐hard 0 1 2 69 0 0 0 0
concwc06-­‐soft 0 0 0 0 80 0 0 0
dolomite 0 0 0 0 0 68 1 0
limest 0 0 0 0 0 0 81 0
sandst 0 0 0 0 0 0 0 81
False  Classified  Samples 3 4 2 1 0 1 1 0
Total  Samples 68 69 75 70 80 69 82 81

Based on sequential forward selection procedure, 2552 complete models were created to rank which parameters have the
most impact on the formation prediction model. The result is illustrated in Figure 7. It seems from the figure, the vibrations
have great impact on the model and can prove the statement that the characteristics of the formations have a great impact on
vibrations. In the other words, monitoring the vibrations (mostly axial vibration) can aid to increase the possibility of
recognizing the formation type. The operational drilling parameters such as rotary speed, torque and weight on bit have also
influences in predicting the formations since they are affecting the rate of penetration and the mechanical specific energy.

Figure 7. Result of sequential forward selection RMS error, correct classification rate (CCR)

Conclusions
Studying the results of experimental tests determines that there are linear relations between few drilling parameters,
vibrations data and formation characteristics. However, linear equations are not a sufficient means to describe a relation
between them in a unique definition.
SPE/IADC 166793 9

To recognize this relation, non-linear models are a good choice to find the relation between formation characteristics,
drilling parameters and drill string vibrations. Artificial Neural Networks are one of the most popular non-linear models to
recognize complex patterns and relations between different parameters without any assumptions about the nature and
distribution of the data. Results of the experimental tests were used as input channels as well as output channels to create the
neural networks. The formation prediction model proved that it is possible to use the drilling parameters and drill string
vibration data as input channel to predict and recognize type of formation.
To overcome the exhaustive search for determining the parameters that have the most impact on the model, sequential
forward selection method is a right choice to rank the input channels based on their impact on the model.
However, the model is not a universal model and it should be trained with different suitable data bases, resulting from both,
laboratory and real rig experimental tests. Correct selection of input data has a great impact on the output results.
References
Anderson, D., McNeill, G., “Artificial Neural Network Technology”, The Data & Analysis Center for Software Technical Report, 1992.
Bishop, C.M., “Neural Networks for Pattern Recognition” Oxford University Press, 1995.
cVision Software Manual, NGS - Neuro Genetic Solutions GmbH, 2012.
Dashevskiy, D., Dubinsky, V. and Macpherson, J.D., “Application of Neural Networks for Predictive Control in Drilling Dynamics”, SPE
Annual Technical Conference and Exhibition, 3-6 October, Houston, Texas, U.S.A., 1999.
Esmaeili, A., Elahifar, B., Fruhwirth, R.K., Thonhauser, G., “ROP Modeling using Neural Network and Drill String Vibration Data”, SPE
Kuwait International Petroleum Conference and Exhibition, 10-12 December, Kuwait, Kuwait, 2012.
Esmaeili, A., Elahifar, B., Fruhwirth, R.K., Thonhauser, G., “Effect of Formations Compressive Strength on Drill String Vibrations”, 6th
International Petroleum Technology Conference, 26-28 November, Beijing, China, 2013.
Guo, Y., Hansen, R.O. and Harthill, N., “Artificial Intelligence I: Neural Networks in Geophysics”, SEG Annual Meeting, 25 - 29October,
New Orleans, Louisiana, U.S.A., 1992.
Hagan, M.T., H.B. Demuth and M. Beale., “Neural Network Design”, PWS Publishing Company, 1996.
McCulloch, W.W. and Pitts, W., “A Logical Calculus of Ideas Imminent in Nervous Activity”. Bulletin of Mathematical Biophysics, Vol. 5,
pp. 115−133, 1943.
Mohaghegh, S., Arefi, R., Ameri, S., and Rose, D., “Design and Development of an Artificial Neural Network for Estimation of Formation
Permeability”, SPE Petroleum Computer Conference, 31 July-3 August, Dallas, U.S.A., 1994.
Teale, R., “The Concept of Specific Energy in Rock Drilling”, International Journal of Rock Mechanics and Mining Sciences &
Geomechanics, Vol. 2, Issue 1, pp. 57–73, 1965.
Zilouchian, A., Jamshidi, M. (Edited), “Intelligent Control Systems Using Soft Computing Methodology”, ISBN: 978-1-4200-5814-7, DOI:
10.1201/9781420058147.ch2, 2011.