Journal of

Manufacturing and
Materials Processing

Article
Machining Chatter Prediction Using a Data
Learning Model
Harish Cherukuri 1 , Elena Perez-Bernabeu 2 , Miguel Selles 3 and Tony Schmitz 4, *
1 Department of Mechanical Engineering and Engineering Science, University of North Carolina at Charlotte,
Charlotte, NC 28223, USA; harish.cherukuri@uncc.edu
2 Department of Applied Statistics, Operational Research, and Quality, Universitat Politècnica de València,
03802 València, Spain; elenapb@eio.upv.es
3 Department of Mechanical Engineering and Materials, Universitat Politècnica de València, 03802 València,
Spain; maselles@dimm.upv.es
4 Department of Mechanical, Aerospace, and Biomedical Engineering, University of Tennessee, Knoxville,
TN 37996, USA
* Correspondence: tschmit2@utk.edu




Received: 17 April 2019; Accepted: 6 June 2019; Published: 8 June 2019


Abstract: Machining processes, including turning, are a critical capability for discrete part production.
One limitation to high material removal rates and reduced cost in these processes is chatter, or unstable
spindle speed-chip width combinations that exhibit a self-excited vibration. In this paper, an artificial
neural network (ANN)—a data learning model—is applied to model turning stability. The novel
approach is to use a physics-based process model—the analytical stability limit—to generate
a (synthetic) data set that trains the ANN. This enables the process physics to be combined with data
learning in a hybrid approach. As anticipated, it is observed that the number and distribution of
training points influences the ability of the ANN model to capture the smaller, more closely spaced
lobes that occur at lower spindle speeds. Overall, the ANN is successful (>90% accuracy) at predicting
the stability behavior after appropriate training.

Keywords: turning; machine learning; neural network; stability; chatter

1. Introduction
Material removal processes, including turning and milling, are widely applied in industry.
While significant advances have been achieved over the last decades, one limitation to high material
removal rates is chatter, or unstable cutting conditions. The result is poor surface quality and potential
damage to the tool, workpiece, and machine.
The control of stability in turning operations is crucial in industry and ongoing research efforts
address chatter avoidance. Siddhpura and Paurobally [1] completed a literature review of chatter
prediction in turning. They classified the techniques for chatter stability prediction as stability lobe
diagrams, Nyquist plots, and finite element analyses. Their study also discussed the experimental
techniques for chatter stability prediction and detection and separated them into three main groups:
signal acquisition and processing techniques; chip analysis; and artificial intelligence techniques,
including neural networks, hidden Markov models, and fuzzy logic. The authors noted that the
number of publications featuring artificial intelligence techniques was low at the time of publication
(only 11 from 1978 to 2012), although additional work has been done since then.
Chanda and Dwivedy [2] developed the governing nonlinear equations of motion for turning,
considering both the workpiece and the tool to be flexible. The regenerative effect due to inherent
time delay was considered. Copenhaver et al. [3] described a periodic sampling-based method for

J. Manuf. Mater. Process. 2019, 3, 45; doi:10.3390/jmmp3020045 www.mdpi.com/journal/jmmp

J. Manuf. Mater. Process. 2019, 3, 45 2 of 15

identifying the stability of modulated tool path turning (MTP). They compared a periodic sampling
metric with the traditional frequency domain approach, where the frequency spectrum is analyzed to
identify the turning stability.
Filippov et al. [4] studied the transition from the stable turning to chatter using acoustic emission
signals. High frequency peaks appeared in transition to the chatter mode. A mathematical model of
turning was presented by Gerasimenko et al. [5], where the stability limit for turning a thin-walled
cylindrical part was defined. Gouskova et al. [6] determined the stability of a continuous cutting
process for an arbitrary arrangement of two cutters. Gyebroszki et al. [7] created two models and
combined them later, a surface regeneration model for turning operations and a chip formation model.
They compared the results obtained and found that the time scale for the second model (chip formation)
was smaller than the first one.
Hajdu et al. [8] incorporated noise and uncertainties in chatter predictions. In their approach,
they did not apply filtering or parameter identification to the frequency response functions (FRFs) and
presented a frequency domain method for the analysis of stability in machining. In the application of
this method to a model for orthogonal cutting, the authors found that the robust stability boundaries
are lower to those showed by the FRF alone. Huang et al. [9] also considered uncertainties in their
analysis. They used a probabilistic method (Monte Carlo simulation) and found that, when compared
with a deterministic solution, Monte Carlo simulation provided better results by considering the
influence of random parameters.
Mousavi et al. [10] presented a numerical approach for the prediction of a manipulator behavior
in machining tasks. They established theoretical stability limits taking into account the variability
of the robot dynamics within its workspace. This made an auto-adaptation of cutting parameters
and robot configurations along the trajectory possible. The authors generated a stability diagram for
chatter in milling taking into account the kinematic redundancy variable. The theory was validated
with experimental robotic machining trials. Liu et al. [11] investigated the probability of stability for
turning. The authors defined and represented a reliability lobe diagram to identify stable and unstable
zones, rather than the traditional stability lobe diagram (SLD). The reliability was calculated using the
FOSM (first-order second moment) and Monte Carlo methods, and was compared to the traditional
stability lobe diagram.
Khasawneh and Munch [12] proposed a new approach for determining the stability of stochastic
dynamical systems by examining their time series using topological data analysis. Two statistical
approaches (three sigma edit rule analysis and principal component analysis) were used by Jiménez et
al. [13] to predict chatter, obtaining accuracy rates over 75%.
The receptance coupling method was used by Jasiewicz and Powalka [14] for determining the
lathe workpiece dynamics and inverse receptance coupling was applied for the spindle dynamics.
Lu et al. [15] focused on turning process and proposed a model for chatter prediction for turning
a tailstock-supported flexible rod. They considered the tool position effect along the workpiece axis.
The variation of theoretical and experimental tool locations was within 9%.
Tyler et al. [16] studied the process damping force in turning to propose a stability model. This was
dependent on several variables, including the chip width and cutting speed. Experiments were
completed to validate the time domain simulation, which provided accurate predictions.
Chatter prediction using machine learning techniques has also been studied, including neural
networks, support vector machines, and others. Ahmad et al. [17] developed two different models of
extreme learning techniques using random weights and hidden nodes. Lamraoui et al. [18] applied
a neural network (NN) and the input data based on signal analysis was used to predict milling stability.
Gupta et al. [19] based their research on the combination of several artificial techniques, such as support
vector regression (SVR), genetic algorithms (GA) or artificial neural networks (ANN). They trained the
model using the process parameters as input values. Power, tool wear and surface roughness were the
output variables.
J. Manuf. Mater. Process. 2019, 3, 45 3 of 15

Jurkovic et al. [20] compared the performance of three machine learning methods for the prediction
ofJ. Manuf. Mater.parameters
operating Process. 2019, 3,
in45high-speed turning. Observed parameters were the surface roughness 3 of 15
(Ra), cutting force (Fc ), and tool life (T). Polynomial (quadratic) regression, SVR, and ANN were used.
were used.regression
Polynomial Polynomial regression demonstrated
demonstrated the best performancethe best in performance
for Fc and Ra in for Fc and
prediction, Ra prediction,
while the ANN
while the ANN showed the best
showed the best performance for T prediction. performance for T prediction.
Khasawnehetetal.
Khasawneh al.[21]
[21]combined
combinedseveral
severaldeterministic
deterministicand andstochastic
stochasticmodels
modelstotocreate
createpersistence
persistence
diagrams of turning stability. The approach was intended for chatter
diagrams of turning stability. The approach was intended for chatter classification using signals classification using signals
produced by complicated and noisy manufacturing systems. Yao et
produced by complicated and noisy manufacturing systems. Yao et al. [22] developed a vector for al. [22] developed a vector for
chatterprediction
chatter predictionininmachining.
machining.ItItwas wasbased
basedon onthethewavelet
waveletpacket
packetenergy
energyratio
ratioininthe
thefrequency
frequency
band and the standard deviation of the wavelet transform. This vector was then used to generatea
band and the standard deviation of the wavelet transform. This vector was then used to generate
asupport vectormachine
support vector machine(SVM) (SVM) forfor pattern
pattern recognition,
recognition, which which was classified
was classified in threeincategories:
three categories:
stable,
stable, transition, and chatter. Zagorski et al. [23] employed two NNs
transition, and chatter. Zagorski et al. [23] employed two NNs for chatter detection in milling: for chatter detection in milling:
radial
radial
basis basis function
function (RBF) and (RBF) and multi-layered
multi-layered perceptronperceptron (MLP). and
(MLP). Kumar Kumar and[24]
Singh Singh
used[24]
anused
ANNanbasedANN
onbased
feed on feed forward
forward backpropagation
backpropagation to predicttostability
predictinstability
turning,inalso
turning, alsoonfocusing
focusing the metal onremoving
the metal
removing index. The authors applied a tangent sigmoid activation
index. The authors applied a tangent sigmoid activation function for model training. function for model training.
InInthis
thispaper,
paper,ananANN ANNisisused
usedtotomodel
modelstability
stabilitybehavior
behaviorininturning,
turning,where
wherethe thephysics-based
physics-based
analyticalstability
analytical stabilitylimit
limitisisapplied
appliedtotogenerate
generateaadata datasetsetthat
thattrains
trainsthe
theANN.
ANN.The Themotivation
motivationfor forthis
this
effort is that the ANN model inputs are the spindle speed and depth of cut,
effort is that the ANN model inputs are the spindle speed and depth of cut, while the analytical stability while the analytical
stability
limit model limit model
inputs areinputs are the structural
the structural dynamicsdynamics
and forceand forceWhen
model. model. When experimental
experimental stability stability
data is
data is collected at a selected spindle speed depth of cut pair, a
collected at a selected spindle speed depth of cut pair, a model which can accept this data model which can accept this data
directly is
directly is preferred. The ANN enables this
preferred. The ANN enables this convenient model updating. convenient model updating.

2.2.Background
Background
Duringturning,
During turning,aasharp
sharpcutting
cuttingedge
edgeisisused
usedtotoremove
removematerial
materialininthe theform
formofofa achip.
chip.Figure
Figure11
shows an orthogonal cutting operation, where only the normal, F
shows an orthogonal cutting operation, where only the normal, Fn , and tangential, Ft , components of theof
n , and tangential, F t, components
the resultant
resultant force, force, F, are considered
F, are considered [25]. In the
[25]. In general, general,
cuttingthe cutting
force vectorforce vector
includes theincludes the third
third component
component
along along the
the workpiece workpiece
rotation axis, rotation axis, but the
but the orthogonal orthogonal
(planar) treatment(planar) treatmenttoisdescribe
is sufficient sufficienttheto
describe the process dynamics. The figure also identifies: (1) The mean
process dynamics. The figure also identifies: (1) The mean chip thickness, hm , or commanded feed perchip thickness, h m , or
commanded feed per revolution for the facing operation pictured, and (2) the
revolution for the facing operation pictured, and (2) the force angle, β, between F and Fn . The sideforce angle, β, between
F and
view ofFthis
n. The side view
operation of this
(inset operation
in Figure (inset inthe
1) identifies Figure
chip1)width,
identifies the chipthe
b. Together, width,
chipb.thickness
Together,and the
chip thickness and chip width define the area of material
chip width define the area of material to be removed, A = bhm . to be removed, A = bh m .

Figure 1. Orthogonal cutting operation showing the cutting force with its normal and tangential
Figure 1. Orthogonal
components cutting
[25]. Adapted withoperation showing
permission the cutting
from Springer force
Nature, with its Dynamics:
Machining normal andFrequency
tangential
components [25]. Adapted with permission from Springer Nature, Machining
Response to Improved Productivity by Schmitz, T.L. and Smith, K.S., 2019. Dynamics: Frequency
Response to Improved Productivity by Schmitz, T.L. and Smith, K.S., 2019.

The cutting force can be approximated as the product of the chip area and the process dependent
specific (per unit area) force coefficient, Ks [26]. It depends on the workpiece material, tool geometry,
and, to a lesser extent, the cutting speed (peripheral velocity of the rotating workpiece) and chip
thickness.
𝐹 𝐾𝐴 𝐾 𝑏ℎ (1)
J. Manuf. Mater. Process. 2019, 3, 45 4 of 15

The cutting force can be approximated as the product of the chip area and the process dependent
specific (per unit area) force coefficient, Ks [26]. It depends on the workpiece material, tool geometry, and,
to a lesser extent, the cutting speed (peripheral velocity of the rotating workpiece) and chip thickness.
J. Manuf. Mater. Process. 2019, 3, 45 4 of 15

F = Ks A = Ks bhm (1)
The normal and tangential components, Fn and Ft, can be expressed using F and the force angle:
The normal and tangential components, Fn and Ft , can be expressed using F and the force angle:
𝐹 cos 𝛽 𝐹 cos 𝛽 𝐾 𝑏ℎ 𝑘 𝑏ℎ (2)
Fn = cos βF = cos βKs bhm = kn bhm (2)
and
and 𝐹 sin 𝛽 𝐹 sin 𝛽 𝐾 𝑏ℎ 𝑘 𝑏ℎ , (3)
Ft = sin βF = sin βKs bhm = kt bhm , (3)
where the cutting force coefficients, kn and kt, are introduced which incorporate both Ks and β. A
where the
common cutting used
approach characterizekthese
forcetocoefficients, kt , are dependent
n andprocess introducedvalueswhichis incorporate
to prescribe known and β.
both Kscutting
A common approach used to characterize these process dependent values is
conditions and measure the force components directly [26]. As an alternative, the material behavior to prescribe known cutting
conditions
can be defined andusing
measure the force components
a constitutive directly [26]. Asand
model (e.g., Johnson–Cook) an alternative, the material
the cutting force behavior
predicted using
can be defined using
finite element simulations. a constitutive model (e.g., Johnson–Cook) and the cutting force predicted using
finiteThe
element simulations.
cutting force causes deflections of the cutting tool. Because the tool has stiffness and mass,
it canThe cutting
vibrate. force
If the causes
tool deflections
is vibrating as itofremoves
the cutting tool. Because
material, the tool has
these vibrations are stiffness
imprintedandonmass,
the
it can vibrate. If the tool is vibrating as it removes material, these vibrations
workpiece surface as a wavy profile. Figure 2 shows an exaggerated view, where the initial impact are imprinted on the
workpiece
with surface as surface
the workpiece a wavy profile. Figure
causes the tool2 shows
to begin an vibrating,
exaggerated view,
and thewhere the initial
oscillations impact
in the with
normal
the workpiece
direction to besurface
copied causes
onto thethe workpiece.
tool to beginWhenvibrating, and the oscillations
the workpiece begins itsinsecond
the normal direction
revolution, theto
be copied onto the workpiece. When the workpiece begins its second revolution,
vibrating tool encounters the wavy surface produced during the first revolution. Therefore, the chip the vibrating tool
encounters
thickness atthe
anywavy surface
instant produced
depends during
both on the first
the tool revolution.
deflection Therefore,
at that time andthe thechip thickness
workpiece at any
surface
instant depends both on the tool deflection at that time and the workpiece
from the previous revolution(s). Vibration of the tool therefore leads to a variable chip thickness surface from the previous
revolution(s).
which, according Vibration of the
to Equation tool therefore
1, yields a variable leads to a force
cutting variable
sincechip
the thickness which, according
force is proportional to the
to Equation
chip thickness. (1),The
yields a variable
cutting cutting the
force governs force since the
current toolforce is proportional
deflection to the chipthe
and, subsequently, thickness.
system
The cutting
exhibits force governs the current tool deflection and, subsequently, the system exhibits feedback.
feedback.

Figure
Figure 2.
2. Description
Descriptionof ofregenerative
regenerativechatter
chatterin in turning.
turning. Initial
Initialtool
tooldeflections
deflectionsare
arecopied
copiedonto
ontothe
the
workpiecesurface
workpiece surfaceand
andareareencountered
encountered in in subsequent
subsequent revolutions.
revolutions. This
Thisvaries
variesthe
thechip
chipthickness
thicknessand
and
cuttingforce
cutting forcewhich,
which,in inturn,
turn,affects
affects the
the resulting
resulting tool
tool deflections
deflections [25].
[25]. Adapted
Adaptedwith
withpermission
permissionfrom
from
SpringerNature,
Springer Nature,Machining
MachiningDynamics:
Dynamics:Frequency
FrequencyResponse
ResponsetotoImproved
ImprovedProductivity
Productivityby
bySchmitz,
Schmitz,
T.L. and
T.L. and Smith,
Smith, K.S.,
K.S., 2019.
2019.

From a modeling standpoint, this “regeneration of waviness” appears as a time-delayed term in

the chip thickness equation. Figure 3 shows an unwrapped view of the turning operation, where the
surface on the left was produced in the previous revolution and the surface to the right of the tool
(offset by the mean feed per revolution) was just cut away by the oscillating tool. Only the vibrations
in the normal direction, y (positive direction out of the cut), are considered here because they have
the most direct influence on the chip thickness.
J. Manuf. Mater. Process. 2019, 3, 45 5 of 15

From a modeling standpoint, this “regeneration of waviness” appears as a time-delayed term in

the chip thickness equation. Figure 3 shows an unwrapped view of the turning operation, where the
surface on the left was produced in the previous revolution and the surface to the right of the tool
(offset by the mean feed per revolution) was just cut away by the oscillating tool. Only the vibrations
J. Manuf. Mater. Process. 2019, 3, 45 5 of 15
J.in the normal
Manuf. Mater. Process. 2019, 3,y 45
direction, (positive direction out of the cut), are considered here because they have 5 of the
15
mostThe direct influence on the chip thickness.
time dependent, instantaneous chip thickness, h(t), is determined using Equation 4. It is seen
The time dependent, instantaneous chip thickness, h(t), is determined
h(t), using Equation 4. It is seen
that aThe timepositive
larger dependent, instantaneous
vibration during thechip thickness,
previous revolution,is determined
y(t – τ), whereusing
τ isEquation
the time (4). It is
for one
that a larger
seen that gives
rotation,
positive
a larger vibration
an positive
increased
during
vibration the previous
during the
chip thickness
revolution,
(i.e.,previous
y(t
revolution,
less material
– τ ),
was removed
where τ
y(t − τ), where
so the τcurrent
is the time for
is the time one
chipforis
rotation,
one givesgives
rotation, an increased
an increased chipchip
thickness
thickness(i.e.,(i.e.,
lessless
material
material was
wasremoved
removed sosothe
the current
current chip
chipisis
thicker). A larger positive current vibration, y(t), on the other hand, yields a thinner chip, see Figure
thicker).
thicker). AAlarger
larger positive
positive current
current vibration, y(t), on on the
the other
other hand,
hand,yields
yieldsaathinner
thinnerchip,
chip,see
seeFigure
Figure4.
4.
4.
ℎh(𝑡t) = ℎhm + 𝑦y(𝑡t − τ𝜏) − y𝑦(t𝑡) (4)
(4)
ℎ 𝑡 ℎ 𝑦 𝑡 𝜏 𝑦 𝑡 (4)

Figure
Figure 3.
3. Depiction
Depictionofofturning
turningwhere
wherethethesurface
surfacefrom
fromthe
theprevious revolution,
previous revolution,shown
shown to to
thethe
leftleft
of
Figure 3. Depiction of turning where the surface from the previous revolution, shown to the left of
the tool, is removed by the vibrating cutter to produce a new wavy surface to the right of
of the tool, is removed by the vibrating cutter to produce a new wavy surface to the right of the the tool [25].
the tool, is removed by the vibrating cutter to produce a new wavy surface to the right of the tool [25].
Adapted with permission
tool [25]. Adapted from Springer
with permission Nature,
from Springer Machining
Nature, Dynamics:
Machining Dynamics:Frequency
FrequencyResponse
Responseto
Adapted with permission from Springer Nature, Machining Dynamics: Frequency Response to
Improved
to Improved Productivity
Productivitybyby
Schmitz, T.L.
Schmitz, and
T.L. andSmith, K.S.,
Smith, 2019.
K.S., 2019.
Improved Productivity by Schmitz, T.L. and Smith, K.S., 2019.

Figure 4. The
The figure demonstrates the instantaneous chip thickness calculation. It It depends
depends on the
Figure 4. The figure demonstrates the instantaneous chip thickness calculation. It depends on the
mean feed per revolution, the current deflection, and the vibration during the previous revolution of
mean feed per revolution, the current deflection, and the vibration during the previous revolution of
the workpiece (to the left of the
the tool)
tool) [25].
[25]. Adapted
Adaptedwith
withpermission
permission from
from Springer
Springer Nature,
Nature, Machining
Machining
the workpiece (to the left of the tool) [25]. Adapted with permission from Springer Nature, Machining
Dynamics: Frequency
Frequency Response
Response to Improved Productivity by Schmitz, T.L. and Smith,
Smith, K.S.,
K.S., 2019.
2019.
Dynamics: Frequency Response to Improved Productivity by Schmitz, T.L. and Smith, K.S., 2019.
The relative phasing between the surface waviness from one pass to the next determines the
The relative phasing between the surface waviness from one pass to the next determines the
level of force variation and whether the operation is stable or unstable (chatter occurs). Figures 5 and
level of force variation and whether the operation is stable or unstable (chatter occurs). Figures 5 and
6 Figure 5; Figure 6 show two possibilities. In Figure 5, the wavy surfaces between two revolutions are in phase.
6 Figure 5; Figure 6 show two possibilities. In Figure 5, the wavy surfaces between two revolutions are in phase.
Therefore, even though vibration is present during material removal, the chip thickness variation
Therefore, even though vibration is present during material removal, the chip thickness variation
(vertical distance between the two curves) is negligible and there is no appreciable force variation.
(vertical distance between the two curves) is negligible and there is no appreciable force variation.
This enables stable cutting at larger chip widths. Considering that the tool tends to vibrate at its
This enables stable cutting at larger chip widths. Considering that the tool tends to vibrate at its
natural frequency, it is intuitive that matching the workpiece rotating frequency (spindle speed) to
J. Manuf. Mater. Process. 2019, 3, 45 6 of 15

The relative phasing between the surface waviness from one pass to the next determines the
level of force variation and whether the operation is stable or unstable (chatter occurs). Figures 5
and 6 show two possibilities. In Figure 5, the wavy surfaces between two revolutions are in phase.
Therefore, even though vibration is present during material removal, the chip thickness variation
(vertical distance between the two curves) is negligible and there is no appreciable force variation.
This enables stable cutting at larger chip widths. Considering that the tool tends to vibrate at its natural
frequency, it is intuitive that matching the workpiece rotating frequency (spindle speed) to the tool’s
natural frequency will lead to this preferred “in phase” situation. However, this is counter-intuitive
based on a traditional understanding of resonance, where driving the system at its natural frequency is
typically avoided because the vibration amplitude is large. While it is understood that the vibration
magnitude is larger, it has been demonstrated through both theory and experiment that increased
material removal rate is beneficial. In milling, a phenomenon referred to as surface location error
(SLE) has been studied, where the final location of the machined surface depends on the location of
the tool in its vibration cycle when it is leaving the surface. It has been shown that SLE is large at
resonance and sensitive to changes in the system dynamics (e.g., natural frequency) [25]. Figure 6
shows a less favorable phase relationship where there is significant variation in the chip thickness.
This leads to unstable cutting at smaller chip widths than the previous case due to the force variations
and subsequent tool deflections.
Depending on the feedback system “gain”, or chip width b, and spindle speed, Ω, the turning
operation will either be stable or exhibit chatter, which causes large vibrations and forces and leads
to poor surface finish and, potentially, tool/workpiece damage. In stable machining, the vibrations
diminish from revolution to revolution. In unstable machining, the vibrations grow from revolution to
revolution until limited in some way. Surprisingly, the vibrations may become large enough that the
tool jumps out of the cut, losing contact with the workpiece. The vibrations in unstable cutting may
be at least as large as the chip thickness and it is not surprising that these large vibrations may result
in damage to the machine, tool, and workpiece. The governing relationships for this behavior are
provided in Equations (5)–(7) [25]. In these equations, blim is the limiting chip width to avoid chatter,
fc is the chatter frequency (should it occur), FRF is the frequency response function that describes the
tool’s dynamic response, N is the integer number of waves of vibration imprinted on the workpiece
ε
surface in one revolution, and 2π is any additional fraction of a wave, where ε is the phase (in rad)
between current and previous tool vibrations.

−1
blim = (5)
2Ks cos βRe[FRF]

fc ε
=N+ (6)
Ω 2π
!
Re[FRF]
ε = 2π − 2 tan−1 (7)
Im[FRF]
The deterministic stability model described in the previous paragraphs inherently includes
uncertainty [27–30]. For example, the actual turning tool clamping conditions may vary from one
setup to the next. This results in uncertainty in the FRF which, in turn, leads to uncertainty in the
stability limit. Propagation of input to output uncertainties may be completed using Monte Carlo
simulation, for example [27]. This output uncertainty motivates the ANN approach presented in the
next section. This model is defined in the desired test domain of spindle speed-chip width so that
experimental stability results may be collected and the ANN stability model may be updated.
 𝑓 𝜀
𝑁 (6)
 2𝜋

𝑅𝑒 𝐹𝑅𝐹
J. Manuf. Mater. Process. 2019, 3, 45 𝜀 2𝜋 2 tan (7)15
7 of
𝐼𝑚 𝐹𝑅𝐹

Figure 5. The surface waviness between revolutions is in phase. Negligible chip thickness variation
Figure 5. The surface waviness between revolutions is in phase. Negligible chip thickness variation is
is obtained [25]. Adapted with permission from Springer Nature, Machining Dynamics: Frequency
obtained [25]. Adapted with permission from Springer Nature, Machining Dynamics: Frequency
J. Manuf. Mater. Process.
Response 2019, 3, Productivity
to Improved 45 by Schmitz, T.L. and Smith, K.S., 2019. 7 of 15
Response to Improved Productivity by Schmitz, T.L. and Smith, K.S., 2019.

Figure 6. Less favorable phase relationship between revolutions yields significant chip thickness
Figure 6. Less favorable phase relationship between revolutions yields significant chip thickness
variation [25]. Adapted with permission from Springer Nature, Machining Dynamics: Frequency
variation [25]. Adapted with permission from Springer Nature, Machining Dynamics: Frequency
Response to Improved Productivity by Schmitz, T.L. and Smith, K.S., 2019.
Response to Improved Productivity by Schmitz, T.L. and Smith, K.S., 2019.
3. Artificial Neural Networks
The deterministic stability model described in the previous paragraphs inherently includes
The machine
uncertainty [27–30]. learning approachthe
For example, applied
actualhere for chatter
turning predictionconditions
tool clamping follows the may
supervised
vary fromlearning
one
model, where the learning algorithm uses known input-output pairs for training.
setup to the next. This results in uncertainty in the FRF which, in turn, leads to uncertainty in the Once trained,
the model
stability can Propagation
limit. be used to predict outputs
of input for new
to output input data.
uncertainties mayWhen the outputusing
be completed (typically
Monte discrete
Carlo
values)
simulation, for example [27]. This output uncertainty motivates the ANN approach presented in the
is used to create categories or classes, the problem is called a classification problem. When the
output is a real,
next section. Thiscontinuous value (or
model is defined in values),
the desiredit istest
a regression
domain ofproblem. Since chatter
spindle speed-chip prediction
width so that
involves
experimentalpredicting
stability whether
resultsa may
givenbeset of inputand
collected variables
the ANN (spindle speed,
stability Ω, and
model maylimiting chip width,
be updated.
blim ) leads to chatter or not, a binary classification problem is to be solved. It is also supervised
since the prediction
3. Artificial is based on pairs of values (Ω, blim ) for which the stability is known a priori.
Neural Networks
Furthermore, the model developed in this paper applies an ANN. An overview of ANNs is presented
in theThe machine
following learning approach applied here for chatter prediction follows the supervised
paragraphs.
learning model, where the learning
ANNs consist of a collection algorithm
of basic uses neurons
units called known input-output pairs(Figure
arranged in layers for training.
7). TheOncefirst
trained, the model can be used to predict outputs for new input data. When
(left) layer is the input layer and the last (right) layer is the output layer. The layers in between the output (typically
are
discrete values) is used to create categories or classes, the problem is called a
hidden layers. A neural network can have zero or more hidden layers. In a feedforward neural network,classification problem.
When
the the output
neurons in oneislayera real,
arecontinuous
connected to value
the (or values),
neurons it isnext
in the a regression
layer and problem. Since chatter
the information flows
prediction involves predicting whether a given set of input variables
forward from the input to the output through the hidden layers. When there are many hidden (spindle speed, Ω, and limiting
layers,
chipnetwork
the width, isblimcalled
) leads to chatter
a deep neuralornetwork
not, a binary
(DNN).classification
The connections problem is tothe
between beneurons
solved. are
It is also
called
supervised since the prediction is based on pairs of values (Ω, blim) for which the stability is known a
priori. Furthermore, the model developed in this paper applies an ANN. An overview of ANNs is
presented in the following paragraphs.
ANNs consist of a collection of basic units called neurons arranged in layers (Figure 7). The first
(left) layer is the input layer and the last (right) layer is the output layer. The layers in between are
J. Manuf. Mater. Process. 2019, 3, 45 8 of 15

synapses. A neuron, the basic building block of ANNs, consists of a set of input values, xi , i = 1, . . . , n,
a set of weights, wi , i = 1, . . . , n, and a transfer (or activation) function, f (see Figure 8). A linear
P
transformation consisting of the weighted sum of all the inputs, wi xi , and a bias, b, is calculated as:
n
X
z = b+ wi xi (8)
i=1
J. Manuf. Mater. Process. 2019, 3, 45 8 of 15
J. Manuf. Mater. Process. 2019, 3, 45 8 of 15
for each neuron. The output, h, is calculated from this neuron through the (usually) nonlinear transfer
neuron, i, fin
function, (z).that
Typically,
layer, theeach neuron
output in a givenaslayer
is calculated has
hi = f(zi), the same
where is ztransfer function
i calculated usingand for each
Equation 8.
neuron, i, in that layer, the output is calculated as hi = f(zi), where is zi calculated using Equation 8.
neuron, i, in serve
The outputs that layer,
as thethe output
inputs foriseachcalculated as hi = f in
of the neurons (z ),
thewhere is zi calculated
next layer, which can using
use aEquation
different(8).
or
The outputs serve as the inputs for each of the neurons inithe next layer, which can use a different or
The outputs
the same serve as
transfer the inputs
function. for process
This each of the neurons in until
is continued the nextthelayer,
output which can
layer isuse a different
reached whereor the
the same transfer function. This process is continued until the output layer is reached where the
neurons compute the output variables, yyii,,ii ==1, 1,… , p the
neurons
same compute
transfer the output
function. variables,
This process is continued until (p is the number
output of outputs).
layer is reached whereFor
theaneurons
binary
…, p (p is the number of outputs). For a binary
compute the output variables, y , i = 1, . . . , p (p is the number of outputs).
classifier, there is usually only one neuron in the output layer and therefore p is taken to be unity.
i For a binary classifier,
classifier,
there there isonly
is usually usually
one only one
neuron neuron in the output
and layer and ptherefore p is taken to be unity.
However, two neurons can also beinusedthe output
for binary layer therefore
classification [23]. is taken to be unity. However,
However,
two neurons twocan neurons
also becan also
used forbebinary
used for binary classification
classification [23]. [23].

Figure 7. An example
Anexample
exampleofofofananartificial neural
artificial network
neural (ANN).
network This ANN
(ANN). has fourfour
inputs (features), one
Figure
Figure 7. An
7. an artificial neural network (ANN). ThisThis
ANN ANN
has has inputs
four inputs (features),
(features), one
hidden
one layer withwith
hidden three neurons, and twotwooutputs.
hidden layerlayer
with threethree neurons,
neurons, and and outputs.
two outputs.

Figure 8. AAsingle
singleneuron
neuron consists of of
consists thethe
inputs xi, the
inputs x , weights wi, and
the weights w ,the
andtransfer function
the transfer f(z), which
function f (z),
Figure 8. A single neuron consists of the inputs xi, thei weights wi, andi the transfer function f(z), which
produces
which the scalar
produces value value
the scalar h. Note
h. that
Notezthat
is defined with with
z is defined the bias b absorbed
the bias intointo
b absorbed the the
summation
summation by
produces the scalar value h. Note that z is defined with the bias b absorbed into the summation by
setting
by x0 =x10 =
setting and w0 =wb.0 = b.
1 and
setting x0 = 1 and w0 = b.
In supervised learning, the training data (input data and the corresponding output data) is used
In supervised learning, the training data (input data and the corresponding output data) is used
to train the ANN model. The training starts with an initial assumption
assumption on
on the weights w
the weights wii. The input
to train the ANN model. The training starts with an initial assumption on the weights wi. The input
data is processed by the ANN and output is predicted. The error between the predicted outputs and
data is processed by the ANN and output is predicted. The error between the predicted outputs and
the known outputs is calculated using a cost (or loss) function which can be the sum of the squares
the known outputs is calculated using a cost (or loss) function which can be the sum of the squares
of the errors between predicted and observed outputs, for example. Since the predicted values
of the errors between predicted and observed outputs, for example. Since the predicted values
depend on the weights and biases, it is clear that the loss function, E, is also a function of the weights
depend on the weights and biases, it is clear that the loss function, E, is also a function of the weights
and biases for a given set of training data, i.e., E = E ( wi , b) . By absorbing the bias b into the weights
and biases for a given set of training data, i.e., E = E ( w , b) . By absorbing the bias b into the weights
J. Manuf. Mater. Process. 2019, 3, 45 9 of 15

data is processed by the ANN and output is predicted. The error between the predicted outputs and
the known outputs is calculated using a cost (or loss) function which can be the sum of the squares of
the errors between predicted and observed outputs, for example. Since the predicted values depend
on the weights and biases, it is clear that the loss function, E, is also a function of the weights and
biases for a given set of training data, i.e., E = E(wi , b). By absorbing the bias b into the weights as
an additional parameter, E can be assumed to be a function of only the weights wi . If the error is not
acceptable, the weights are updated through various methods. One approach is the gradient descent
method, where the weight updates are computed using the derivatives of the error function with
respect to the weights:
∂E  j

j+1 j
wi = wi − η (9)
∂wi 


In Equation (9), the superscript j indicates the jth iteration and η is the learning rate that is used to
control the magnitudes of the corrections applied to wi . Too large a value of η will lead to convergence
issues and too small a value increases the computational time and cost. The updated weights are
again used for predictions and calculating the error in predictions. This process is repeated until the
error is less than a preselected value or a maximum number of iterations has been reached. Although
Equation (9) captures the essence of weight updates, in a typical ANN with multiple hidden layers,
the gradient calculation is quite complicated and involved. The backpropagation algorithm may
be used to compute these gradients. In the standard backpropagation algorithm, the learning rate
is kept constant. A modification of this algorithm, called the resilient backpropagation algorithm,
uses separate learning rates for each weight and, in addition, these rates can be altered during the
training process to accelerate the convergence. Furthermore, the adjustments to weights do not include
the partial derivatives of the error function with respect to weights. Instead, only the signs of the
derivatives are used in place of the derivatives [31]. In the present work, the resilient backpropagation
algorithm was used for updating the weights.
When the training is complete, the performance of the model is evaluated using test data with
known outputs. Additional cross-validation methods are also used to further evaluate the model
performance. If the predictions from the test data and cross-validations are satisfactory, the ANN
model is used for predictive purposes on new sets of input data.

4. ANN Model for Chatter Prediction

In this work the R neural net package, neuralnet [31], was used to build the ANN model. The input
parameters for the neural network were the spindle speed and limiting chip width, see Figure 9.
The data set for training and testing was obtained from the stability algorithm described in Section 2.
The corresponding model parameters were: k = 1 × 106 N/m, c = 315 N-s/m, m = 2.5 kg, Ks = 700 N/mm2 ,
and β = 70 deg. The data set was generated by considering uniformly distributed random values for the
pairs (Ω, blim ) in the range of 1000 rpm to 4000 rpm and 0 mm to 2.5 mm. For each set of values, the cut
was labeled as stable or unstable (chatter) using the stability limit. The total number of points in the
training set was varied to evaluate the predictive capability on the ANN model. The training and test
sets were first rescaled using the max-min normalization method. The value, x, for each of the inputs
was mapped to the range [0, 1] through the transformation (x − xmin )/(xmax − xmin ). Since there were
no outliers in the input data, this rescaling method was acceptable. The normalized training data was
used to train the ANN model. The non-normalized training set consisting of 2001 randomly generated
points is shown in Figure 9a. The activation function for the neural network considered in the present
work is taken to be the binary logistic function with output between 0 and 1. Output less than 0.5 is
taken to be stable and greater than or equal to 0.5 is taken to indicate chatter. The cross entropy loss
function is used for E [32]. Furthermore, two hidden layers were considered with six neurons in each
layer. The trained neural network with the corresponding weights is shown in Figure 9b.
J. Manuf. Mater. Process. 2019, 3, 45 10 of 15

J. Manuf. Mater. Process. 2019, 3, 45 10 of 15

(a)

(b)
Figure 9.9. (a)
(a)The
Thetraining
training
setset consisting
consisting of 2001
of 2001 points
points was to
was used used to (b)
train, train, (b)hidden
a two a two layer
hidden layer
artificial
artificial
neural neural network.
network.

The actual
actual decision
decisionboundary
boundaryand
andthe
thepredicted decision
predicted boundaries
decision areare
boundaries shown Figure
shown 10. As
Figure 10. the
As
figures indicate,
the figures thethe
indicate, decision boundary
decision boundaryis is
reasonably reproduced
reasonably reproducedbybythe
theANN
ANNmodel.
model. However,
for the points near the lobe peaks and some
some troughs,
troughs, the
the predicted
predicted boundary
boundary is not
not accurate.
accurate. It is to
be expected then, that when the model is used for predictive purposes, the predictions may not be
J. Manuf. Mater. Process. 2019, 3, 45 11 of 15
J. Manuf. Mater. Process. 2019, 3, 45 11 of 15
J. Manuf. Mater. Process. 2019, 3, 45 11 of 15
be expected then, that when the model is used for predictive purposes, the predictions may not be
be expected then, that when the model is used for predictive purposes, the predictions may not be
accurate when the input data is near these peaks and troughs. However, away from these locations,
accurate when the input data is near these peaks and troughs. However, away from these locations,
the accuracy
accurate of the
when predictions
data can be expected to be high. This isHowever,
confirmed by evaluating
thesethe ANN
the accuracy ofthe
theinput
predictionsis can
nearbethese peaks
expected toand troughs.
be high. This is confirmedaway byfrom
evaluating locations,
the ANN
model
the using a of
accuracy test
theset. The predictions
predictions can beby the ANN
expected model
to be high.onThis
the is
test set are shown
confirmed in Figurethe11. ANN
The
model using a test set. The predictions by the ANN model on the test set are by evaluating
shown in Figure 11. The
test set
model consists
using a of 501
test set.points
The randomly
predictions distributed
by the ANN in the
model feature
on space,
the test see
set Figure
are 11a.
shown The
in ANN
Figure 11.
test set consists of 501 points randomly distributed in the feature space, see Figure 11a. The ANN
model
The predictions
test set consistsareofshown
501 as arandomly
points confusiondistributed
matrix in Figure
in the 11b. Only
feature three
space, seepoints
Figure out ofThe
11a. 501ANN
are
model predictions are shown as a confusion matrix in Figure 11b. Only three points out of 501 are
misidentified
model with two
predictions false negatives and onematrix
false positive (with
11b.respect to chatter).
pointsThe accuracy
misidentified witharetwoshown as a confusion
false negatives and one falseinpositive
Figure (with Only three
respect to chatter). out
Theofaccuracy
501 are
ofmisidentified
the model is with
498/501 = 99.4%.
two false negatives and one false positive (with respect to chatter). The accuracy of
of the model is 498/501 = 99.4%.
the model is 498/501 = 99.4%.

Chatter
Chatter

Stable
Stable

(a) (b)
(a) (b)
Figure 10. The decision boundary separates the stable region from the chatter region. (a) The actual
Figure10.
Figure 10.The
Thedecision
decisionboundary
boundaryseparates
separatesthe
thestable
stableregion
regionfrom
fromthe
thechatter
chatterregion.
region.(a)
(a)The
Theactual
actual
boundary, and (b) the predicted decision boundary.
boundary,and
boundary, and(b)
(b)the
thepredicted
predicteddecision
decisionboundary.
boundary.

Predicted
Predicted

Stable Chatter
Stable Chatter
blim (mm)
blim (mm)

Stable 215 2 217

Stable 215 2 217
Actual
Actual

Chatter 1 283 284

Chatter 1 283 284

Ω (rpm)
Ω (rpm) 216 285
216 285
(a) (b)
(a) (b)
Figure11.
Figure (a)The
11.(a) Thetest
testdata
dataset
setconsisting
consistingofof501
501points,
points,and
and(b)
(b)the
thepredictions
predictionsbybythe
thetrained
trainedneural
neural
Figure 11. (a) The test data set consisting of 501 points, and (b) the predictions by the trained neural
network on this data (shown as a confusion
network on this data (shown as a confusion matrix). matrix).
network on this data (shown as a confusion matrix).
ANNmodels
ANN modelstrained
trainedwith
witha afewer
fewernumber
numberofofpoints
pointsororwith
witha asingle
singlehidden
hiddenlayerlayerwere
werefound
found
ANN models trained with a fewer number of points or with a single hidden layer were found
totobebeless
lessaccurate.
accurate.InInFigure
Figure12a,
12a,a asingle
singlehidden
hiddenlayer
layerneural
neuralnetwork
networkwith
withfour
fourneurons
neuronsisisshown.
shown.
to be less
Thedecision accurate.
decisionboundaries In Figure
boundariespredicted 12a, a single
predictedbybythis hidden
thismodel
modelforlayer neural
fordifferent network
differentnumbers with four
numbersofoftotal neurons
totaldata
datapoints is shown.
pointsininthe
the
The
The decision
training set areboundaries
shown in predicted
Figure 12b. by this model
Clearly, the for different
predicted numbers
decision of total
boundaries data
differ frompoints
the in the
actual
training set are shown in Figure 12b. Clearly, the predicted decision boundaries differ from the actual
training set
boundary are many
with shown intheFigure
lobes12b. Clearly, the predicted decision boundaries differ from the actual
boundary with many ofofthe lobes missing
missing altogether.
altogether. Therefore,
Therefore, predictions
predictions forinput
for inputdata
datanear
near the
the
boundary
actual with
decision many
boundary of the
willlobes
be missing
inaccurate. altogether. Therefore, predictions for input data near the
actual decision boundary will be inaccurate.
actual decision boundary will be inaccurate.
J. Manuf. Mater. Process. 2019, 3, 45 12 of 15
J. Manuf. Mater. Process. 2019, 3, 45 12 of 15

(a)

(b)

12. (a) One hidden layer neural

Figure 12. neural network
network with
with four
four neurons.
neurons. (b) The decision
decision boundaries
boundaries
predicted by the neural network. The decision boundaries
boundaries in the top row correspond to 201 and 601
points in the
the training
training set. The bottom
set. The bottom two
two correspond
correspond to 2001 points each, but with two different
random distributions.

To
To further
further examine
examine the
the sensitivity
sensitivity of
of the
the ANN performance to
ANN performance its architecture,
to its architecture, seven
seven ANNs were
ANNs were
evaluated of hidden layers, HL, and neurons, NE, where HL
evaluated with different numbers of hidden layers, HL, and neurons, NE, where HL was varied from1
with different numbers was varied from
1 to 3 and NE from 4 to 8. The decision boundary for each architecture was calculated by considering
 J. Manuf. Mater. Process. 2019, 3, 45 13 of 15

J. Manuf. Mater. Process. 2019, 3, 45 13 of 15

to 3 and NE from 4 to 8. The decision boundary for each architecture was calculated by considering
an an evenly
evenly spaced
spaced gridgrid in the
in the blim)bdomain.
(Ω, (Ω, lim ) domain.
This This
domaindomain was divided
was divided into ainto
fineagrid
fineconsisting
grid consisting
of
of equispaced
1001 1001 equispacedpointspoints in direction
in each each direction for a of
for a total total of 1,002,001
1,002,001 points.
points. The results
The results are presented
are presented in
in Figure 13 in the form of various decision boundaries obtained by the predictions
Figure 13 in the form of various decision boundaries obtained by the predictions of the trained ANN of the trained
ANNon
models models
theseon these points.
points. Visual Visual comparison
comparison to the totrue
the boundary
true boundary demonstrates
demonstrates thatthat
the the
modelmodel
architecture
architecture is closely
is closely related
related to the
to the modelmodel accuracy.
accuracy. Of Of
thethe seven
seven architectures
architectures considered,
considered, thethetwotwo
hidden layer ANN with six neurons per layer best replicates the true stability
hidden layer ANN with six neurons per layer best replicates the true stability boundary. boundary. Interestingly,
the other two
Interestingly, thearchitectures with two hidden
other two architectures withlayers either did
two hidden not converge
layers either didin not
the specified
convergenumber
in the of
epochs or required adjustments to the stopping criterion. As expected, the
specified number of epochs or required adjustments to the stopping criterion. As expected, the training time increased
with increasing
training complexity
time increased with of the ANNcomplexity
increasing architecture.ofThe
thetwoANN andarchitecture.
three hiddenThelayertwo
models
and required
three
one to
hidden three
layer hours required
models using a standard laptop
one to three with
hours 8 GBaRAM.
using standard laptop with 8 GB RAM.

Figure 13. Results for study of ANN sensitivity to architecture.

Figure 13. Results for study of ANN sensitivity to architecture.
5. Conclusions
5. Conclusions
This paper provided an analysis of the application of artificial neural networks (ANN) to modeling
stabilitypaper
This provided
behavior an analysis
in turning. of the application
The physics-based, analyticalof stability
artificiallimit
neural
wasnetworks (ANN)ato
used to generate data
modeling stability behavior in turning. The physics-based, analytical stability limit
set that trained the ANN. It was observed that the number and distribution of training points was used to
influence
generate a data set that trained the ANN. It was observed that the number and distribution of training
points influence the ability of the ANN model to capture the smaller, more closely spaced lobes that
occur at lower spindle speeds. Overall, the ANN is successful (>90% accuracy) at predicting the
J. Manuf. Mater. Process. 2019, 3, 45 14 of 15

the ability of the ANN model to capture the smaller, more closely spaced lobes that occur at lower
spindle speeds. Overall, the ANN is successful (>90% accuracy) at predicting the stability behavior
after appropriate training. Deep and narrow ANNs were found to be more accurate than shallow and
wide ANNs. A sensitivity study showed that the ANN performance is closely linked to the selected
number of hidden layers and neurons per layer. Future work will include incorporating experimental
data into the ANN models to augment the data obtained from the physics-based models.

Author Contributions: The individual author contributions follow: conceptualization, T.S.; methodology, H.C.,
E.P.-B., and M.S.; software, T.S. and H.C.; validation, H.C. and T.S.; formal analysis, T.S.; investigation, T.S., H.C.,
E.P.-B., and M.S.; resources, T.S.; data curation, T.S. and H.C.; writing—original draft preparation, T.S., H.C., E.P.-B.,
and M.S.; writing—review and editing, T.S.; visualization, T.S. and H.C.; supervision, T.S.; project administration,
T.S.; funding acquisition, T.S., E.P.-B., and M.S.
Funding: The authors gratefully acknowledge financial support from the UNC ROI program. Elena Perez-Bernabeu
and Miguel Selles also acknowledge support from Universitat Politècnica de València (PAID-00-17). Additionally,
some of the neural network and cross validation figures are based on the TikZ codes provided on StackExchange-TeX
by various users. Harish Cherukuri would like to thank them for their valuable advice.
Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design of the
study, in the collection, analyses, or interpretation of data, in the writing of the manuscript, or in the decision to
publish the results.

References
1. Siddhpura, M.; Paurobally, R. A review of chatter vibration research in turning. Int. J. Mach. Tools Manuf.
2012, 61, 27–47. [CrossRef]
2. Chanda, A.; Dwivedy, S.K. Nonlinear dynamic analysis of flexible workpiece and tool in turning operation
with delay and internal resonance. J. Sound Vib. 2018, 434, 358–378. [CrossRef]
3. Copenhaver, R.; Schmitz, T.; Smith, S. Stability analysis of modulated tool path turning. Cirp Ann. Manuf.
Technol. 2018, 67, 49–52. [CrossRef]
4. Filippov, A.V.; Rubtsov, V.E.; Tarasov, S.Y.; Podgornykh, O.A.; Shamarin, N.N. Detecting transition to chatter
mode in peakless tool turning by monitoring vibration and acoustic emission signals. Int. J. Adv. Manuf.
Technol. 2018, 95, 157–169. [CrossRef]
5. Gerasimenko, A.; Guskov, M.; Gouskov, A.; Lorong, P.; Shokhin, A. Analytical modeling of a thin-walled
cylindrical workpiece during the turning process. Stability analysis of a cutting process. J. Vibroengineering
2017, 19, 5825–5841. [CrossRef]
6. Gouskov, A.M.; Guskov, M.A.; Tung, D.D.; Panovko, G.Y. Modeling and investigation of the stability of
a multicutter turning process by a trace. J. Mach. Manuf. Reliab. 2018, 47, 317–323. [CrossRef]
7. Gyebrószki, G.; Bachrathy, D.; Csernák, G.; Stepan, G. Stability of turning processes for periodic chip
formation. Adv. Manuf. 2018, 6, 345–353. [CrossRef]
8. Hajdu, D.; Insperger, T.; Stepan, G. Robust stability analysis of machining operations. Int. J. Adv. Manuf.
Technol. 2017, 88, 45–54. [CrossRef]
9. Huang, X.; Hu, M.; Zhang, Y.; Lv, C. Probabilistic analysis of chatter stability in turning. Int. J. Adv. Manuf.
Technol. 2016, 87, 3225–3232. [CrossRef]
10. Mousavi, S.; Gagnol, V.; Bouzgarrou, B.C.; Ray, P. Dynamic modeling and stability prediction in robotic
machining. Int. J. Adv. Manuf. Technol. 2017, 88, 3053–3065. [CrossRef]
11. Liu, Y.; Li, T.X.; Liu, K.; Zhang, Y.M. Chatter reliability prediction of turning process system with uncertainties.
Mech. Syst. Signal Process. 2016, 66–67, 232–247. [CrossRef]
12. Khasawneh, F.A.; Munch, E. Chatter detection in turning using persistent homology. Mech. Syst. Signal
Process. 2016, 70–71, 527–541. [CrossRef]
13. Jiménez Cortadi, A.; Irigoyen, I.; Boto, F.; Sierra, B.; Suárez, A.; Galar, D. A statistical data-based approach to
instability detection and wear prediction in radial turning processes. Eksploat. I Niezawodn. Maint. Reliab.
2018, 20, 405–412. [CrossRef]
14. Jasiewicz, M.; Powalka, B. Prediction of turning stability using receptance coupling. AIP Conf. Proc. 2018,
1922, 100005. [CrossRef]
15. Lu, K.; Lian, Z.; Gu, F.; Liu, H. Model-based chatter stability prediction and detection for the turning of
a flexible workpiece. Mech. Syst. Signal Process. 2018, 100, 814–826. [CrossRef]
J. Manuf. Mater. Process. 2019, 3, 45 15 of 15

16. Tyler, C.; Troutman, J.R.; Schmitz, T. A coupled dynamics, multiple degree of freedom process damping
model, Part 1: Turning. Precis. Eng. 2016, 46, 65–72. [CrossRef]
17. Ahmad, N.; Janahiraman, T.V.; Tarlochan, F. Modeling of surface roughness in turning operation using
extreme learning machine. Arab. J. Sci. Eng. 2015, 40, 595–602. [CrossRef]
18. Lamraoui, M.; Barakat, M.; Thomas, M.; El Badaoui, M. Chatter detection in milling machines by neural
network classification and feature selection. J. Vib. Control 2015, 21, 1251–1266. [CrossRef]
19. Gupta, A.K.; Guntuku, S.C.; Desu, R.K.; Balu, A. Optimisation of turning parameters by integrating genetic
algorithm with support vector regression and artificial neural networks. Int. J. Adv. Manuf. Technol. 2015, 77,
331–339. [CrossRef]
20. Jurkovic, Z.; Cukor, G.; Brezocnik, M.; Brajkovic, T. A comparison of machine learning methods for cutting
parameters prediction in high speed turning process. J. Intell. Manuf. 2018, 29, 1683–1693. [CrossRef]
21. Khasawneh, F.A.; Munch, E.; Perea, J.A. Chatter classification in turning using machine learning and
topological data analysis. IFAC Pap. 2018, 51, 195–200. [CrossRef]
22. Yao, Z.; Mei, D.; Chen, Z. On-line chatter detection and identification based on wavelet and support vector
machine. J. Mater. Process. Technol. 2010, 210, 713–719. [CrossRef]
23. Zagórski, I.; Kulisz, M.; Semeniuk, A.; Malec, A. Artificial neural network modelling of vibration in the
milling of AZ91D alloy. Adv. Sci. Technol. Res. J. 2017, 11, 261–269. [CrossRef]
24. Kumar, S.; Singh, B. Ascertaining of chatter stability using wavelet denoising and artificial neural network.
Proc. Inst. Mech. Eng. Part C 2018, 233, 39–62. [CrossRef]
25. Schmitz, T.; Smith, K.S. Machining Dynamics: Frequency Response to Improved Productivity, 2nd ed.; Springer:
New York, NY, USA, 2019.
26. Tlusty, G. Manufacturing Equipment and Processes; Prentice-Hall: Upper Saddle River, NJ, USA, 2000.
27. Karandikar, J.; Zapata, R.; Schmitz, T. Incorporating stability, surface location error, tool wear, and uncertainty
in the milling super diagram. Trans. NAMRI/SME 2010, 38, 229–236.
28. Schmitz, T.; Karandikar, J.; Kim, N.H.; Abbas, A. Uncertainty in machining: Workshop summary and
contributions. J. Manuf. Sci. Eng. 2011, 133, 051009. [CrossRef]
29. Schmitz, T.; Ziegert, J.; Canning, J.S.; Zapata, R. Case study: A comparison of error sources in milling. Precis.
Eng. 2008, 32, 126–133. [CrossRef]
30. Kim, H.S.; Schmitz, T. Bivariate uncertainty analysis for impact testing. Meas. Sci. Technol. 2007, 18,
3565–3571. [CrossRef]
31. Fritsch, S.; Guenther, F.; Guenther, M.F. Package ‘neuralnet’. Training of Neural Networks. Available online:
ftp://64.50.236.52/.1/cran/web/packages/neuralnet/neuralnet.pdf (accessed on 31 May 2019).
32. Intrator, O.; Intrator, N. Interpreting neural-network results: A simulation study. Comput. Stat. Data Anal.
2001, 37, 373–393. [CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).