materials

Article
Laser Direct Metal Deposition of 2024 Al Alloy:
Trace Geometry Prediction via Machine Learning
Fabrizia Caiazzo 1 ID
and Alessandra Caggiano 2,3, * ID

1 Department of Industrial Engineering, University of Salerno, 84084 Fisciano (SA), Italy; [email protected]
2 Department of Industrial Engineering, University of Naples Federico II, 80125 Naples, Italy
3 Fraunhofer Joint Laboratory of Excellence on Advanced Production Technology (Fh-J_LEAPT UniNaples),
80125 Naples, Italy
* Correspondence: [email protected]; Tel.: +39-0817682371; Fax: +39-0817682362

Received: 1 March 2018; Accepted: 18 March 2018; Published: 19 March 2018

Abstract: Laser direct metal deposition is an advanced additive manufacturing technology suitably
applicable in maintenance, repair, and overhaul of high-cost products, allowing for minimal distortion
of the workpiece, reduced heat affected zones, and superior surface quality. Special interest is growing
for the repair and coating of 2024 aluminum alloy parts, extensively utilized for a wide range of
applications in the automotive, military, and aerospace sectors due to its excellent plasticity, corrosion
resistance, electric conductivity, and strength-to-weight ratio. A critical issue in the laser direct metal
deposition process is related to the geometrical parameters of the cross-section of the deposited metal
trace that should be controlled to meet the part specifications. In this research, a machine learning
approach based on artificial neural networks is developed to find the correlation between the laser
metal deposition process parameters and the output geometrical parameters of the deposited metal
trace produced by laser direct metal deposition on 5-mm-thick 2024 aluminum alloy plates. The results
show that the neural network-based machine learning paradigm is able to accurately estimate the
appropriate process parameters required to obtain a specified geometry for the deposited metal trace.

Keywords: laser direct metal deposition; aluminum alloy; machine learning; artificial neural network

1. Introduction
In recent years, additive manufacturing processes, characterized by layer upon layer construction
of parts, have emerged as an alternative to conventional processes for the manufacturing of various
metal materials like steel, Inconel, aluminum and titanium alloys [1–4].
Due to progress in computation power and systems technology, laser-based additive
manufacturing, employing a laser beam to provide thermal energy for the melting and consolidating
of the additive materials, has notably advanced and is receiving a great deal of attention due to its
high potential for industrial applications [1,2]. The most common laser-based additive manufacturing
processes use powder material include selective laser sintering, laser beam melting, and laser direct
metal deposition. Selective laser melting (SLM) has been employed for processing various metal alloys
for the biomedical, aerospace and automotive industries, being able to produce complex-shape parts
in a highly efficient way and to provide properties comparable or superior to those produced by other
conventional methods [3,5].
Laser direct metal deposition (DMD) is an advanced additive manufacturing technology
which is attracting increasing interest due to its suitable applicability in maintenance, repair and
overhaul of critical high-cost products, such as those employed in the aerospace and automotive
industry. These complex products may be subject to manufacturing-induced damages or to severe
operating conditions (temperature, wear, and mechanical stresses) hindering the product’s operational

Materials 2018, 11, 444; doi:10.3390/ma11030444 www.mdpi.com/journal/materials

Materials 2018, 11, 444 2 of 11

functionality: in both cases, to avoid scrap part replacement, which create high costs, proper part
recovery operations are required.
In laser direct metal deposition, a laser beam is used as a focused heat source to scan the surface
and create a melting pool over an existing metal substrate. Since the added metal impinging the molten
pool is fed simultaneously with the laser action (i.e., in single stage processing) in the form of wire
or loose powder, a deposited metal trace is generated with metallurgical bonding to the substrate as
a result of fusion and diffusion phenomena [6].
Single weld tracks are placed next to each other in order to form a single layer with thickness
varying from 0.1 mm to several millimeters depending on the process parameters (velocity, powder
feed rate, and laser power) [1,7]. In order to coat wide surfaces on 3D complex geometries, side
overlapping of the individual laser traces is required. The process can be utilized in the repair of worn
out high value components, the building of new components, and the application of wear resistant
and corrosion resistant coatings [8].
In comparison with conventional coating and repair techniques such as arc welding and plasma
spraying, which involve a large temperature increase in the part and a weakening of the base as
a result of the wide temperature distribution in the working region, laser direct metal deposition
allows for minimal distortion of the workpiece, reduced heat affected zones (HAZs), and superior
surface quality [1,9]. Moreover, the coating adherence and its tribological behavior are reported to
be higher in laser DMD [10]. Another interesting aspect of DMD technology is the possibility to
achieve enhanced productivity, higher process automation, and reduction of the overall processing
time, which are main targets within adaptive and flexible manufacturing environments typical of the
factories of the future [11].
At present, two possible DMD feedstock kinds are offered: wire and powder. The former is
generally preferred for its lower cost and inferior probability of oxide content [12], whereas the latter
is preferred for its flexibility in material selection and the higher precision for local repair [13] with
better surface quality and bonding strength [6,14].
As regards the substrate materials to be processed via laser direct metal deposition, interest
is growing for the repair [15] and coating [16] of 2024 aluminum alloy components. This alloy is
characterized by excellent plasticity, corrosion resistance, electric conductivity, and strength-to-weight
ratio. For this reason, among all of the commercial high strength heat-treatable and age-hardenable
aluminum alloys, the 2024 alloy is the most extensively utilized for a wide range of applications and
manufacturing areas, especially in the automotive, military, and aerospace industries.
In this research work, laser direct metal deposition of 2024 Al alloy is investigated through an
experimental campaign under different process conditions carried out on 5-mm-thick T3 temper 2024
Al alloy plates, generating single deposited metal traces with a length of 100 mm.
On the final part, many quality parameters can be of interest, including material density,
mechanical material properties, surface quality, and dimensional and geometrical accuracy [1].
With reference to dimensional and geometrical accuracy, the objective of this research work is the
prediction of the resulting geometry of the deposited metal trace cross-section, which is a critical
issue related to the laser metal deposition process. In the literature, research efforts have been spent
to predict the cross-section geometry of the metal deposition via mathematical modeling based on
prior experimental work. El Cheikh et al. [17] implemented a mathematical model to predict the
clad cross-section dimensions and obtain an analytical description of the clad geometry, establishing
analytical relationships between the radius and the centre of the disk-shaped clad cross-sections and
the process parameters [17]. Since a number of processing parameters are involved in laser direct
metal deposition, statistical analysis of the responses was employed in other research works [18,19].
In the present work, an alternative approach to deal with the complexity of the laser direct
metal deposition modelling was proposed. A machine learning paradigm based on artificial neural
networks (ANN) was developed to find correlations between the input laser metal deposition process
parameters and the output geometrical parameters of the deposited metal trace produced by laser
Materials 2018, 11, 444 3 of 11

DMD of 2024 aluminum alloy. Machine learning is an artificial intelligence method which allows
Materials 2018, 11, x FOR PEER REVIEW 3 of 11
for
data-driven formulation of complex models that lend themselves to predictions or decisions from
sample
sample inputs
inputs revealing
revealingthe thestructural
structural patterns
patterns embedded
embedded in indata
data[20,21].
[20,21].InInthis
thisstudy,
study, thethe specific
specific
Materials 2018, 11, x FOR PEER REVIEW 3 of 11
aim of of
aim thethemachine
machinelearning
learningapproach
approach isis toto estimate
estimate via ANN
ANN the the appropriate
appropriateprocess
processparameters
parameters
required
required to to
sample obtain
obtain
inputs a deposited
arevealing
deposited tracewith
trace with
the structural aacross-section
cross-section
patterns embedded inof given
ofdata
given geometrical
geometrical
[20,21]. parameters.
parameters.
In this study, the specific TheTheresults
results
of the aim
of the of the machine
experimental
experimental learning approach
campaignofoflaser
campaign is
laserDMDto
DMD ofestimate via ANN
of 2024 Al alloy the
alloy areappropriate
areemployed process
employedininorder parameters
ordertototrain
trainthethe
ANNANN
required to obtain a deposited trace with a cross-section of given geometrical parameters. The results
according
according toto a two-phaseprocedure
a two-phase procedurefor for the
the estimation
estimation of
of the
the appropriate
appropriateprocessprocessparameters
parameters andand
of the experimental campaign of laser DMD of 2024 Al alloy are employed in order to train the ANN
following
following verification.
verification.
according to a two-phase procedure for the estimation of the appropriate process parameters and
following verification.
2. Experimental
2. Experimental Procedure
Procedure
2. Experimental Procedure
The
The experimentalcampaign
experimental campaignof oflaser
laser direct
direct metal
metal deposition
depositionon on2024
2024Al Alalloy
alloyplates
plateswas carried
was carried
out using Thea experimental
laser campaign
deposition line,ofwhich
laser direct
is a metal deposition
complex system oncomprising
2024 Al alloyseveral
plates was carried
basic components,
out using a laser deposition line, which is a complex system comprising several basic components, as
out using a laser deposition line, which is a complex system comprising several basic components,
as shown
shown in Figure
1. 1.1.
asinshown
Figurein Figure

Figure 1. Main components in the laser deposition line.

Figure 1. Main components in the laser deposition line.
Figure 1. Main components in the laser deposition line.
A fiber-delivered Yb:YAG disc laser source operating in continuous wave emission with the
Atechnical
A
features reported
fiber-delivered
fiber-delivered Yb:YAG
Yb:YAG
in disc
Tablelaser
1 was employed.
laser source
disc robot source
The displacement
operating
operating in continuous
in
of thewave
continuous laser emission
head is
wave emission with the
accomplished by a six-axis industrial with dedicated controller; an in-built feeding nozzle is with the
technical
technical features
features reported ininTable 11was employed. The displacement of the laser head is
displaced togetherreported Table(Figure
with the laser head was
2). employed. The displacement of the laser head is
accomplished by a six-axis industrial robot with
accomplished by a six-axis industrial robot with dedicated dedicatedcontroller;
controller;an an in-built
in-built feeding
feeding nozzle
nozzle is
is displaced together with the laser
Table 1. head
Main (Figure
technical
displaced together with the laser head (Figure 2). 2).
features of the laser source.

Parameter Value
Table
Table1.1.Main
Main technical
technical features of
of the
the laser
lasersource.
source.
Maximum output powerfeatures
(kW) 4.0
Operating nominal wavelength (nm) 1030
Beam ParameterParameter
Parameter
Product (mm × mrad) Value
8.0Value
Maximum
Core diameter output power
of the delivering
Maximum output power
(kW)
fibre (µm)
(kW) 300 4.04.0
Spot Operating
sizeOperating nominal
of the laser wavelength
nominal wavelength(mm)
beam on the surface (nm)
(nm) 1030
3.0 1030
Beam
BeamParameter
Parameter Product
Product(mm
(mm× × mrad)
mrad) 8.08.0
Core Core diameter
diameter of the
of the delivering
delivering fibre(µm)
fibre (µm) 300300
SpotSpot
sizesize of the
of the laser
laser beam
beam onon the
the surface(mm)
surface (mm) 3.03.0

Figure 2. Laser head performing direct metal deposition.

Figure 2. Laser head performing direct metal deposition.

Figure 2. Laser head performing direct metal deposition.
Materials 2018, 11, 444 4 of 11
Materials 2018, 11, x FOR PEER REVIEW 4 of 11

To supply the
To supply the metal
metal powder,
powder, aa three-way
three-way feeding
feeding nozzle
nozzle waswas used,
used, which
which receives
receives the metal
the metal
powder from a feeder with an oscillating conveyor (Figure 3). Namely, three stream
powder from a feeder with an oscillating conveyor (Figure 3). Namely, three stream cones of metal cones of metal
powder enclosingthe
powder enclosing thelaser
laser beam
beam are provided;
are provided; each stream
each stream is injected
is injected by its argon
by its separate separate argon
conveying
conveying flow. Argon gas, flowing coaxially with the laser beam, was employed
flow. Argon gas, flowing coaxially with the laser beam, was employed to shield the melting pool to shield the
melting pool from the environment. In agreement with the common practice in processing
from the environment. In agreement with the common practice in processing highly reflective metals highly
reflective metals such
such as aluminum as copper
and aluminum and acopper
alloys, tiltingalloys,
angle aoftilting
4◦ was angle of 4°
used forwas
the used
laserfor the to
head laser head
prevent
to prevent back-reflections from entering the
back-reflections from entering the optical train. optical train.

Figure
Figure 3.
3. Scheme
Scheme of
of the
the three-way
three-way feeding
feeding nozzle.
nozzle.

With the scope to repair real components using materials with the same or similar characteristics
With the scope to repair real components using materials with the same or similar characteristics
as the substrate metal, 2024 aluminum alloy powder was utilized in this study. The spherical shaped
as the substrate metal, 2024 aluminum alloy powder was utilized in this study. The spherical shaped
powder had particle sizes ranging between 20 and 60 µm, as certified by the powder supplier. Since
powder had particle sizes ranging between 20 and 60 µm, as certified by the powder supplier.
a steady feeding rate must be provided constantly, the powder was preliminary dried in furnace at 180
Since a steady feeding rate must be provided constantly, the powder was preliminary dried in furnace
°C for 2◦ h so as to properly flow via the conveyor.
at 180 C for 2 h so as to properly flow via the conveyor.
The workpieces to be processed consisted of 5-mm-thick 2024 aluminum alloy plates in T3 heat
The workpieces to be processed consisted of 5-mm-thick 2024 aluminum alloy plates in T3 heat
treatment state. The laser DMD process was performed so as to produce cladding traces in the form
treatment state. The laser DMD process was performed so as to produce cladding traces in the form of
of single metal deposition along an overall scanning length of 100 mm under diverse process conditions.
single metal deposition along an overall scanning length of 100 mm under diverse process conditions.
Since numerous process variables are involved in direct metal deposition, a systematic approach
was Since numerous
employed process
to select the variables are involved
process conditions of inthedirect metal deposition,
experimental campaign. a systematic
The mainapproach
process
was employed to select the process conditions of the experimental
parameters of the experimental plan were selected with reference to the literature and campaign. The main
past process
experience:
parameters of the experimental plan were selected with reference to the
namely, laser power, P, scanning speed, v, and powder feeding rate, ṁ, were considered. literature and past experience:
namely,
Thelaser power, P, scanning
thermophysical speed,
properties v, material,
of the and powder such feeding ṁ, were considered.
rate, conductivity,
as thermal thermal expansion
coefficient and melting point, have a significant influence on the geometrythermal
The thermophysical properties of the material, such as thermal conductivity, expansion
and mechanical
coefficient and melting point, have a significant influence on the geometry
characteristics of the deposition. However, since in this research work only one type of material was and mechanical
characteristics
employed, bothofforthethe
deposition.
substrate as However,
well as the since in thisthese
powder, research work only
properties one type
remained of material
constant in the
experimental campaign and therefore they were not considered in the implementation ofconstant
was employed, both for the substrate as well as the powder, these properties remained the ANN in
the experimental campaign and therefore they were not considered in the implementation of the
methodology.
ANNLaser
methodology.
power was varied among five different levels comprised between 1200 W and 3000 W,
Laser
scanning speedpower was was varied
varied among
among fivedifferent
seven differentlevels
levels comprised
between between
150 and 1200 W and
600 mm/min, and powder
3000 W,
scanningrate
feeding speed
was was varied
varied amongseven
among sevendifferent
differentlevels
levels between
between 1504 and and10600 mm/min,
g/min. and powder
For each process
feeding rate was varied among seven different levels between 4 and
condition, three repetitions were carried out, and they were performed in random order so 10 g/min. For each process
as to
condition, three repetitions
reduce systematic errors. were carried out, and they were performed in random order so as to reduce
systematic
In ordererrors.
to evaluate the geometry of the deposited metal trace cross-section, cross-cutting and
In order to evaluatewas
mechanical preparation the geometry
carried outofon thethree
deposited
samples metal
for trace
each cross-section,
process condition.cross-cutting
Polishingand to
mechanical preparation was carried out on three samples for each process
mirror finish and chemical etching with a solution of 10% hydrofluoric acid and 15% nitric acid incondition. Polishing to
mirror finish and chemical etching with a solution of 10% hydrofluoric
water at room temperature were performed. Figure 4 shows the view and the cross-section of a acid and 15% nitric acid
in water atmetal
deposited room temperature
trace produced were
by oneperformed. Figure 4 tests
of the experimental shows the view
of laser DMDand the cross-section
carried out at P = 3 kW, of
va =deposited
200 mm/min,metal ṁ =trace produced by one of the experimental tests of laser DMD carried out at
4 g/min.
P = 3 kW, v = 200 mm/min, ṁ = 4 g/min.
Materials 2018, 11, 444 5 of 11
Materials 2018, 11, x FOR PEER REVIEW 5 of 11

Figure4.4.Deposited
Figure Depositedmetal
metaltrace
traceview
viewand
andcross-section
cross-section(P
(P==33kW,
kW,vv==200
200mm/min, ṁ=
mm/min, ṁ = 44 g/min).
g/min).

The characteristic geometrical parameters of the deposited trace cross-section were measured as
The characteristic geometrical parameters of the deposited trace cross-section were measured as
indicated in Figure 5. The measurement values for each process condition are reported in (Table 2).
indicated in Figure 5. The measurement values for each process condition are reported in (Table 2).
Table 2. Measured values of the cross-section geometrical parameters for the three repetitions of the
Table 2. Measured values of the cross-section geometrical parameters for the three repetitions of the
diverse experimental process conditions.
diverse experimental process conditions.
Process Conditions Measured Geometrical Parameters
P (W) v Process
(mm/min) ṁ (g/min)
Conditions Width (mm) Depth (mm)
Measured Geometrical ParametersHeight (mm)
1200P (W) 400
v (mm/min) 4
ṁ (g/min) 2.93, 3.13,
Width (mm)3.21 0.25, 0.26,
Depth (mm) 0.25 0.96, 1.02,
Height (mm)0.96
1500 400 4 3.90, 3.78, 3.45 0.48, 0.45, 0.45 1.11, 1.11, 1.02
1200 400 4 2.93, 3.13, 3.21 0.25, 0.26, 0.25 0.96, 1.02, 0.96
1500 400 6 3.61, 3.45, 3.78 0.38, 0.36, 0.37 1.43, 1.43, 1.52
1500 400 4 3.90, 3.78, 3.45 0.48, 0.45, 0.45 1.11, 1.11, 1.02
1500 400 8 3.69, 3.52, 3.26 0.11, 0.11, 0.11 1.99, 2.06, 1.90
1500 400 6 3.61, 3.45, 3.78 0.38, 0.36, 0.37 1.43, 1.43, 1.52
15001500 400 400 10 8 4.57, 4.30, 4.03
3.69, 3.52, 3.26 0.20, 0.22, 0.21
0.11, 0.11, 0.11 2.06,
1.99, 2.18,
2.06, 1.902.18
20001500 200 400 3 10 5.06, 5.40, 5.26
4.57, 4.30, 4.03 0.68, 0.68, 0.76
0.20, 0.22, 0.21 2.06, 2.18, 2.180.88
0.84, 0.83,
20002000 200 200 10 3 5.30, 5.22, 5.32
5.06, 5.40, 5.26 0.77, 0.73, 0.74
0.68, 0.68, 0.76 2.02,
0.84, 2.12,
0.83, 0.882.20
20002000 300 200 3 10 5.02,
5.30,5.48, 5.48
5.22, 5.32 1.87, 0.73,
0.77, 1.81, 0.74
1.75 1.38,
2.02, 1.43,
2.12, 2.201.33
20002000 400 300 8 3 5.02,
4.28, 5.48,
3.86, 5.48
4.07 1.87,
0.59, 1.81,
0.60, 1.75
0.58 1.38, 1.43,
1.78, 1.331.79
1.68,
20002000 500 400 8 8 4.28,
4.01, 3.86,
4.09, 4.07
3.73 0.59,
0.64, 0.60,
0.65, 0.58
0.66 1.78, 1.68,
1.52, 1.791.51
1.44,
20002000 500 500 3 8 4.01,
4.49, 4.09,
4.12, 3.73
4.32 0.64,
1.21, 0.65,
1.26, 0.66
1.19 1.52, 1.44,
1.12, 1.511.15
1.07,
25002000 100 500 5 3 4.49,
11.18, 4.12,10.77
10.16, 4.32 1.21,
2.74, 1.26,
2.91, 1.19
2.84 1.12, 1.07,
0.85, 1.150.81
0.80,
2500 100 5 11.18, 10.16, 10.77 2.74, 2.91, 2.84 0.85, 0.80, 0.81
2500 150 5 8.95, 8.93, 7.62 1.50, 1.51, 1.52 1.08, 1.06, 0.98
2500 150 5 8.95, 8.93, 7.62 1.50, 1.51, 1.52 1.08, 1.06, 0.98
2500 150 7 9.31, 9.32, 8.08 1.51, 1.51, 1.54 2.13, 2.07, 1.99
2500 150 7 9.31, 9.32, 8.08 1.51, 1.51, 1.54 2.13, 2.07, 1.99
25002500 200 200 5 5 7.94,
7.94,8.43, 7.78
8.43, 7.78 1.50, 1.53,
1.50, 1.53, 1.47
1.47 1.07,
1.07, 1.08,
1.08, 1.001.00
25002500 200 200 7 7 8.25,
8.25,8.38, 9.35
8.38, 9.35 1.66, 1.68,
1.66, 1.68, 1.52
1.52 1.80,
1.80, 1.82,
1.82, 2.052.05
25002500 250 250 5 5 8.58,
8.58,8.54, 7.84
8.54, 7.84 1.71, 1.65, 1.71 0.81,
0.81, 0.81,
0.81, 0.730.73
25002500 400 400 3 3 5.56,
5.56,5.34, 5.00
5.34, 5.00 2.13, 2.11, 2.15 1.12,
1.12, 1.06,
1.06, 1.131.13
30003000 150 150 5 5 8.95,
8.95, 9.07,
9.07, 7.78
7.78 2.36, 2.38, 2.53 1.20, 1.11,
1.20, 1.171.17
1.11,
3000 3000 150 150 7 7 9.62, 9.49, 9.11
9.62, 9.49, 9.11 2.59, 2.61, 2.60
2.59, 2.61, 2.60 1.60, 1.52, 1.471.47
1.60, 1.52,
3000 3000 200 200 5 5 8.56, 8.16, 7.44
8.56, 8.16, 7.44 2.50, 2.43, 2.36
2.50, 2.43, 2.36 1.00, 1.01, 0.980.98
1.00, 1.01,
3000 3000 200 200 7 7 8.66, 8.54, 9.21
8.66, 8.54, 9.21 2.36, 2.40, 2.47
2.36, 2.40, 2.47 1.35, 1.28, 1.391.39
1.35, 1.28,
3000 200 4 8.17, 7.82, 8.25 2.39, 2.44, 2.34 0.81, 0.82, 0.74
3000 200 4 8.17, 7.82, 8.25 2.39, 2.44, 2.34 0.81, 0.82, 0.74
3000 200 6 9.33, 9.11, 9.23 1.90, 1.85, 1.74 1.16, 1.22, 1.22
30003000 200 200 6 3 9.33, 9.11, 9.23
7.24, 7.75, 7.51
1.90, 1.85, 1.74
2.21, 2.23, 2.37
1.16, 1.22, 1.22
0.69, 0.66, 0.66
30003000 200 200 3 10 7.24, 7.75, 7.51
9.51, 9.91, 8.48 2.21, 2.23, 2.37
2.39, 2.33, 2.27 0.69,
1.86, 0.66,
1.78, 1.950.66
30003000 200 300 10 3 9.51,
6.71,9.91, 8.48
7.04, 7.37 2.39, 3.31,
3.14, 2.33, 3.24
2.27 1.86,
1.05, 1.78,
1.03, 1.141.95
30003000 300 500 3 3 6.71,
5.23,7.04, 7.37
5.16, 5.76 3.14, 2.21,
2.19, 3.31, 2.35
3.24 1.05,
0.93, 1.03,
1.00, 0.951.14
30003000 500 600 3 3 5.23,
4.97,5.16, 5.76
4.68, 4.37 2.19, 0.90,
0.94, 2.21, 0.92
2.35 0.93,
0.28, 1.00,
0.29, 0.300.95
30003000 600 600 3 10 5.83,
4.97, 5.68,
4.68, 5.00
4.37 1.05,
0.94, 1.06,
0.90, 1.01
0.92 1.30, 1.30,
0.28, 1.240.30
0.29,
3000 600 10 5.83, 5.68, 5.00 1.05, 1.06, 1.01 1.30, 1.30, 1.24
Materials 2018, 11, 444 6 of 11
Materials 2018, 11, x FOR PEER REVIEW 6 of 11

Figure 5.
Figure 5. Scheme
Scheme of
of the
the deposited
deposited trace
trace cross-section
cross-sectiongeometrical
geometricalparameters.
parameters.

3. Artificial
3. ArtificialNeural
NeuralNetwork-Based
Network-BasedMachine
MachineLearning
Learningfor
forProcess
ProcessParameters
ParametersEstimation
Estimation
In this
In this research
research work,
work, aamachine
machine learning
learning paradigm
paradigm basedbased on on artificial
artificial neural
neural networks
networks (ANN)
(ANN)
was developed to find correlations between the input laser direct metal
was developed to find correlations between the input laser direct metal deposition process parameters deposition process
parameters
(laser power, (laser power,
P, scanning P, scanning
speed, speed,
v, and powder v, and
feeding powder
rate, ṁ) and feeding
the output rate, ṁ) and parameters
geometrical the output
geometrical parameters (width, w, depth, d, and height, h) of the deposited trace
(width, w, depth, d, and height, h) of the deposited trace produced by direct metal deposition (Figure produced by direct5).
metal Thedeposition
specific aim(Figure
was 5).
to estimate via ANN the appropriate process parameters required to obtain
The specific
a deposited trace aim
withwas to estimate
given w, d, and via
h ANN the appropriate
geometrical parameters. process
The parameters required
results obtained to obtain
through the
a deposited trace with given w, d, and h geometrical parameters. The results
experimental campaign were employed to train the ANN according to a two-phase procedure, the first obtained through the
experimental
for the estimation campaign were employed
of the appropriate to train
process the ANN
parameters, and according
the second to for
a two-phase procedure,
verification purpose. the
first for the estimation of the appropriate process parameters, and the second for verification purpose.
3.1. ANN Data Processing for Process Parameters Estimation
3.1. ANN Data Processing for Process Parameters Estimation
In the first phase, the geometrical parameters (width, depth, and height) of the deposited metal
trace In the firstby
produced phase,
each the geometrical
experimental parameters
test were combined (width,to depth,
construct andfeature
height)pattern
of the vectors
deposited metal
(FPV) to
trace
be fedproduced
as input for by ANN
each experimental
learning [22].test
Due were
to thecombined
fact thatto30construct
differentfeature patternof
combinations vectors (FPV)
laser DMD
to be fed
process as input for
parameters ANN
were learning [22].tested,
experimentally Due toand the each
fact that 30 different
experimental combinations
condition of laser DMD
was repeated three
process parameters were experimentally tested, and each experimental
times, a training set of n = 90 FPV (one FPV for each experimental test) was generated. condition was repeated three
times, a training set of n = 90 FPV (one FPV for each experimental test) was generated.
FPVi = [wi , di , hi ] i = 1, . . . , n (1)
FPVi = [wi, di, hi] i = 1, …, n (1)
To
To perform
perform supervised
supervised ANN
ANN learning,
learning, each
each three-features
three-features FPV
FPVii was
was associated
associated to
to the
the
corresponding process parameters combination (Pi i, vii, ṁii), representing the output of the ANN.
corresponding process parameters combination (P , v , ṁ ), representing the output of the ANN.

Outputi =
Outputi [Pii,, vvii,,ṁ
= [P ṁi]i ] i i==1,1,. …,
. . , nn (2)
(2)
As regards the ANN architecture, a three-layer cascade-forward backpropagation ANN was set
As regards
up with a number theofANN
inputarchitecture, a three-layer
layer nodes equal to 3, i.e.,cascade-forward
matching the number backpropagation ANNand
of input features, wasa
set up with a number of input layer nodes equal to 3, i.e., matching
number of output layer nodes equal to 3, i.e., corresponding to the number of output processthe number of input features,
and a number
parameters. of output
Different layer nodes
numbers equallayer
of hidden to 3, nodes
i.e., corresponding
were tested, from to the1×number
to 3× theofnumber
output process
of input
parameters. Different numbers of hidden layer nodes were tested, from 1 ×
layer nodes, with the aim to identify the optimal ANN architecture providing the best performance to 3 × the number of input
layer nodes,
rate. The with the aim to identify
Levenberg–Marquardt the optimal
algorithm ANN architecture
was selected as ANN trainingproviding the best performance
function.
rate. The Levenberg–Marquardt
In ANN learning, the training algorithm
subsetwas selectedfor
is utilized ascalculating
ANN training the function.
gradient and updating the
ANN weights and biases, while the testing subset is used to evaluate thegradient
In ANN learning, the training subset is utilized for calculating the performanceand updating the
of the trained
ANN weights
ANN [22,23]. and biases, while the testing subset is used to evaluate the performance of the trained
ANNIn [22,23].
this research work, the training set of n = 90 FPV was employed to train the ANN, while the
testing this
In research
procedure waswork, the training
performed using set of n = 90
a different FPVFPVsetwas employed
constructed to train
using the ANN,
the average while
values of
the testing procedure was performed using a different FPV set constructed
the geometrical parameters (width, depth, and height) of the deposit calculated over the three using the average values
of the geometrical
repetitions of each parameters
experimental (width, depth,
condition. and30
Since height)
differentof the deposit calculated
experimental conditions over thetested,
were three
repetitions of each experimental condition. Since 30 different experimental
the testing set, built with the average values of measured deposit width, depth, and height, was madeconditions were tested,
the
of ptesting
= 30 set,
FPV. built
The with the average
overall patternvalues of measured
recognition deposit width,
performance over thedepth, and height,
testing set waswas made of
eventually
pestimated
= 30 FPV.by The overall pattern recognition performance
aggregating the p recognition rates obtained. over the testing set was eventually estimated
by aggregating
The pattern p recognition
therecognition rates obtained.
performance of each ANN architecture was evaluated in terms of root
mean square error (RMSE) between ANN predicted values (ŷt) and target values (y).
Materials 2018, 11, 444 7 of 11

The pattern recognition performance of each ANN architecture was evaluated in terms of root
mean square error (RMSE) between ANN predicted values (ŷt ) and target values (y).
s
Materials 2018, 11, x FOR PEER REVIEW 2 7 of 11
Materials 2018, 11, x FOR PEER REVIEW ∑nt=1 (ŷt − y) 7 of 11
RMSE = (3)
n
∑ ( − )
= ∑ ( − ) (3)
The best performance was obtained for the = ANN architecture with a number of hidden (3) layer
nodes equal to nine, i.e., three times the number of hidden layer nodes. The aggregated RMSE values
The best performance was obtained for the ANN architecture with a number of hidden layer
by The
providednodes this best
equal
performance was
architecture
to nine, i.e.,over
three the
obtained
times
for the cases
p =the30number
testing ANN architecture
of hiddenwere
layerequal with
nodes.toThe a number
61.6 of hidden
W for power,
aggregated RMSE18.0
layer
mm/min
values
nodes
for speed, equal
and 0.38 to nine, i.e., three times the number of hidden layer nodes. The aggregated RMSE values
provided by g/min for powder
this architecture overfeeding
the p = 30rate.
testing cases were equal to 61.6 W for power, 18.0
provided by this architecture over the p = 30 testing cases were equal to 61.6 W for power, 18.0
Themm/min
results for
arespeed,
shown and in0.38
Figures
g/min6–8, which feeding
for powder for each of the p = 30 experimental conditions,
report, rate.
mm/min for speed, and 0.38 g/min for powder feeding rate.
The results
the experimental laser are shown
direct in Figures
metal 6–8, which
deposition report,parameters
process for each of the(laser
p = 30 experimental conditions,
power, P, scanning speed, v,
The results are shown in Figures 6–8, which report, for each of the p = 30 experimental conditions,
the experimental laser direct metal deposition process parameters (laser power, P, scanning speed, v,
and powder feeding rate,
the experimental ṁ)direct
laser and themetalprocess parameters
deposition estimated(laser
process parameters by the ANN
power, with nine
P, scanning hidden
speed, v, layer
and powder feeding rate, ṁ) and the process parameters estimated by the ANN with nine hidden
nodes. From the Figures,
and powder feeding it canṁ)beand
rate, clearly observed
the process that the
parameters experimental
estimated by the ANN andwith
the nine
estimated
hidden values
layer nodes. From the Figures, it can be clearly observed that the experimental and the estimated
are verylayer
closenodes.
values to
areeach
From the Figures,
very other,
close toindicating
it can be the
each other,that
clearly
indicating
observed
neural that the
network
that the
experimental
evaluation
neural network wasand
evaluation
the estimated
accurately performed.
was accurately
The meanvalues are very
absoluteThe close
percentage to each other, indicating
error percentage
of the ANN that the
estimation neural
was network
as low as evaluation
2.0% foras was
laser accurately5.8% for
performed. mean absolute error of the ANN estimation was as low 2.0%power,
for laser
performed. The mean absolute percentage error of the ANN estimation was as low as 2.0% for laser
scanningpower,
speed, 5.8%
and for5.5
scanning
% for speed,
powder andfeeding
5.5 % forrate.
powder feeding rate.
power, 5.8% for scanning speed, and 5.5 % for powder feeding rate.

Figure 6. Experimental
Experimental
Figure 6.Figure (blue(blue marks)
marks) and and ANNestimated
ANN estimated (orange
(orangemarks)
marks)power values,
power P, for P,
values, each
forofeach of
6. Experimental (blue marks) and ANN estimated (orange marks) power values, P, for each of
the p = 30 experimental conditions.
the p = 30
theexperimental conditions.
p = 30 experimental conditions.

Figure 7. Experimental (blue marks) and ANN estimated (orange marks) speed values, v, for each of
Figure 7.Figure 7. Experimental
Experimental
the (blue(blue
p = 30 experimental
marks)
marks) and
and
conditions.
ANNestimated
ANN estimated (orange
(orangemarks)
marks)speed values,
speed v, for each
values, v, forofeach of
the p = 30 experimental conditions.
the p = 30 experimental conditions.
Materials 2018, 11, 444 8 of 11
Materials 2018, 11, x FOR PEER REVIEW 8 of 11

Figure 8.Figure 8. Experimental

Experimental (blue
(blue marks)and
marks) and ANN
ANN estimated
estimated(orange marks)
(orange powderpowder
marks) feeding rate, ṁ, forrate, ṁ,
feeding
each of the p = 30 experimental conditions.
for each of the p = 30 experimental conditions.
3.2. ANN Data Processing for Geometrical Parameters Verification
3.2. ANN Data Processing for Geometrical Parameters Verification
The second phase of the ANN procedure was employed to verify the results provided by the
ANN
The second in the first phase,
phase of thebyANN
performing the reverse
procedure was data processing,
employed i.e., estimating
to verify the geometrical
the results provided by the
parameters (width, depth, and height) of the deposit produced using the process
ANN in the first phase, by performing the reverse data processing, i.e., estimating the parameters estimated
geometrical
in the first ANN phase.
parameters (width, depth, and height) of the deposit produced using the process parameters estimated
In this phase, the laser DMD process parameters (laser power, P, scanning speed, v, and powder
in the first ANN
feeding phase.
rate, ṁ) were combined to build up the feature pattern vectors (FPV) to be fed as input for
In this phase,
ANN learning. the laser DMD process parameters (laser power, P, scanning speed, v, and powder
feeding rate,Asṁ)30were combined
different to buildofup
combinations thedirect
laser feature pattern
metal vectors
deposition (FPV)parameters
process to be fed as input for
were
experimentally tested, and each experimental condition was repeated three times, also in this case
ANN learning.
the training set was made of n = 90 FPV, one FPV for each experimental test.
As 30 different combinations of laser direct metal deposition process parameters were
FPVi = [Pi, vi, ṁ
experimentally tested, and each experimental i]
condition wasirepeated
= 1, …, n three times, also in this
(4) case the
training set was
To perform n = 90 FPV,
made ofsupervised ANNonelearning,
FPV foreach eachthree-features
experimental test.
FPV i was associated with the

corresponding deposit geometrical parameters, representing in this case the output of the ANN.
FPV
Output
vi ,i, ṁ
i = [Pi i=, [w ]
di,i hi]
i =i =1,1,. …,
. . ,nn (5)
(4)

Three-layer
To perform cascade-forward
supervised ANN backpropagation
learning, eachANN were setup with
three-features FPVthe same architectures and
i was associated with the
transfer function utilized for the first phase.
corresponding deposit geometrical parameters, representing in this case the output of the ANN.
As regards the testing FPV set, it was constructed using as input the p = 30 combinations of
process parameters (P, v, ṁ) estimated by the ANN in the first phase. The ANN output during the
testing phase should provide Output
estimated i , di , hof
i = [wvalues i ] the depositi= 1, . . . , n parameters (width, depth,
geometrical (5)
and height) sufficiently close to the average values provided as input in the first phase. The overall
Three-layer cascade-forward
pattern recognition performance backpropagation
over the testing set ANN were setup
was eventually with the
estimated bysame architectures
aggregating the p and
transfer recognition
function utilized for the first phase.
rates obtained.
As before,
As regards the pattern
the testing FPV recognition
set, it wasperformance
constructed of eachusing
ANN architecture
as input the waspevaluated in terms
= 30 combinations of
process of root mean square
parameters (P, v, error (RMSE) between
ṁ) estimated by theANN ANN predicted
in the values (ŷt) and The
first phase. targetANN
valuesoutput
(y). during the
The best performance was obtained for the ANN architecture with a number of hidden layer
testing phase should provide estimated values of the deposit geometrical parameters (width, depth,
nodes equal to nine, i.e., three times the number of hidden layer nodes. The RMSE values provided
and height) sufficiently
by this architectureclose to the
over the average
p = 30 testing values
cases were provided
equal toas input
0.59 mm forin the first phase.
the deposit width,The0.53overall
pattern mm
recognition performance
for the deposit depth, andover0.14 mmthefor
testing set was
the deposit eventually estimated by aggregating the
height.
p recognition rates obtained.
As before, the pattern recognition performance of each ANN architecture was evaluated in terms
of root mean square error (RMSE) between ANN predicted values (ŷt ) and target values (y).
The best performance was obtained for the ANN architecture with a number of hidden layer
nodes equal to nine, i.e., three times the number of hidden layer nodes. The RMSE values provided by
this architecture over the p = 30 testing cases were equal to 0.59 mm for the deposit width, 0.53 mm for
the deposit depth, and 0.14 mm for the deposit height.
Materials 2018,11,
Materials2018, 11,444
x FOR PEER REVIEW 99ofof11
11

The results of the second phase are reported in Figures 9–11, which show, for each of the p = 30
The resultsconditions,
experimental of the second
thephase are reported
average in Figures
experimental 9–11, which
geometrical show, for
parameters the p =and
each ofdepth,
(width, 30
experimental conditions, the average experimental geometrical parameters (width, depth,
height) and the predicted geometrical parameters estimated by the ANN with nine hidden layer and height)
and the The
nodes. predicted geometrical
Figures parameters
indicate that estimated
the neural networkbyevaluation
the ANN with nine hidden
accuracy, despite layer
some nodes.
minor
The Figures indicate that the neural network evaluation accuracy, despite some minor
deviations, is generally high, providing for a successful verification of the results obtaineddeviations,
in the
isANN
generally high, providing for a
process parameters estimation.successful verification of the results obtained in the ANN process
parameters estimation.

Figure9.9.Experimental
Figure Experimentalaverage
average(blue
(bluemarks)
marks)and
andANN
ANNestimated
estimated(orange
(orangemarks)
marks)deposit
deposit width, w,
width, w,
for each of the p = 30 experimental conditions.
for each of the p = 30 experimental conditions.

Figure10.
Figure 10.Experimental
Experimentalaverage
average(blue
(bluemarks)
marks)and
andANN
ANNestimated
estimated(orange
(orangemarks)
marks)deposit
deposit depth, d,
depth, d,
foreach
for eachofofthe
thepp==30
30experimental
experimentalconditions.
conditions.
Materials 2018, 11, 444 10 of 11
Materials 2018, 11, x FOR PEER REVIEW 10 of 11

Figure11.
Figure 11.Experimental
Experimentalaverage
average(blue
(bluemarks)
marks)andandANN
ANNestimated
estimated(orange
(orangemarks)
marks)deposit
depositheight, h,
height,h,
foreach
for eachofofthe
thepp==30
30experimental
experimentalconditions.
conditions.

4.4.Conclusions
Conclusions

InInthis
thisresearch
researchwork,
work,laser
laserdirect
directmetal
metaldeposition
depositionof of2024
2024Al Alalloy
alloywas
wasinvestigated
investigatedthroughthroughan an
experimental campaign under different process conditions carried out on 5-mm-thick
experimental campaign under different process conditions carried out on 5-mm-thick T3 temper 2024 T3 temper 2024
Alalloy
Al alloyplates,
plates,generating
generatingsingle
singledeposited
depositedmetalmetaltraces
tracesfor
foraalength
lengthofof100
100mm.
mm.AAcritical
criticalissue
issueininthe
the
laser DMD process is related to the geometrical parameters of the cross-section of
laser DMD process is related to the geometrical parameters of the cross-section of the deposited metal the deposited metal
tracethat
trace thatshould
shouldbe becontrolled
controlledtotomeetmeetthethepart
partspecifications.
specifications. To Tothis
thisaim,
aim,complex
complexmathematical
mathematical
modelling is required including in the model all the relevant processing factors. In
modelling is required including in the model all the relevant processing factors. In this research work, this research work,
a machine learning approach based on artificial neural networks was developed
a machine learning approach based on artificial neural networks was developed to find the correlation to find the
correlation
between thebetween
laser DMDtheprocess
laser DMD process parameters
parameters and the output andgeometrical
the output geometrical
parameters parameters of the
of the deposited
deposited metal trace on 2024 aluminum alloy plates. The specific aim was to
metal trace on 2024 aluminum alloy plates. The specific aim was to identify via ANN the appropriate identify via ANN the
appropriate process parameters required to obtain a deposited trace with
process parameters required to obtain a deposited trace with given geometrical parameters in terms of given geometrical
parameters
width, in terms
w, depth, d, andofheight,
width,h.w,Thedepth,
resultsd, showed
and height, h. The
that the results showed
ANN-based machinethat the ANN-based
learning paradigm
machine learning paradigm is able to accurately estimate the correct laser
is able to accurately estimate the correct laser power, scanning speed and powder feeding power, scanning speed
rateand
to
powdera feeding
achieve specifiedrate to achieve
geometry for thea deposited
specified geometry
metal trace, forwith
the mean
deposited metal
absolute trace, with
percentage mean
errors of
absolute
the ANN percentage
estimation as errors
low of
asthe
2.0%ANN estimation
for laser power,as5.8%
low for
as 2.0% for laser
scanning power,
speed, and 5.8%
5.5% for
for scanning
powder
speed, and
feeding rate. 5.5% for powder feeding rate.

Acknowledgments: The
Acknowledgments: The Fraunhofer
Fraunhofer Joint Joint Laboratory
Laboratory of
of Excellence
Excellence on on Advanced
Advanced Production
Production Technology
Technology
(Fh-J_LEAPT
(Fh-J_LEAPTUniNaples)
UniNaples) atatthe
theDepartment
Departmentof ofChemical,
Chemical,Materials,
Materials, andandIndustrial
IndustrialProduction
Production Engineering,
Engineering,
University
UniversityofofNaples
NaplesFederico
FedericoII,II,isisgratefully
gratefullyacknowledged
acknowledgedforforits
itssupport
supporttotothis
thisresearch
researchactivity.
activity.
Author Contributions: Fabrizia Caiazzo mainly designed and performed the experiments; Alessandra Caggiano
Author Contributions: Fabrizia Caiazzo mainly designed and performed the experiments; Alessandra Caggiano
mainly analyzed the data; both authors equally contributed to the discussion of the results and the writing of
mainly analyzed the data; both authors equally contributed to the discussion of the results and the writing of
the paper.
the paper.
Conflicts of Interest: The authors declare no conflict of interest.
Conflicts of Interest: The authors declare no conflict of interest.
References
References
1. Schmidt, M.; Merklein, M.; Bourell, D.; Dimitrov, D.; Hausotte, T.; Wegener, K.; Overmeyer, L.; Vollertsen, F.;
1. Schmidt,
Levy, G.N.M.; Merklein,
Laser M.; Bourell,
based additive D.; Dimitrov,inD.;
manufacturing Hausotte,
industry andT.;academia.
Wegener, CIRP
K.; Overmeyer,
Ann. 2017,L.;66,Vollertsen,
561–583.
F.; Levy,
[CrossRef] G.N. Laser based additive manufacturing in industry and academia. CIRP Ann. 2017, 66, 561–583.
2.2. Gibson, I.; Rosen, D.W.; Stucker, B. Additive Manufacturing Technologies; Springer: New York, NY,
Gibson, I.; Rosen, D.W.; Stucker, B. Additive Manufacturing Technologies; Springer: New York, NY, USA, 2015.USA,
2015.
Materials 2018, 11, 444 11 of 11

3. Attar, H.; Ehtemam-Haghighi, S.; Kent, D.; Wu, X.; Dargusch, M.S. Comparative study of commercially pure
titanium produced by laser engineered net shaping, selective laser melting and casting processes. Mater. Sci.
Eng. A 2017, 705, 385–393. [CrossRef]
4. International Organization for Standardization. Standard Terminology for Additive Manufacturing—General
Principles—Part 1: Terminology; ISO/ASTM 52900:2015(E); ASTM: West Conshohocken, PA, USA, 2015.
5. Attar, H.; Ehtemam-Haghighi, S.; Kent, D.; Okulov, I.V.; Wendrock, H.; Bönisch, M.; Volegov, A.S.; Calin, M.;
Eckert, J.; Dargusch, M.S. Nanoindentation and wear properties of Ti and Ti-TiB composite materials
produced by selective laser melting. Mater. Sci. Eng. A 2017, 688, 20–26. [CrossRef]
6. Toyserkani, E.; Khajepour, A.; Corbin, S. Laser Cladding; CRC Press: Boca Raton, FL, USA, 2005.
7. Ocylok, S.; Alexeev, E.; Mann, S.; Weisheit, A.; Wissenbach, K.; Kelbassa, I. Correlations of Melt
Pool Geometry and Process Parameters during Laser Metal Deposition by Coaxial Process Monitoring.
Phys. Procedia 2014, 56, 228–238. [CrossRef]
8. Dutta, B.; Singh, V.; Natu, H.; Choi, J.; Mazumder, J. Direct Metal Deposition. Adv. Mater. Proc. 2009, 167,
29–31.
9. Carcel, B.; Serrano, A.; Zambrano, J.; Amigó, V.; Cárcel, A.C. Laser cladding of TiAl intermetallic alloy on
Ti6Al4V, process optimization and properties. Phys. Procedia 2014, 56, 284–293. [CrossRef]
10. Khanna, K.S.; Kanungo, S.; Gasser, A. Hard coatings based on thermal spray and laser cladding. Int. J.
Refract. Met. Hard Mater. 2009, 27, 485–491. [CrossRef]
11. Del Val, J.; Comesana, R.; Lusquinos, F.; Boutinguiza, M.; Riveiro, A.; Quintero, F.; Pou, J. Laser cladding of
Co-based superalloy coatings: Comparative study between Nd:YAG laser and fibre laser. Surf. Coat. Technol.
2010, 204, 1957–1961. [CrossRef]
12. Hussein, N.I.S.; Segal, J.; McCartney, D.G.; Pashby, I.R. Microstructure formation in Waspaloy multilayer
builds following direct metal deposition with laser and wire. Mater. Sci. Eng. A 2008, 497, 260–269. [CrossRef]
13. Riveiro, A.; Mejías, A.; Lusquiños, F.; del Val, J.; Comesaña, R.; Pardo, J.; Pou, J. Laser cladding of aluminium
on AlSi 304 stainless steel with high-power diode lasers. Surf. Coat. Technol. 2014, 253, 214–220. [CrossRef]
14. Danesh Manesh, H.; Mashreghi, A.; Ehtemam Haghighi, S.; Khajeh, A. Investigation of cold pressure welding
of aluminum powder to internal surface of aluminum tube. Mater. Des. 2009, 30, 723–726. [CrossRef]
15. Pantelakis, S.G.; Chamos, A.N.; Kermanidis, A.T. A critical consideration for the use of Al-cladding for
protecting aircraft aluminum alloy 2024 against corrosion. Theor. Appl. Fract. Mech. 2012, 57, 36–42.
[CrossRef]
16. Tian, Y.S. Laser fabrication of nickel aluminide coatings on Al2024 alloy. Mater. Sci. Technol. 2013, 29, 314–318.
[CrossRef]
17. El Cheikh, H.; Courant, B.; Hascoet, J.; Guillen, R. Prediction and analytical description of the single laser
track geometry in direct laser fabrication from process parameters and energy balance reasoning. J. Mater.
Process. Technol. 2012, 212, 1832–1839. [CrossRef]
18. Caiazzo, F. Laser-aided Directed Metal Deposition of Ni-based superalloy powder. Opt. Laser Technol. 2018,
103, 193–198. [CrossRef]
19. Angelastro, A.; Campanelli, S.L.; Casalino, G. Statistical analysis and optimization of direct metal laser
deposition of 227-F Colmonoy nickel alloy. Opt. Laser Technol. 2017, 94, 138–145. [CrossRef]
20. Gao, R.; Wang, L.; Teti, R.; Dornfeld, D.; Kumara, S.; Mori, M.; Helu, M. Cloud-enabled Prognosis for
Manufacturing. CIRP Ann. Manuf. Technol. 2015, 64, 749–772. [CrossRef]
21. Alpaydin, E. Introduction to Machine Learning; MIT Press: Cambridge, MA, USA, 2014.
22. Bishop, C.M. Neural Networks for Pattern Recognition; Clarendon Press: Oxford, UK, 1995.
23. Abe, S. Pattern Classification: Neuro-Fuzzy Methods and Their Comparison; Springer: London, UK, 2001.

© 2018 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/).