Measurement 85 (2016) 88–99

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Optimization of micro milling of hardened steel with different

grain sizes using multi-objective evolutionary algorithm
Carlos Henrique Lauro a, Sérgio Luiz Moni Ribeiro Filho b, Denison Baldo c,
Sérgio Augusto Araújo da Gama Cerqueira b, Lincoln Cardoso Brandão b,⇑
a
Dept. of Mechanical Engineering, University of Aveiro, Campus Santiago, 3810-193 Aveiro, Portugal
b
Federal University of São João del Rei, Department of Mechanical Engineering, Center for Innovation in Sustainable Manufacturing, São João del Rei, Brazil
c
Federal Institute of Technological Education of Minas Gerais, Rua Bernardo Mascarenhas, 1283 Juiz de Fora, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Modern manufacturing processes need high production rates, low costs, and high product
Received 20 October 2015 quality. Generally, surface roughness is a good reference to determine the performance in
Received in revised form 6 February 2016 machined products. The use of optimization systems can determine the optimum machin-
Accepted 10 February 2016
ing parameters in the machining process, especially in milling operations. The present
Available online 24 February 2016
study integrates the least square model based on feed rate, cutting speed, and grain size
with a genetic optimization algorithm to provide the optimal process parameter. The
Keywords:
NSGA II algorithm was applied due to its coverage and easily to optimize the micro milling
Genetic algorithm
Micro milling
of hardened steel. The responses were Fy Force and Mz Torque. The results show that the
Least-square method feed rate was the most significant factor for minimizing Fy force and Mz Torque.
Grain size Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction set is the main factor that will define the high quality of the
products. Generally, surface roughness is a good reference
The DIN 1.2344 steel is one of the materials widely used to determine the performance of machined products.
to produce dies and molds in the casting of non-ferrous According to Mukherjee and Ray [17], traditional tech-
and ferrous products [4]. The DIN 1.2344 steel shows great niques, such as Taguchi and RSM, are commonly used to
hardness after thermal treatment and it have good ability optimize machining processes. However, modern tech-
to maintain the hardness and strength when submitted niques that are considered as non-conventional
at great temperatures during the hot work [11]. The com- approaches, such as genetic algorithms (GA), particle
position of DIN 1.2344 steel can also vary in order to pro- swarm optimization, and mixture designs have been used
vide special properties in specific applications. as an alternative for estimating the optimal result in
Manufacturing processes seek to obtain three main machining [23,12,29].
objectives: high production rates, low costs, and high pro- Several researchers have developed studies on micro
duct quality [29]. The big challenge in the industrial sector machining specially in milling with focus on size effect,
is reach an optimum point in manufacturing processes that cutting force, and surface roughness. Saedon et al. [21]
meets these three objectives. The first point that should be used a Response Surface Methodology to define the opti-
mum input parameters to predict the tool wear, and tool
⇑ Corresponding author. life. According to the authors, the RSM was a good tool to
E-mail addresses: [email protected] (C.H. Lauro), sergiolmrf@gmail. estimate the region of maximum tool life. The total under-
com (S.L.M. Ribeiro Filho), [email protected] (D. Baldo), standing of cutting mechanism in micro machining is not
[email protected] (S.A.A.d.G. Cerqueira), [email protected] yet complete and need of many future studies to fill the
(L.C. Brandão).

http://dx.doi.org/10.1016/j.measurement.2016.02.011
0263-2241/Ó 2016 Elsevier Ltd. All rights reserved.
 C.H. Lauro et al. / Measurement 85 (2016) 88–99 89

gaps of process. Arruda et al. [3] used a mixture design to Maiyar et al. [14] used the Taguchi technique based on
optimize the surface roughness in the milling of molds grey relational analysis to predict the optimal cutting
with different strategies. The design was capable of define parameters during the milling of Inconel 718 alloy. This
the optimum input parameters. optimization technique was capable of reduce simultane-
The need for miniaturized components and devices of ously, the surface roughness in 9.5% and increase the mate-
high complexity contributes to the development of micro- rial removal rate in 65%. Manufacturing processes, mainly
fabrication in a variety of materials including materials the machining processes can be optimized using artificial
with different grain sizes. Therefore, the reduction in techniques. However, turning and milling are the main
dimension means that cutting phenomena and mecha- processes studied together with special materials such as
nisms imply variations in cutting power and forces in the Inconel, titanium alloy, duplex stainless steel, and nickel
micro process. In this case, monitoring investigations have alloys.
been helpful in the conduction in this problem [7,20]. Koyee et al. [12] studied the turning of duplex stainless
However, in machining processes, the milling of macro steel of the grades EN 1.4462 and EN 1.4410 DEDM using
components is completely different from micro milling the response surface methodology and multilayer percep-
components. The effect size during the cutting mechanism tron artificial neural network. According to the authors,
is complex, not only because of the microstructure of the after the three phases, the machinability of EN 1.4462
materials, but also due to the perfect sharpness of the cut- was higher than the machinability of EN 1.4410 and the
ting edge. Unlike the macro milling, in micro milling the use of lower cutting speed, intermediate feed rate, and
main source of compliance can account for up to 80–90% lower depth of cut ranges tend to maximize the opera-
of the total compliance at the tool tip [25]. Oliveira et al. tional sustainability index in dry and wet cutting.
[19] studied the size effect and minimum chip thickness However, there are few studies on optimization in
in micro milling in AISI 1045 steel. According to the micro machining processes such as turning and milling.
authors, the cutting section area has great influence in For this reason, this study presents an optimization tech-
the size effect phenomenon and in the minimum uncut nology to minimize force and torque in the micro milling
chip thickness. Moreover, studies on cutting forces are process. Workpieces of hardened DIN 1.2344 steel were
essential to define the stability in micro milling because used in micro milling process and a Genetic Algorithm
the increase of cutting forces provides a decrease in the (GA) was applied to find the best-input parameters, focus-
stability during the process, mainly for small edge radii ing on minimal Fy force and Mz torque. The Fy force was
and high-uncut chip thickness [1]. used as reference because the displacement of the cutter
Thus, the application of optimization systems during was carried out in Y direction generating a groove with
the machining process can be a powerfully tool to define the same width of cutter.
input parameters and predict the responses such as surface
roughness, cutting forces, burr formation, and tool wear, 2. Materials and methods
especially in milling operations. Genetic Algorithms have
been widely used in several manufacturing sectors to opti- 2.1. Parameter optimization using GA combined with the
mize the machining parameters and improve the efficiency Least-Square Method – LSM
of the processes. Ganesan and Mohankumar [9] studied the
genetic algorithm to provide an improvement on program- The development of GA was based on the Least Square
ing of CNC machine centers. According to the authors, sev- Method (LSM) mathematic model. The LSM searches the
eral production strategies could be optimized using the best fit into a data group, minimizing the sum of the
non-dominated genetic algorithm NSGA-II. Authors as squared differences between the estimated and observed
Zhuang et al. [31] have studied the balance of tools in values (generally, this difference is called residual). The
milling processes. They used a model complex, nonlinear, least-square method can be linear or non-linear, and the
and multi-variable with an objective function to maximize response is a linear function of the independent variables.
the material removal rate (MMR). This way, it was possible Thus, the relation between ‘‘y” and the three-predictor
to define the optimum input parameters and the allowable variables can be written as a Taylor series.
range of machining conditions considering the stability
y ¼ b0 þ b1 X 1 þ b2 X 2 þ . . . þ bk X k þ e ð1Þ
and cutting speed of the process.
Yildiz [28] used a hybrid approach based on the differ- When the database has 3 input variables and 27 obser-
ential evolution algorithm to optimize cutting parameters vations, the model can be written in matrix form, as
in milling operations. According to the author, the opti- follows.
mization method found better milling parameters solu- 0 1
0 1 0 1 b10 1
tions compared to other approaches in the references. y1 1 x11 x21    x271 Bb C e1
The method can be easily applied to several milling con- B C B CB 2 C B C
@ y2 A ¼ @ 1 x12 x22    x272 A:B C þ @ e2 A
straints and considering various objectives. Commonly, @  A
optimization techniques are not only constrained by tech-
y3 1 x13 x23    x273 e3
b27
nological parameters, such as surface roughness or cutting
force, but also by costs or production times. However, where xji is the value of the jth variable of the ith observa-
when non-technological constraints are used, the opti- tion, which can be written in the short form as in Eq. (2)
mization should be tightly linked to the quality of product. y¼Xbþe ð2Þ
90 C.H. Lauro et al. / Measurement 85 (2016) 88–99

The solution of LSM can be found using the minimiza-

P
tion of the sum of the square of the errors ni¼1 e2i and it
should be rewritten as e e. Thus, replacing e with ðy  XbÞ,
0

we get the following Eq. (3).

0
SðbÞ ¼ ðy  XbÞ  ðy  XbÞ ð3Þ
The minimization occurs when the derivation S(b) is
carried out in relation to (b) and when it is equal to zero,
instead of can be seen in Eq. (4):
1
b ¼ ð2X 0  yÞ  2X 0  Xb ð4Þ
NSGA II is a very famous multi-objective optimization
algorithm and it was applied in this work because of its
coverage and strength [8]. The objective function was the Work piece
optimum combination of the input parameters Fy force f
and torque. The constraints were defined according to
the limit of the high-speed machining head with Fig. 2. Finite element model for explicit simulation with deformation.
50,000 rpm and 300 W coupled to a machining center.
The constraints were based on the high-speed machining
Table 1
head because experimental tests were carried out to vali-
Cutting parameters used in simulation tests.
date the LSM. The constraints are shown below (see Fig. 1).
Test Spindle Angular velocity Feed Feed rate
number (rpm) (rad/s) (mm/rot) (m/s)
 0.001 (mm/rot) < Feed rate < 0.005 (mm/rot);
 15 (mm/rot) < Cutting speed < 77 (mm/rot); 1 21,000 2199.11 0.001 0.70  103
3.50  103
 10 (lm) < Grain Size < 1000 (lm). 2 21,000 2199.11 0.005
3 49,000 5131.27 0.001 1.63  103
4 49,000 5131.27 0.005 8.16  103
Thus, 1000 individuals who evolved during 1000 inter-
actions formed 1,000,000 generations generating the initial according to the Johnson-Cook model. The numerical tests
population of the model. The LSM of Fy Force and Torque were carried out according to the Johnson-Cook thermo-
were integrated with a genetic optimization to provide plastic flow for both mesh size [18]. The tool was consid-
optimal process parameters. Fig. 2 exhibits the integrated ered a rigid body that did not suffer deformations or
optimization scheme. damage. The friction coefficient in the cutting region
between tool and workpieces was 0.1, according to Woon
2.2. Preliminary experimental design and Finite Element et al. [26]. The acquisition rate for numerical and experi-
analysis mental tests was the same, and the responses were Fy cut-
ting force and Mz torque.
The experimental tests were carried out using a micro To reduce the computational effort, the tests were con-
cutter of 0.5 mm in diameter to generate grooves in the ducted in a two-dimensional space using a simulation time
workpieces of hardened DIN 1.2344 steel. The Mz torque of 0.05 s. In addition, four displacements for the tool were
and Fy force were modeled in the micro milling process used, and these displacements were performed in meters
using a finite-element method and the design of experi- per second, corresponding to the feed rate. Two angular
ments, as can be seen in Table 1. speeds were also applied for the tool in radians per second.
The numerical tests used type CPE4R square elements The values used were converted in the International Sys-
with four nodes to generate the mesh. The mesh size was tem units because of the required standardization of the
correlated with the experimental grain size. This means AbaqusTM software.
that the mesh sizes of 50 and 400 lm were used. The The micro milling process was simulated with a
workpieces were considered deformable and isotropic dynamic mechanical finite element using an explicit inte-
gration. The simulating of chip formation was carried out
Feed rate
Roughness with a strain gradient for FE models. A description more
detailed for the strain gradient in plasticity model can be
Least Y Force
found in the works of Liu and Melkote [13]. Tables 2 and
Cutting Speed Square
Torque 3 exhibit the Johnson-Cook model for hardened materials
Model such as the DIN 1.2344 steel with hardness of 49 HRC,
Grain Size Burr Length according to Ng and Aspinwall [18] and Yan et al. [27].

2.3. Experimental tests

Optimization NSGA II
Requirement A high-speed machining head with 50,000 rpm and
300 W of main power was coupled to a machining center.
Fig. 1. Integrated scheme for optimization. The use of the high-speed machining head provided an
 C.H. Lauro et al. / Measurement 85 (2016) 88–99 91

Table 2 measuring cutting forces in Fy Force and Mz Torque. The

Johnson Cook’s constants [6]. dynamometer was set up with the sensibility of 26 pC/N
A – Initial Von Mises Stress (MPa) 715 for micro the milling experiments, according to Afazov
B – Modulus of strain hardening (MPa) 329 et al. [2]. The frequency was 6.5 kHz, because it should
C – Strain rate sensitivity 0.03 be at least four times as large as the natural frequency of
m – Coefficient of thermal expansion 1.13
n – Cold work rate 1.5
the dynamometer, according to Shaw [22].

Johnson Cooḱs parameter [18].

D1 – Johnson Cook’s parameter 0.8 2.4. Genetic algorithm
D2 - Johnson Cook’s parameter 2.11
D3 - Johnson Cook’s parameter 0.5
Genetic algorithms are computational models that use
D4 - Johnson Cook’s parameter 0.0002
D5 - Johnson Cook’s parameter 0.61
Darwin’s theory as an adaptive process of optimization.
The concepts of genes, chromosomes, crossover, mutation,
and natural selection combined with computational tech-
niques provide great adaptability and compatibility to
Table 3
Parameters of the material (DIN 1.2344 steel) [27]. search for optimum solutions. Based on this heuristic
method, the evolution starts in a solution that is created
Specific heat (J/kg K) 560
randomly and is carried to several later generations. In
Coefficient of thermal expansion (N/s/mm/C) 40
Thermal conductivity (W/m K) 37 each generation, the adaptation of every solution in the
Density (kg/m3) 7800 population is evaluated. Some individuals are selected for
Young’s modulus (GPa) 211 the subsequent generation and recombined or mutated to
Poisson’s ratio 0.28
form a new population. The new population is then used
as input to the next iteration of the algorithm.
Genetic Algorithms belong to the class of the probabilis-
increase in spindle speed and, consequently, it generated tic methods used in optimization techniques. Although
high cutting speeds. The displacement of the tool was GAs are not random, the concept of probability can be used
achieved by movement over the X, Y, and Z axes of the because the GAs seek out regions in space where optimal
machining center. Workpieces of hardened DIN 1.2344 points are likely.
steel with dimensions of 11  11  11 mm were used in Evolutive algorithms differ from the traditional opti-
the tests. The grinding process was applied previously dur- mization algorithm basically in four aspects, according to
ing the manufacture of the workpieces to ensure the accu- Srinivas and Patnaik [24], Goldberg [10], and Michalewicz
racy and a perfect assembly on the fixture device. The [16], namely:
chemical composition of the workpieces was 0.40% C,
1.00% Si, 0.35% Mn, 5.20% Cr, 1.50% Mo, 0.90% V, and Fe  They are based on a space with encoded solutions, not
in balance. The workpieces were submitted to different on a direct search space.
heat treatments to produce two grain sizes. Fig. 3 shows  The results are presented as a set of solutions rather
a layout of the experimental tests. than as a single solution.
Table 4 shows the two grain sizes that were obtained  They do not require any knowledge derived from the
after heat treatment, where the austenite grain size was problem.
measured in micrometers according to the ASTM standard.  They use probabilistic transitions and not deterministic
The values for cutting speed, feed rate, axial depth of rules.
cut, and radial depth of cut were defined according to the
manufacturer of the tools. The dynamometer used in the
2.4.1. Individual
tests was a KistlerTM model 9272, which is capable of
The individual is an abstract representation of the
genetic code. The genetic code is a spatial representation
of the problem to be solved, generally in binary code. Using
High-speed genetic metaphors, each sequence of the solution corre-
Special device for machining head sponded to a chromosome and each element of the solu-
assembly of the tion is equivalent to a gene. The gene position
work piece corresponds to a position of the gene into the chromo-
Feed some; that is, the index within the sequence. The genetic
code should be a rearrangement with the ability to show
Dynamometer
the whole set of possible values in the search space, and
it should have a finite size.
Work piece
2.4.2. Selection
The procedure of ordination consisted of classifying the
individuals ‘‘S” from a group in several levels
N i1 ; N i2 ; . . . ; N id according to the degree of dominance of
the individuals where ‘‘d” is the number of levels and
Fig. 3. Layout of experimental tests. ‘‘Ni” the number of dominance. The algorithm used a
92 C.H. Lauro et al. / Measurement 85 (2016) 88–99

Table 4
Grain size after heat treatment.

Workpiece Grain size (lm) Grain size (ASTM) Hardness (HRC) Temperature (°C) Time (s) Cooling
G1 497 0 44 1250 7200 Oil
G2 40 6 46 1025 2700

selection process based on a tournament. Thus, considering Nm ¼ pm  nc  N ð7Þ

the residue of the relative average fitness, the ranking of
where ‘‘nc” is the size of the individual, ‘‘pm” is the proba-
each individual ‘‘S” depends on the frontier N id to which
bility of the mutation, and ‘‘N” is the number of individuals
it belongs and it is away from the crowd (crowdistS). In this
in the population. This work used a crossover probability of
approach, two solutions are compared in order to choose
90% and a mutation of 10%. Considering the probabilities of
which one will generate offspring in the new population.
recombination and mutation, empirical studies have
A solution ‘‘i” is chosen over a solution ‘‘j” if:
shown that good results are usually obtained with a high
value of recombination (P0.70) and a low value for muta-
 ‘‘i” has a lover ranking than ‘‘j”, in other words,
tion (P0.10).
ranki < rankj.
 If both solutions have the same ranking and ‘‘i” has a
greater distance value from the crowd, in other words, 2.5. Adequacy of the LSM model
ranki < rankj and crowdisti > crowdistj.
The technical statistical analysis of variance (ANOVA)
The distance calculation allows the solutions dispersed was used to confirm the adequacy of the LSM models with
in the crowd to be prioritized, ensuring the diversity of a confidence interval of 95%. The analysis of variance
solutions. The new generation is sourced using the selec- (ANOVA) was carried out to verify whether the main
tion operators per tournament, recombination and muta- and/or the interaction factors were statistically significant.
tion in the N idþ1 population. The crowd distance is the The P-values in Table 3 indicate which of the effects
average distance of two adjacent individuals to each indi- throughout the system are statistically significant based
vidual in the population for all objectives, according to on examination of the experimental data from replicates.
Eq. (5) below: The effect is considered significant when the P-value is less
than or equal to 0.05. A a-level of 0.05 is the level of signif-
crowdists ¼ crowdists þ f obj ðS þ 1Þ  f obj ðS  1Þ; for S icance, which implies that there is 95% of probability of the
¼ 1 until N  1 ð5Þ effect being significant. In addition, the results are pre-
sented using main effect and interaction plots. Table 5
shows a summary of P-values for Fy Force and Mz Torque.
2.4.3. Recombination and mutation R2 (adj) is defined as the ratio of the sum of squares
The recombination is a process that involves combining explained by a regression model and the total sum of
two or more individuals. This process simulates the cross- squares around the mean in the ANOVA. Therefore, the
over phenomenon that is the change of fragments between regression equation can be explained by measuring of the
paired chromosomes. These pairs are selected randomly by percentage of variability in the responses (Fy Force and
the crossover probability ‘‘Pc” and the individual descen- Mz torque) that is explained by predicting variables (Cut-
dants are generated by recombination. The place of recom- ting speed, feed rate, and grain size). Based on this, the val-
bination is selected randomly and the crossover number ues for R2 (adj) were 83.20% and 91.9% for the models of Fy
‘‘Nc” is defined according to Eq. (6) below. Force and Mz Torque. Accordingly, the mathematical mod-
els for the prediction of Fy force and Mz torque are pre-
sented below.

Fy Force ½N ¼ 0:224  0:00059  Cutting Speed ½m=min þ 40:1  Feed Rate ½mm=rev:  0:000189  Grain Size ½lm ð8Þ
Torque ½Nm ¼ 0:0505  0:000136  Cutting Speed ½m=min  6:12  Feed Rate ½mm=rev:  0:000015  Grain Size ½lm ð9Þ

N  PC 2.6. Accuracy of the models

NC ¼ ð6Þ
2
where ‘‘N” is the number of individuals of the population Mathematical models of the LSM were used to predict
and ‘‘Pc” is the probability of crossover. The mutation pro- Fy force and Mz Torque of the micro milling process for
cess selects a position in a chromosome and changes the each variation of the responses cutting speed, feed rate,
value of the genetic representation randomly to another and grain size. The theoretical values were compared with
possible allele. The number of mutations can be written the experimental values for Fy force and Mz Torque. Table 6
according to Eq. (7). shows the experimental and predictive values.
 C.H. Lauro et al. / Measurement 85 (2016) 88–99 93

Table 5 Table 7
Analysis of variance for micro milling models. Percentage error for experimental and simulated values of micro milling
models.
Response Degree of Square Mean P-
freedom Sum Square value Run Cutting Feed Grain Percentage error
speed rate size
Fy force (N) 3 1.8594 0.3719 0.031 Fy (%) Torque
Mz torque 3 2.625 0.525 0.042 (%)
(N m)
1 33 0.001 39.9 1.790 11.364
2 33 0.001 497.0 0.000 2.273
3 77 0.005 39.9 3.325 4.545
4 77 0.005 497.0 7.417 15.909
The average error rate of these models with 8 sets in the
5 33 0.005 39.9 1.535 6.818
LSM was of 2.297% and 7.386% for Fy force and Mz Torque 6 33 0.005 497.0 1.023 6.818
models. Table 7 shows the percentage accuracy for micro 7 77 0.001 39.9 1.023 0.000
milling models. 8 77 0.001 497.0 3.069 11.364
For the validation of the tested data, maximum errors of
7.417% and 15.909% were considered. Thus, the experi-
mental values followed the same trend models as the
LSM. The percentage accuracy calculated by:
!
yexp  ytheoretical
D¼ :100 ð10Þ
yexp

where yexp is the experimental value and ytheoretical is the

predicted value.

3. Results and discussion

A computer program was developed to simulate a GA-

based optimization of micro milling processes using LSM
models. The optimal force in the Y direction during the
micro milling tests was 73.7–75 mm/min for cutting speed,
0.0009–0.0001 mm/rev for feed rate and 900–1000 lm for
grain sizes, as can be seen in Fig. 4. Fig. 5 reveals that the Fig. 4. Detail of the measuring of the burr length.
requirements for cutting speed and grain size are high (in
the range of 73.5–75 m/min and 900–1000 lm) in order
to minimize the Y force. On the other hand, the optimal
values for feed rate are low (in the range of 0.0009– In the same way, a high cutting speed decreases the cut-
0.0001 mm/rev). ting forces. According to Zaman et al. [30], peak-to-peak
The values found in the GA simulation are in close agree- values of cutting forces can vary from 0.0024 to 7.057 N
ment with the experimental tests carried out by Chae and in the milling of pre-hardened steel with an average hard-
Park [5] where the peak-to-peak values of the reference ness of 32.5 HRC and using a micro cutter of 1 mm in diam-
forces are approximately 0.2 N. In this work, the range of eter. Afazov et al. [2] developed a study to predict the
values found for force in the Y direction was 0.09–0.39 N. relation between the cutting forces, uncut chip thickness,
Based on this, it can be supported that lower cutting forces and cutting speed. The uncut thickness also occurs in the
occurred with feed rates below 0.0049 mm/rev and that the macro-milling process. However, its influence on this pro-
optimal point was less than 0.0019 mm/rev. cess has little influence because of the scale factor.

Table 6
Comparison of experimental and simulated values of micro milling models.

Run Cutting speed Feed rate Grain size Experimental values Simulated values
Fy force (N) Torque (N m) Fy force (N) Torque (N.m)
1 33 0.001 39.9 0.230 0.044 0.237 0.039
2 33 0.001 497.0 0.151 0.033 0.151 0.032
3 77 0.005 39.9 0.385 0.007 0.372 0.009
4 77 0.005 497.0 0.256 0.009 0.285 0.002
5 33 0.005 39.9 0.391 0.012 0.397 0.015
6 33 0.005 497.0 0.315 0.005 0.311 0.008
7 77 0.001 39.9 0.215 0.033 0.211 0.033
8 77 0.001 497.0 0.137 0.021 0.125 0.026
94 C.H. Lauro et al. / Measurement 85 (2016) 88–99

Fig. 5. Relationship between feed rate, cutting speed, and grain size for Force in Y direction.

-3
x 10
Optimal cutting speed (mm/rot)

80 2.5
Optimal feed rate (mm/rev)

60
2

40

1.5
20

0 1
0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Optimal Y Force (N) Optimal Y Force (N)

1000
Optimal grain size (mm)

900

800

700

600
0.04 0.05 0.06 0.07 0.08 0.09 0.1

Optimal Y Force (N)

Fig. 6. Optimal Fy Force for cutting speed, grain size and feed rate.

The micro-milling forces proposed by the authors were showed very good agreement between predicted and trial
determined based on the finite-element method and con- tests. Furthermore, it can be stated that the best option
firmed experimentally. The results showed that the cutting to minimize the cutting force in micro milling is use low
forces in micro-milling of AISI 4340 steel were 0.8 N for the feed rate and high cutting speeds. Thus, the prediction of
simulation and 0.7 N for the experimental tests. The results cutting forces with an auxiliary methodology, such as
 C.H. Lauro et al. / Measurement 85 (2016) 88–99 95

Fig. 7. Relationship between feed rate, cutting speed, and grain size for Torque.

-3
x 10
Optimal cutting speed (mm/rot)

80 5
Optimal feed rate (mm/rev)

70

60 4.5

50

40 4

30

20 3.5
0 0.5 1 1.5 2 2.5 3 3.5 0 0.5 1 1.5 2 2.5 3 3.5
-3 -3
Optimal Torque (Nm) x 10 Optimal Torque (Nm) x 10

1000
Optimal grain size (mm)

900

800

700

600

500

400
0 0.5 1 1.5 2 2.5 3 3.5
-3
Optimal Torque (Nm) x 10

Fig. 8. Optimal Y Force for cutting speed, grain size and feed rate.

genetic algorithms, can provide good information and a spindle speed of 50,000 rpm coupled in a machining cen-
avoid the break of the micro cutter before the beginning ter can provide an adequate solution. Therefore, the com-
of the micro-milling process. mercial high-speed machining head used in this study
Figs. 6 and 7 exhibit the optimum point in the micro- had a torque of 0.061 N.m and a drive power of about
milling process considering torque. Commonly, mechanical 300 W.
or hydraulic collets with a diameter of 6 mm should be One should bear in mind that electrical or pneumatic
considered in micro-milling. In addition, the micro- systems are applied to operate high speed machining
cutters have diameters lower than 0.5 mm. In this range, heads, and that they are supplied with low power from
96 C.H. Lauro et al. / Measurement 85 (2016) 88–99

Fig. 9. Relationship between feed rate, cutting speed, and grain size for burr length.

-3
x 10
Optimal cutting speed (mm/rot)

80 5
Optimal feed rate (mm/rev)

75
4
70
3
65
2
60

55 1
360 370 380 390 400 410 420 360 370 380 390 400 410 420
Optimal length burr (µm) Optimal length burr (µm)

1000
Optimal grain size (mm)

950

900

850

800

750

700
360 370 380 390 400 410 420
Optimal length burr (µm)

Fig. 10. Optimal burr length for cutting speed, grain size and feed rate.

the main motor. It is important to understand not only the generations from distinctive ranges considering the input
operating mechanism of high-speed head machining, parameters. However, as can be seen in Fig. 8, the conver-
but also its power in order to define input parameters gence profile of the solution indicates that it has reached
according to torque of the device. The parameter the different number of generations for the fitness of the
levels that optimize Torque are as follows: feed rate input parameters. The fitness for burr length, Fy force and
(0.0045–0.005 mm/rev), cutting speed (73.7–75 m/min) Mz Torque occurred in generation G = 24, G = 20.
and grain size (900–1000 lm). Generation represents the evolution of the individual
Fig. 8 exhibits the improved fitness curve for Fy force and that produces a set of descendants that will constitute
Mz Torque. The best solution was applied in 100 consecutive the next generation. According to Martínez-Romo et al.
 C.H. Lauro et al. / Measurement 85 (2016) 88–99 97

Fig. 11. Relationship between feed rate, cutting speed, and grain size for surface roughness Ra.

-3
x 10
Optimal cutting speed (mm/rot)

80 4
Optimal feed rate (mm/rev)

3.5
60
3

40 2.5

2
20
1.5

0 1
0.04 0.045 0.05 0.055 0.06 0.065 0.04 0.045 0.05 0.055 0.06 0.065

Optimal Surface roughness Ra (µm) Optimal Surface roughness Ra (µm)

1000
Optimal grain size (mm)

950

900

850

800
0.04 0.045 0.05 0.055 0.06 0.065
Optimal Surface roughness Ra (µm)

Fig. 12. Optimal surface roughness Ra for cutting speed, grain size and feed rate.

[15] fitness generation refers to the derivation of features solution at the generation numbers ranging from 20 to
from input parameters, in this case feed rate and cutting 24. Accordingly, this result can be ascribed to the fact that
speed, which are favorable in terms of class segregation the algorithm proposed in this work is consistent with a
in the feature space. fitness of the optimal points for feed rate and cutting
Thus, as GAs seek to produce a desirable solution for the speed.
problem, it can be supported that, evaluating a preselected Finally, Fig. 9 shows the analysis of variance for cutting
number of iterations, the last generation will be the most speed (mm/min), feed rate (mm/rev) and austenitic grain
optimal solution for the problem. It is noteworthy that size (um) for Fy Force and Mz Torque. According to Fig. 9,
the model proposed in this work produced an ideal the feed rate is the most significant factor of influence on
98 C.H. Lauro et al. / Measurement 85 (2016) 88–99

357.5 -0.12

Fitness burr length (mm)

Fitness Y Force (N)

357 -0.125

356.5
-0.13
356
-0.135
355.5
-0.14
355
0 20 40 60 80 100 0 20 40 60 80 100
Generation number Generation number
-3
x 10
-4

Fitness Roughness (Ra)

Fitness Torque (Nm)

0.039
-4.5
0.038

-5 0.037

0.036
-5.5

0 20 40 60 80 100 0 20 40 60 80 100
Generation number Generation number

Fig. 13. Fitness improvement curve.

Fig. 14. Sensibility analysis for micro milling process.

Torque (N m) and Fy Force (58% and 51.2%, respectively) The results showed that the increase of grain size pro-
(see Figs. 10–14). vided the decrease of cutting force, which affected the
responses, mainly when milling the hardened DIN 1.2344
steel. In addition, the variation of the austenitic grain size
4. Conclusions was of influence not only on the physical properties of
the materials, but also on the scale of miniaturization of
A Genetic algorithm is an optimization technique mechanical components. Moreover, the grain size may be
inspired by evolutionary behavior observable in nature. of significant influence on process of micro machining.
GAs have shown to be an efficient and robust optimization In general, hardened AISI H13 steel exhibited different
algorithm in multiobjective optimization problems (MOPs). cutting conditions to optimize the responses. In order to
In this context, LSM models integrated with genetic opti- minimize the Fy Force in micro milling of DIN 1.2344 steel,
mization were proposed in order to obtain a range of solu- the feed rate, cutting speed and grain size must be set at:
tions that provide useful information to the user during the 0.001–0.0015 mm/rev, 73.7–75 mm/min and 900–1000 lm,
selection of machining parameters. respectively. The same parameters could optimize the Mz
 C.H. Lauro et al. / Measurement 85 (2016) 88–99 99

Torque in micro milling, except for the feed rate, which [15] J.C. Martínez-Romo, F.J. Luna-Rosas, M. Mora-González, C.A. Luna-
Ortega, V. López-Rivas, Optimal feature generation with genetic
must be set at 0.0045–0.005 mm/rev. It can be supported
algorithms and FLDR in a restricted-vocabulary speech recognition
that, the feed rate was the most significant factor in order system, Bio-Inspired Comput. Algorithms Their Appl. 420 (2012).
to minimize Fy Force and Mz Torque. [16] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution
Programs, Springer-Verlag, 1992.
[17] I. Mukherjee, P.K. Ray, A review of optimization techniques in metal
cutting processes, Comput. Ind. Eng. 50 (2006) 15–34.
References [18] E. Ng, D.K. Aspinwall, Modelling of hard part machining, J. Mater.
Process. Technol. 127 (2002) 222–229.
[1] S.M. Afazov, D. Zdebski, S.M. Ratchev, J. Segal, S. Liu, Effects of micro- [19] F.B. de Oliveira, A.R. Rodrigues, R.T. Coelho, A.F. de Souza, Size effect
milling conditions on the cutting forces and process stability, J. and minimum chip thickness in micromilling, Int. J. Mach. Tools
Mater. Process. Technol. 213 (2013) 671–684. Manuf. 89 (2015) 39–54.
[2] S.M. Afazov, S.M. Ratchev, J. Segal, Modeling and simulation of [20] A. Rivero, L.N.L. de Lacalle, M.L. Penalva, Tool wear detection in dry
micro-milling cutting forces, J. Mater. Process. Technol. 210 (15) high-speed milling upon the analysis of machine internal signals,
(2010) 2154–2162. Mechatronics 18 (2008) 627–633.
[3] E.M. Arruda, L.C. Brandão, S.L.M. Ribeiro Filho, J.A. Oliveira, [21] J.B. Saedon, S.L. Soo, D.K. Aspinwall, A. Barnacle, N.H. Saad, Prediction
Integrated optimization using mixture design to confirm the and optimization of tool life in micro milling AISI D2 (62HRC)
finishing of AISI P20 using different cutting strategies and ball nose hardened steel, international symposium on robotics and intelligent
end mills, Measurement 47 (2014) 54–63. sensors 2012 (IRIS 2012), Proc. Eng. 41 (2012) 1674–1683.
[4] J.C. Benedyk, Ferrous FeUH – H13, High Performance Alloys Dabase [22] M.C. Shaw, Metal Cutting Principles, 2nd ed., Oxford University
(2008). Press, USA, 2005. 672.
[5] J. Chae, S.S. Park, High frequency bandwidth measurements of micro [23] S.P. Silva, S.L.M. Ribeiro Filho, L.C. Brandão, Particle swarm
cutting forces, Int. J. Mach. Tools Manuf. 47 (2007) 1433–1441. optimization for achieving the minimum profile error in honing
[6] L. Chen, T.I. El-Wardany, M. Nasr, M.A. Elbestawi, Effects of edge process, Precision Eng. 38 (2014) 759–768.
preparation and feed when hard turning a hot work die steel with [24] M. Srinivas, L.M. Patnaik, Genetic Algorithms: A Survey, Computer
polycrystalline cubic boron nitride tools, CIRP Annals – Manuf. 27 (6) (1994) 17–26.
Technol. 55 (2006) 89–92. [25] L. Uriarte, A. Herrero, M. Zatarain, G. Santiso, L.N.L. de Lacalle, A.
[7] L.N.L. de Lacalle, A. Lamikiz, J.A. Sánchez, I.F. Bustos, Simultaneous Lamikiz, J. Albizuri, Error budget and stiffness chain assessment in a
measurement of forces and machine tool position for diagnostic of micromilling machine equipped with tools less than 0.3 mm in
machining tests, IEEE Trans. Instrum. Meas. 54 (2005) 2329–2335. diameter, Precison Eng. 31 (2007) 1–12.
[8] K. Deb, A Fast and elitist multiobjective genetic algorithm: NSGA-II, [26] K.S. Woon, M. Rahman, F.Z. Fang, K.S. Neo, K. Liu, Investigations of
IEEE Trans. Evol. Comput. 6 (2) (2002) 182–197. tool edge radius effect in micromachining: a FEM simulation
[9] H. Ganesan, G. Mohankumar, Optimization of machining techniques approach, J. Mater. Process. Technol. 167 (2007) 316–337.
in CNC turning centre using genetic algorithm, Arab. J. Sci. Eng. 38 (6) [27] H. Yan, J. Hua, R. Shivpuri, Flow stress of AISI H13 die steel in hard
(2013) 1529–1538. machining, Mater. Des. 28 (2007) 272–277.
[10] D.E. Goldberg, Genetic Algorithms in Search. Optimization and [28] A.R. Yildiz, A new hybrid differential evolution algorithm for the
Machine Learning, Addison-Wesley, 1989. selection of optimal machining parameters in milling operations,
[11] P.D. Harvey, Engineering Properties of Steel: Hot Work Tool Steels, Appl. Soft Comput. 13 (2013) 1561–1566.
ASM International, 1982, pp. 457–462. [29] A.M. Zain, H. Haron, S. Sharif, Application of GA to optimize cutting
[12] R.D. Koyee, U. Heisel, R. Eisseler, S. Schmauder, Modeling and conditions for minimizing surface roughness in end milling
optimization of turning duplex stainless steels, J. Manuf. Process. 16 machining process, Exp. Syst. Appl. 37 (2010) 4650–4659.
(2014) 451–467. [30] M.T. Zaman, A.S. Kumar, M. Rahman, S. Sreeram, A three-
[13] K. Liu, S.N. Melkote, Finite element analysis of the influence of tool dimensional analytical cutting force model for micro end milling
edge radius on size effect in orthogonal micro-cutting process, Int. J. operation, Int. J. Mach. Tools Manuf. 46 (2006) 353–366.
Mech. Sci. 49 (5) (2007) 650–660. [31] K. Zhuang, X. Zhang, D. Zhang, H. Ding, On cutting parameters
[14] L.M. Maiyar, R. Ramanujam, K. Venkatesan, J. Jerald, Optimization of selection for plunge milling of heat-resistant-super-alloys based on
machining parameters for end milling of Inconel 718 super alloy precise cutting geometry, J. Mater. Process. Technol. 213 (2013)
using taguchi based grey relational analysis, Proc. Eng. 64 (2013) 1378–1386.
1276–1282.