See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/257775391

Application of artificial neural network and optimization algorithms for

optimizing surface roughness, tool life and cutting forces in turning operation

Article in Journal of Mechanical Science and Technology · May 2013

DOI: 10.1007/s12206-013-0327-0

CITATIONS READS

66 1,354

3 authors:

Farshid Jafarian Mehran Taghipour

Mahallat Institute of Higher Education, Mahallat, Iran University of Surrey
35 PUBLICATIONS 775 CITATIONS 29 PUBLICATIONS 451 CITATIONS

SEE PROFILE SEE PROFILE

Hossein Amirabadi
University of Neyshabur
35 PUBLICATIONS 508 CITATIONS

SEE PROFILE

All content following this page was uploaded by Farshid Jafarian on 31 May 2014.

The user has requested enhancement of the downloaded file.

 Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477
www.springerlink.com/content/1738-494x
DOI 10.1007/s12206-013-0327-0

Application of artificial neural network and optimization algorithms for optimizing

surface roughness, tool life and cutting forces in turning operation†
F. Jafarian1,*, M. Taghipour2 and H. Amirabadi1
1
Department of Mechanical engineering, University of Birjand, Birjand, Iran
2
Faculty of Electrical and computer engineering, University of Birjand, Birjand, Iran

(Manuscript Received June 18, 2011; Revised March 26, 2012; Accepted January 12, 2013)

----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Abstract

Our goal is to propose a useful and effective method to determine optimal machining parameters in order to minimize surface rough-
ness, resultant cutting forces and maximize tool life in the turning process. At first, three separate neural networks were used to estimate
outputs of the process by varying input machining parameters. Then, these networks were used as optimization objective functions.
Moreover, the proposed algorithm, namely, GA and PSO were utilized to optimize each of the outputs, while the other outputs would
also be kept in the suitable range. The obtained results showed that by using trained neural networks with genetic algorithms as optimiza-
tion objective functions, a powerful model would be obtained with high accuracy to analyze the effect of each parameter on the output(s)
and optimally estimate machining conditions to reach minimum machining outputs.
Keywords: Neural network (ANN); Surface roughness; Genetic algorithm (GA); Cutting forces; Particle swarm optimization (PSO); Tool life
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

ods, such as statistic, intelligent and analytical methods [4, 5].

1. Introduction
Among them, predictive models are capable of estimating
In recent years, industrial producers and manufacturers have complex relationships between machining input parameters
attempted to increase the efficiency, performance and accu- and corresponding outputs [6, 7]. Therefore, these methods
racy of machining operations. These attempts involve some significantly reduce the required experimental tests for predic-
activities such as increasing production rate, decreasing opera- tion outputs of the process. Artificial neural network (ANN),
tion costs and enhancing the quality of production. These ac- fuzzy logic (FL) and regression models are some instances of
tivities can be affected by a number of factors, such as ma- these methods [8, 9]. Among them, ANN is one of the most
chine tool conditions, tool geometry, workpiece material and well known methods which has utilized widely for presenting
also machining parameters [1]. Among these, machining pa- a predictive model of machining processes. In this regard,
rameters such as cutting speed, depth of cut and feed rate play several researchers have succeeded in estimating various out-
a significant role in machining quality as parameters that are puts of the machining processes at specified range of input
controlled by the user. Therefore, suitable selection of these parameters by using artificial neural networks [6, 10]. Hence,
parameters is necessary to reach optimal machining conditions this method can be used as an important and useful tool for
in order to improve production efficiency. reducing time and cost in industrial projects.
So far, several researchers have performed experimental in- Owing to the fact that, in industrial applications, producers
vestigations about the machining operations and evaluated the and engineers are interested in determining optimal input ma-
effect of machining parameters on the outputs of the process chining parameters in order to minimize or maximize out-
[2, 3]. In fact, they have attempted to find suitable machining put(s) of the process, using predictive model can not only
parameters in order to determine optimal conditions of the eliminate requirements of industries. To solve this problem, in
process. But, implementing numerous experimental tests for addition to the predictive methods, some researchers have
finding mentioned conditions is very time consuming and utilized optimization techniques in their investigations [8, 11].
costly. To solve this problem, some of researchers have at- Evolutionary algorithms such as genetic algorithm (GA) and
tempted to model the machining processes by various meth- particle swarm optimization (PSO) are among the most well
*
Corresponding author. Tel.: +98 09357578764, Fax.: +98 5612502160 known techniques which are properly used for optimization of
E-mail address: [email protected] machining processes [11]. Of course, these techniques are
†
Recommended by Associate Editor Song Min Yoo applied to the problems with one objective function of optimi-
© KSME & Springer 2013
1470 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477

zation. In this regard, some researchers have determined suit-

able machining parameters for optimizing one output of the
process by employing integration of predictive and optimizing
methods [8, 12]. It should be noted that although in some
cases, optimizing one specified output of the process is re-
quirement of industries, in several cases machining operation
is evaluated from several aspects and not limited to one pur-
pose. Therefore, some researchers are employing multi objec-
tive optimization (MOO) for simultaneous optimization of (a) (b)
two or more outputs of the process [4, 13].
As mentioned above, some advantages and requirements for Fig. 1. (a) Views of tool flank wear; (b) View of main cutting forces in
turning process.
using the aforementioned methods have been explained. But,
implementing predictive and optimization methods in machin-
ing applications involves some limitations and deficiencies. mentioned outputs separately, a new method is implemented
More in detail, as far as predictive models (such as ANNs) are for optimizing each of specified outputs while other outputs
concerned, numerous data samples are necessary for better are kept in the desired range.
training and increasing performance of the neural networks.
Owing to the fact that there are few data samples at the ma-
2. Overview of machining characteristics
chining operation because of high expense of experimental
tests, the procedure of network training plays an important role Surface roughness of machined workpiece is an important
in accuracy of the predicted results. So far, several researchers factor in evaluating the quality of manufactured products. It
have obtained satisfactory results for estimating output(s) of the has a significant impact on characteristics such as fatigue,
machining operation by using ANNs which are trained by con- corrosion resistance and creep strength. Wear at the tool edge
ventional methods (such as back propagation) [14]. But devel- (flank wear) is also considered as an important factor which
oping a novel method for training the ANNs which yields more directly affects the surface roughness of the machined part.
accuracy of predicted results in spite of few data samples is Increasing flank wear and reaching critical level leads to re-
necessary, especially for machining applications. ducing the quality of the surface roughness [15]. Therefore, to
On the other hand, as far as optimization methods are con- increase optimal production, it is necessary to predict the flank
cerned, although MOO techniques are employed to optimize wear and estimate suitable durability of the tool before reach-
machining outputs at the same time, a problem arises when ing the critical state. A schematic view of the flank wear is
the goal involves optimizing one output of the process while shown in Fig. 1(a). Machining forces also play a significant
the other outputs are maintained in a suitable range. This role in the development of stresses and temperature in the
method (which is not implementable properly by MOO tech- machined part. Furthermore, increasing the forces may lead to
nique) can be applied as a useful tool for obviating some re- unstable state and development of chatter during the machin-
quirements and limitations of process engineers. Therefore, ing process. Since all of these features may affect the machin-
considering a novel and innovative technique for implement- ing surface characteristics, cutting forces are also considered
ing the mentioned method is favorable for engineers and de- as another factor for evaluating performance of the machining
serves to be developed further. process and must be controlled. Fig. 1(b) is a schematic view
According to the reasons mentioned above, our main goal is of the main cutting forces during the turning process. Due to
specialized to presentation of a novel intelligent method (in- the importance of the mentioned points, researchers have done
cluding predictive and optimization method) which is more studies in this regard to determine suitable conditions of the
efficient than other conventional methods for machining ap- machining processes. As mentioned above, applying efficient
plications. intelligent methods for prediction and optimization of machin-
Hence, in the first part of the paper, three artificial neural ing processes contains several advantages, such as reducing
networks were employed to present a predictive model for the cost and time of the production, and process engineers will be
turning process. Surface roughness of the machined workpiece, able to determine optimal cutting conditions in order to in-
resultant cutting forces and tool wear were predicted by ANNs crease performance of the process. Hence, using intelligent
at the specified range of input parameters (including cutting methods is favorable for industrial applications.
speed, feed rate and depth of cut). In this regard, a novel On the other hand, several researchers have performed ex-
method (in addition to the conventional methods) was intro- perimental investigations and obtained valuable data samples,
duced and implemented for training the ANNs. Based on this, but they have not analyzed their data samples properly. There-
evolutionary algorithms (GA and PSO) were utilized for train- fore, it seems necessary to perform researches for developing
ing the ANNs. In the rest of the paper, functions implemented a more suitable analysis of the aforementioned investigations
by ANNs are employed as objective functions of the optimiza- by using several techniques such as novel intelligent methods.
tion algorithms. Finally, in addition to optimizing each of the Based on this, in the performed study, 27 turning experiments
 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477 1471

using tungsten carbide with the grade of P-10 for machining

of S45C steel have been utilized from the Ref. [16]. Three
cutting parameters including cutting speeds in the range of
135-285 m/min, depth of cut in the range of 0.6-1.6 mm and
finally feed rate in the range of 0.08-0.32 mm/rev have been
investigated. Also, three outputs of the turning process includ-
ing average surface roughness (measured by 3D-
hommelewerk), resultant cutting forces (measured by Fig. 2. Typical movement of one particle in solution space.
Kisler5257A) and also tool life (based on the factor of reach-
ing to the average flank wear of the tool VB = 0.3mm or max-
imum flank wear VB (max) = 0.6mm) have been used [16].
Wnew = Wold + Wnew (2)

3. Overview of population-based algorithms where W is the inertial weight, Vnew is the particle velocity, Xcs
3.1 Genetic algorithm is the current particle (solution) of each particle, Ppb and Pgb
are pbest and gbest respectively, r is a random number between
The principle of natural evolution is the main idea of evolu- (0, 1) and C1, C2 are learning factors. Particle velocities on
tionary algorithms (EAs). According to evolution theory, par- each dimension are clamped to a maximum velocity Vmax. If
ticles of population evolve themselves to obtain more adapta- the sum of old velocity and impact of particles recorded ex-
tion with their environment. Therefore, the particles that can periences (see Eq. (1)) would cause the velocity on that di-
adapt themselves have a greater chance to survive. Such algo- mension to exceed (Vmax is a parameter specified by the user),
rithms are stochastic optimization techniques; in these tech- the velocity on that dimension will be limited to Vmax. In Fig. 2,
niques, information of each generation is transferred to the typical movement of one particle in solution space is shown.
next generation by chromosomes. Each chromosome consists To improve the performance of the basic PSO one of the
of some genes, and each of the genes illustrates a special fea- most powerful versions is presented in Ref. [19]. In this ver-
ture or behavior [17]. sion of PSO, the updating step of velocity is changed in com-
Genetic algorithm (GA) is one of the most well known evo- parison with basic PSO. To do this, a constant coefficient is
lutionary algorithms. In GA's process, first of all, an initial utilized instead of inertial weight and learning factors are re-
population is created based on the requirement of problem and placed by two random values [20]. In this approach, the updat-
after that, the objective function is evaluated in order to ing step is calculated by the following equations:
achieve the best solution. Off springs are also created from
parents in the reproduction step. In this step, some actions Vnew = χ × (Vold + ϕ1 × ( Ppb − X cs ) +
happen such as crossover and mutation. Consequently, the (3)
ϕ2 × ( Pgb − X cs ))
best solution is obtained during predetermined iterations.
Wnew = Wold + Wnew (4)

3.2 Particle swarm optimization (PSO) with constraint coef- ϕ > 4
ficient  2k (5)
χ =
PSO simulates the behavior of bird populations. In PSO,
 ϕ − 2 + ϕ 2 − 4ϕ
each single solution is a ‘‘bird’’ in the search space and it’s
called a ‘‘particle’’. For all of the “p” particles, a fitness value where ϕ1 and ϕ2 are random numbers uniformly distrib-
is evaluated by the fitness function to be optimized. The p parti- uted in the range (0, ϕ2 ), and χ is the constant coefficient.
cles are ‘‘flown’’ through the problem space by following the
current optimum p particles. PSO is initialized with random p
4. Overview of artificial neural network (ANN)
particles (solutions) and then searches for optimization by up-
dating generations. In each iteration, each p particle is updated ANNs simulate the simplified model of human brain appli-
by following two ‘‘best’’ values. The first one that is obtained cation. Learning ability is the most important characteristic of
so far by any particle is called "personal best (pbest)". Another the human brain. Therefore, artificial intelligence engineers try
‘‘best’’ value is the best value among all personal bests and all to build a useful tool or software that can use experiences in
iterations. This best value is a global best which is called "gbest". features and can give the best decision in special conditions.
After finding the two best values, the particle updates its veloc- In general, ANNs based on the learning algorithm are di-
ity and positions based on Eqs. (1) and (2) [18]: vided into two major groups: supervised and unsupervised
learning. In supervised learning, the input and desired output
Vnew = W ×Vold + C1 × r ×( Ppb − X cs ) + are available as learning or pattern data and ANN is trained by
(1) using them. In unsupervised learning target, outputs are not
C2 × r × ( Pgb − X cs )
available and the network can only cluster input data, and
1472 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477

Fig. 3. A simple structure of ANN.

whenever new data is entered, unsupervised ANN can assign

it into a corresponding cluster.
Fig. 4. Flowchart of training ANN with GA or PSO.
4.1 Performance of ANN

As can be seen in Fig. 3 a simple model of an artificial neu-

ral network is created by connecting neurons. The relative
position of cells in a network (number of neurons, number of
layers and all connections between cells) is the network topol-
ogy. Indeed, a topology is a hardware connection system of
neurons that by using corresponding software (namely,
mathematical procedure of information flow and weight's
calculation) determines the kind of neural network application.
In this topology, there is one input layer that accepts informa-
tion; also there are some hidden layers that give information
from a previous layer, and finally one output layer that accepts Fig. 5. A simple structure of ANN.
results of calculations and presents output.
Each cell is connected to all cells of the next layer. Self-
connection, connecting to previous layer and jumping from
5.2 Administration of the neural network structure
one layer are forbidden. This is so-called "feed forward", be-
cause information always flows from input to output. About The neural network used in this paper is a multilayer per-
90-95 percent of ANN's applications use this topology. At first, ception neural network in which none of the conventional
synapse weights are random values that will be corrected dur- training methods have been considered. For example, as can
ing especial special iterative training method [21]. be seen in Fig. 5, a two layer neural network, which has two
inputs in the input layer, four neurons in the middle layer and
one neuron in the output layer has been drawn. In this paper,
5. Training neural network by genetic algorithm
the weights and biases of the neural network are the variables
As mentioned in the previous section, different methods are of the optimization algorithm.
used to train (or in other words, update weights and biases), Based on the example mentioned above, the number of
which most of them use analytical and mathematical based variables to be updated can be calculated through Eq. (6).
methods, like back propagation and gradient descent. In this
paper, a novel method has been introduced to train ANN [22]. Number of Variable = I × N Lh + BLh + N Lh × N Lo
(6)
In the proposed method, genetic algorithm has been used to + BLo .
train ANN, as the optimization method. Fig. 4 shows the
flowchart of training ANN using GA. Also, in the continua- In which I is the number of the inputs, NLh is the number of
tion of this section the proposed method for training ANN is neurons of the hidden layer, BLh is the number of biases of the
explained in detail. hidden layer, NLo as the number of neurons in the output layer
and BLo as the number of biases of the output layer. The num-
ber of dimensions of the optimization problem in training the
5.1 Preparation of the genetic algorithm
neural network is therefore equal to the variables calculated
The genetic algorithm has some parameters which should through the above equation.
be set in the initialization step, such as population size, elite At first, input data and target matrix are loaded. Then, the
count, migration fraction, migration interval, crossover and matrix of weights that expresses the initial position of chro-
mutation. mosomes should be loaded. This matrix is 1 × i, as follows:
 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477 1473

Weight = [ P1 , P2 ....Pi ] (7)

where weights are the position of the first chromosome in

solution space. In other words, it states the weights of ANN
for first chromosome. Then this matrix should be divided into
submatrixes which express weights of one layer and its biases,
e.g., for the mentioned example, W1 = [P1-P8] belongs to syn-
aptic weights between input data and first layer and b1 = [P9-
P12] belongs to biases of first layer's neurons. After separating
these matrixes, the structure of ANN can be constructed, as
follows:

S1 = W1 ×[ I ] + b1 (8)
S 2 = log sig ( S1 ) (9)
Z1 = W2 × S 2 + b2 (10) Fig. 6. Flowchart of the proposed method for optimization of machin-
Z 2 = log sig ( Z1 ) (11) ing outputs.

where I is input data. W1 and W2 are weights between input

6. Proposed method
and hidden layer, between hidden layer and output layer, re-
spectively. Matrices b1 and b2 are biases related to hidden Neural networks can be used as a suitable model for com-
layer, output layer, respectively. plex processes. Since machining processes often involve
Finally, MSE value of subtraction of actual output and tar- nonlinear and complicated features, correct prediction of the
get output, Z2, is calculated. This value is the fitness value of behavior of such processes is of great significance. The pro-
MLFFNN for GA and the proposed algorithm will change posed method has two major steps: in the first step, trained
weights until this value becomes a minimum. The fitness ANN with GA has been used to present an efficient and opti-
function can be found in Eq. (12). mum model for machining process; in second step, designed
models for optimization of the outputs of the machining proc-
1 n
Fitness function = ∑ (target i − actuali )2
n i =1
(12) ess are then used.
Fig. 6 shows a block diagram of the whole method pro-
where targets are true values, actuals are prediction values and posed for the optimization of the machining process.
“n” is the number of training data.
6.1 Modeling the machining process
5.3 Optimization of weights and biases by GA
Different methods have been proposed for training neural
In this method the neural network is used as an objective networks, all of which make an effort to update weights and
function in optimization. Each of the chromosomes of the biases of neural networks in a way that minimum error can be
genetic algorithm represents a set of weights and biases of one seen in the output of the network. Therefore, such a process
neural network, and each of them may separately be an an- can fall into the form of an optimization problem, in which the
swer to the optimization problem. Depending on the adjust- goal is to minimize the error in the output of the neural net-
ments made, these chromosomes transform into new chromo- work with the most optimal weights and biases. Genetic algo-
somes, which are evaluated based on the objective function, or rithms can make this task possible as one of the most powerful
the neural network, in other words. optimization methods used by the operator for training the
neural network. In this step, an optimal machining process
model has been created using an artificial neural network
5.4 Evaluation of stop criterion
which is trained by genetic algorithm and experimental data
To stop the optimization process by genetic algorithms a borrowed from Ref. [16] that are listed in Table 1.
number of criteria have been used. Two of these criteria are In Table 1, 21data out of the 27 existing data have been
the number of repetitions of the algorithm and the amount of randomly used as training data, which are applied in each
error in training the neural network (less than 2%), which have stage of evaluation so that in the end an accurate model is
been considered in this paper. The optimization process con- obtained.
tinues until one of the above conditions is reached. After train-
ing the neural network with the help of genetic algorithms, the 6.2 Optimization of the outputs of the turning process
six remaining data are used as test data to evaluate the accu-
racy of the trained network. In this section, the designed models (the trained network)
1474 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477

Table 1. Utilized experimental data which are presented in Ref. [16]. Table 2. Mean error value obtained by trained ANN by GA, FFBPNN
and CFBPNN.
Vc f d Ra F T.L
(m/mm) (mm/rev) (mm) (µm) (N) (s) GA FFBPNN CFBPNN
135 0.08 0.6 1.24 263.0 2645 Trained Mean error Mean error Mean error
ANN (%) (%) (%)
135 0.20 0.6 5.34 403.0 2379
135 0.32 0.6 9.49 550.0 2233 Train Test Train Test Train Test

135 0.08 1.1 1.68 454.0 2604 (F) 2.98 4.36 3.81 4.06 11.13 5.08

135 0.20 1.1 1.92 704.0 2060 (T.L) 6.75 5.17 12.19 9.7 11.25 22.88

135 0.32 1.1 4.06 889.0 1870 (Ra) 2.49 6.72 13.96 27.05 17.61 30.14

135 0.08 1.6 1.86 628.0 2563

135 0.20 1.6 4.12 924.0 2032
135 0.32 1.6 9.44 1198.0 1733
210 0.08 0.6 2.61 212.0 1605
210 0.20 0.6 4.51 389.0 1198
210 0.32 0.6 11.05 502.0 802
210 0.08 1.1 1.01 377.0 1350
210 0.20 1.1 2.75 622.0 1059
210 0.32 1.1 7.49 853.8 734
210 0.08 1.6 2.64 593.0 1310
210 0.20 1.6 6.06 952.0 1031
210 0.32 1.6 14.37 1169.7 602
285 0.08 0.6 0.56 203.0 860
285 0.20 0.6 2.84 363.8 847
285 0.32 0.6 9.70 464.2 216
285 0.08 1.1 0.91 335.0 854
285 0.20 1.1 2.74 573.1 846
285 0.32 1.1 6.12 812.7 212
285 0.08 1.6 1.25 443.0 810
285 0.20 1.6 4.18 856.8 765
285 0.32 1.6 10.17 1099.3 203

are used as objective functions for each of the outputs of the

machining process in the previous section for the optimization
of the outputs of the machining process. GA and PSO optimi-
zation algorithm are used for the optimization of the deter-
mined output. Machining parameters are determined in a way
so that each time one of the outputs is optimized the other
outputs are maintained in a suitable range.
Fig. 7. Convergence curves of training procedure for F, T and Ra.

7. Results and discussion

7.1 Creating predictive model of machining process two conventional ANN: feed forward back propagation neural
network (FFBPNN) and cascade forward back propagation
Our first aim was presentation of an optimum machining neural network (CFBPNN).
process model. Using experimental results, three artificial According to Eq. (12), minimum value of mean error shows
neural networks have been trained by GA for average ma- that ANN is trained better. Table 1 shows trained ANN by GA
chined surface (Ra), resultant forces (F) and tool life (T.L). presents better results than FFBPNN and CFBPNN. Therefore,
Table 2 illustrates the obtained mean error for training data models created by GA and ANN are better choices for the
and testing data by using trained ANN with GA; also, in order machining process as compared to those by FFBPNN and
to confirm the ability of the proposed method, the obtained CFBPNN. Also, Fig. 7 shows the convergence curves of train-
results with trained ANN by GA have been compared with ing procedure for F, T.L and Ra. It should be noted that tool
 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477 1475

Table 3. Estimating optimum value of machining process using GA Table 4. Estimating optimum value of machining process using PSO
without constraint. without constraint.

Machining Input machining parameters Optimum Machining Input machining parameters Optimum
output d (mm) f (mm/rev) Vc (m/min) value output d (mm) f (mm/rev) Vc (m/min) value

(Ra) 1.1601 0.089 278.68 0.447 (Ra) 1.192 0.086 268.31 0.4539
(F) 0.60 0.08 262.670 201.282 (F) 0.60 0.080 250.16 201.7742
(T.L) 0.701 0.081 135 2666.6 (T.L) 0.64 0.085 135.03 2665.032

Fig. 9. Convergence curve of cutting forces (F) using GA.

Fig. 8. Flowchart of optimization of machining process using PSO and

GA.

life should be maximized; however, in this paper we use 1/T

that this term is being minimized.

7.2 Optimization of machining outputs Fig. 10. Convergence curve of cutting forces (F) using PSO.

In this section, obtained results have been presented in two

forms. In the first form, optimum results have been obtained one output simultaneously may be required. Therefore, in this
without any constrains and in the second form, one output is section, each output has been optimized, whereas other out-
being optimized, whereas some constraints have been consid- puts have been considered in the given constraint. For exam-
ered for other outputs. Fig. 8 shows a flowchart of the optimi- ple, when we want to determine optimum value for Ra by
zation process using PSO and GA. using GA, T and F are considered in the given range as shown
in Table 5. Tables 5 and 6 present the obtained optimum val-
7.2.1 Optimum outputs without any constraints ues machining process outputs using GA and PSO, respec-
As shown in Fig. 8, PSO and GA have been used in this op- tively. In these tables, limitations are based on minimum pos-
timization problem without any consideration. Obtained re- sible values for them.
sults using GA and PSO have been presented in Tables 3 and
4, respectively. In addition, convergence curves of the resul-
8. Conclusion
tant forces using GA and PSO have been shown in Figs. 9 and
10, respectively. The main goal of this paper was to develop a new intelligent
method for presenting a predictive and optimizing model of
7.2.2 Optimum outputs with constrains the turning process. Hence, surface roughness of the machined
In many cases, obtaining optimum values for one output is workpiece, resultant cutting forces and tool wear were pre-
not sufficient, and according to machining condition and type dicted by three artificial neural networks. A novel method was
of machining device, evaluating and controlling more than proposed for training the ANNs by using evolutionary algo-
1476 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477

Table 5. Obtained optimum values for machining process outputs specified output, while other outputs were kept in the desired
using GA (by employing constraint). range (see Tables 5 and 6).
Input parameters Finally, we hope that the proposed novel intelligent method
Opt in this research can be employed by other branches of manu-
Output Vc f d Lim1 Lim2
value
m/min mm/rev mm facturing processes instead of other conventional methods.
Rough- T.L
--------- ------ ------ F (N) ---------
ness (s)
Nomenclature------------------------------------------------------------------------
156.82 0.08 0.6 280 2100 1.34
Ra(µm) 206.94 0.10 1.097 600 1200 0.897 Ra : Surface roughness
235.20 0.08 1.048 800 900 0.63
T.L : Tool life
F : Resultants cutting forces
Forces --------- ------ ------ Ra (µm) T.L (s) ---------
Vc : Cutting speed
166.54 0.08 0.6 2200 1.5 235.437
f : Feed rate
F(N) 253.58 0.08 1.034 1000 0.8 332.788 d : Depth of cut
262.87 0.08 1.062 700 0.6 343.739 PSO : Particle swarm optimization
Tool life --------- ------ ------ Ra (µm) F (N) --------- GA : Genetic algorithm
236.21 0.08 1.416 0.85 450 1098.03 ANN : Artificial neural network
T.L(s) 149.07 0.08 0.6 1.5 250 2457.12 Lim : Limitation
228.69 0.08 1.048 0.9 350 1194.58 Opt : Optimum
FFBPNN : Feed forward back propagation neural network
Table 6. Obtained optimum values for machining process outputs CFBPNN : Cascade forward back propagation neural net-
using PSO (by employing constraint). work

Input parameters
Opt References
Output Vc f d Lim1 Lim2
value
m/min mm/rev mm [1] T. Childs, K. Maekawa, T. Obikawa and Y. Yamane, metal
Rough- T.L cutting theory and application, New York, USA (2000).
--------- ------ ------ F (N) ---------
ness (s) [2] A. Javidi, U. Riegger and W. Eichlseder, The effect of ma-
156.82 0.08 0.6 280 2100 1.34 chining on the surface integrity and fatigue life, Interna-
Ra(µm) 159.34 0.08 1.128 600 1200 0.903 tional journal of Fatigue, 30 (2008) 2050-2055.
170.15 0.11 1.116 800 900 0.67 [3] D. I. Lalwani, N. K. Mehta and P. K. Jain, Experimental
Force --------- ------ ------ Ra (µm) T.L (s) --------- investigations of cutting parameters influence on cutting
forces and surface roughness in finish hard turning of
165.94 0.08 0.6 1.5 2200 235.697
MDN250 steel, Journal of Materials Processing Technology,
F(N) 251.91 0.08 1.034 0.8 1000 333.561
206 (2008) 167-179.
284.06 0.08 1.0318 0.6 700 340.600
[4] I. Mukherjee and P. K. Ray, A review optimizations tech-
Tool life --------- ------ ------ Ra (µm) F (N) --------- niques in metal cutting processes, Computers & Industrial
235 0.08 1.0307 0.85 450 1136.27 Engineering, 50 (2006) 15-34.
T.L(s) 149.07 0.08 0.6 1.5 250 2457.18 [5] C. X. J. Feng and X. Wang, Development of empirical mod-
235.77 0.08 1.031 0.9 350 1129.13 els for surface roughness prediction in finish turning, inter-
national journal of Advanced Manufacturing Technology, 20
(2002) 348-356.
rithms. Comparison between predicted results of the men- [6] A. M. Zain, H. Haron and S. Sharif, Prediction of surface
tioned method by other NNs which were trained by conven- roughness in the end milling machining using artificial neu-
tional methods (FFBPNN and CFBPNN which use back ral network, journal of Expert Systems with Applications, 37
propagation method) showed that the proposed method is (2010) 1755-1768.
more suitable than other implemented methods and can be [7] V. S. Sharma, S. Dhiman, R. Sehgal and S. K. Sharma, Es-
properly utilized for predicting outputs of the turning process timation of cutting forces and surface roughness for hard
(see Table 2). Also, functions implemented by ANNs were turning using neural networks, Journal of Intelligent Manu-
employed as objective functions of evolutionary algorithms facturing, 19 (2008) 473-483.
(GA and PSO) to determine suitable input parameters (includ- [8] A. M. Zain, H. Haron and S. Sharif, Application of GA to
ing cutting speed, depth of cut and feed rate) for optimizing optimize cutting conditions for minimizing surface rough-
outputs of the process. In this regard, in addition to optimizing ness in end milling machining process, Expert Systems with
each of the aforementioned outputs separately (see Tables 3 Applications, 37 (2010) 4650-4659.
and 4), a new method was implemented for optimizing one [9] Y. Jiao, S. Lei, Z. J. Pei and E. S. Lee, Fuzzy adaptive net-
 F. Jafarian et al. / Journal of Mechanical Science and Technology 27 (5) (2013) 1469~1477 1477

works in machining process modeling: surface roughness [21] Ch. M. Bishop, Pattern recognition and machine learning,
prediction for turning operations, International Journal of Springer, USA (2006).
Machine Tools & Manufacture, 44 (2004) 1643-1651. [22] M. Taghipour, A. R. Moradi and M. Yazdani-Asrami, Iden-
[10] T. Ozel, Y. Karpat, L. Figueira and J. P. Davim, Modeling tification of magnetizing inrush current in power transform-
of surface finish and tool flank wear in turning of AISI D2 ers using GSA trained ANN for educational purposes, IEEE
steel with ceramic wiper inserts, Journal of Materials Proc- Conference on Open Systems (ICOS), Kuala Lumpur, Ma-
essing Technology, 189 (2007) 192-198. laysia (2010) 23-27.
[11] I. N. Tansel, B. Ozcelik, W. Y. Bao, P. Chen, D. Rincon, S.
Y. Yang and A. Yenilmez, Selection of optimal cutting con-
ditions by using GONNS, International Journal of Machine Farshid Jafarian received his B.Sc. in
Tools & Manufacture, 46 (2006) 26-35. Manufacturing Engineering from Azad
[12] I. N. Tansel, S. Gulmez, M. Demetgul and S. Aykut, Ta- University of Najafabad, Iran, and his
guchi Method-GONNS integration: Complete procedure M.Sc. degree in Manufacturing Engi-
covering from experimental design to complex optimization, neering from University of Birjand,
Expert Systems with Applications, 38 (2011) 4780-4789. Birjand, Iran. He is pursuing the Ph.D.
[13] D. Mandal, S. K. Pal and P. Saha, Modeling of electrical in Manufacturing Engineering at Uni-
discharge machining process using back propagation neural versity of Birjand, Birjand, Iran. His
network and multi-objective optimization using non- research interests include: traditional and nontraditional ma-
dominating sorting genetic algorithm-II, Journal of Materi- chining processes, surface integrity of machined workpiece,
als Processing Technology, 186 (2007) 154-162. FE simulation of machining process and application of intelli-
[14] T. Ozel and A. Nadgir, Prediction of flank wear by using gent methods at the machining processes.
back propagation neural network modeling when cutting
hardened H-13 steel with chamfered and honed CBN tools, Mehran Taghipour-Gorjikolaie re-
International Journal of Machine Tools & Manufacture, 42 ceived his B.Sc. in Electrical Engineer-
(2002) 287-297. ing from University of Mazandaran,
[15] A. Ebrahimi and M. M. Moshksar, Evaluation of ma- Babol, Iran in 2008, and his M.Sc. de-
chinability in turning of micro-alloyed and quenched- gree in Electrical Engineering from
tempered steels: Tool wear, statistical analysis, chip mor- University of Birjand, Birjand, Iran in
phology, journal of materials processing technology, 209 2011. He is pursuing the Ph.D. in Elec-
(2009) 910-921. trical Engineering at University of Bir-
[16] B. Y. Lee and Y. S. Tarng, Cutting-parameter selection for jand, Birjand, Iran. His research interests include machine
maximization production rate or minimizing production cost learning, pattern recognition and artificial intelligence, design
in multistage turning operations, journal of materials proc- of fault detection systems and modeling and simulation of
essing technology, 105 (2000) 61-66. electrical machines.
[17] C. H. Antunes, D. F. Pires, C. Barrico, Á. Gomes and A. G.
Martins, A multi-objective evolutionary algorithm for reac- Hossein Amirabadi received his B.Sc.
tive power compensation in distribution networks, Applied in Mechanical Engineering (Solid
Energy, 86 (2009) 977-984. Mechanics) from Ferdowsi University,
[18] J. Kennedy and R. C. Eberhart, Particle swarm optimization, Mashhad, Iran in 1998, and his M.Sc. in
In Proc. IEEE International Conference on Neural Networks Manufacturing Engineering from IUT
IV (1995) 1942-1948. University, Isfahan, Iran in 2002. He
[19] M. Clerc and J. Kennedy, The particle swarm-explosion, received his Ph.D. in Manufacturing
stability, and convergence in a multidimensional complex Engineering from University of
space, IEEE Transaction on Evolutionary Computation, 6 Mazandaran, Babol, Iran in 2006. After that he joined the
(2002) 58-73. University of Birjand as an Assistant Professor. His research
[20] S. H. Zahiri, Swarm intelligence and fuzzy systems, Nova interests include Machine tools, Metal Cutting, Machining
Publisher, USA (2010). Simulation, surface Finishing, Metrology and CAD/CAM.

View publication stats