Journal of Manufacturing Processes 70 (2021) 560–569

Contents lists available at ScienceDirect

Journal of Manufacturing Processes

journal homepage: www.elsevier.com/locate/manpro

Proper estimation of surface roughness using hybrid intelligence based on

artificial neural network and genetic algorithm
Cem Boga a, Tahsin Koroglu b, *
a
Department of Mechanical Engineering, Adana Alparslan Turkes Science and Technology University, Adana 01250, Turkey
b
Department of Electrical and Electronics Engineering, Adana Alparslan Turkes Science and Technology University, Adana 01250, Turkey

A R T I C L E I N F O A B S T R A C T

Keywords: The surface roughness is a crucial index that is commonly used in the machining process to evaluate the final
CNC milling product quality. This paper investigates the effect of different machining parameters on the surface roughness of
Carbon fiber composite the high-strength carbon fiber composite plate, manufactured by utilizing the vacuum infusion process, under
Surface roughness
dry end milling conditions. Besides, a hybrid intelligence approach consisting of artificial neural network (ANN)
Taguchi design
Artificial neural network
whose parameters are tuned by genetic algorithm (GA) is introduced for accurate estimation of surface rough­
Genetic algorithm ness. To construct a database for the ANN, the experimental milling tests have been carried out according to the
Taguchi optimization method with the design of a mixed orthogonal array L32 (21 × 42). The influence of the
machining parameters such as cutting tools, feed rate, and spindle speed on surface roughness have been
examined by using analysis of variance (ANOVA). The analyses reveal that the cutting tool and the feed rate are
the most effective factors in the surface roughness of the composite material. It is also determined that the
experiment with “A1B2C1” combination (TiAlN coated cutting tool, 5000 rpm spindle speed, and 250 mm/rev
feed rate) gives the optimal result. The proposed hybrid ANN-GA algorithm provides a good prediction corre­
lation ratio (R = 0.96177) indicating that the estimated and the measured surface roughness values are
remarkably close to each other. The mean square error (MSE) specifying the accuracy and adequacy of the
network model is obtained as 0.074 during the 33th iteration of the GA.

1. Introduction properties as presented in Zain et al. [2]. The proper selection of cutting
tools, feed rate, and spindle speed, which are handled in this study, have
In recent years, carbon fiber composites are preferred for use in a considerable impact on surface roughness.
various engineering applications, especially in the automotive, aero­ There have been numerous research studies focused on the analysis
space, and defense industries due to their unique mechanical and and optimization of machining parameters affecting the surface rough­
physical properties such as high-strength, lightweight, and resistance to ness and the estimation of surface roughness by using these parameters.
corrosion. Conventional machining operations such as milling, drilling, Karabulut [3] investigated the effect of milling parameters on surface
edge cutting, turning, and grinding are applied to carbon fiber com­ roughness and cutting force for AA7039/Al2O3 metal matrix composites
posites to make them ready for use in these applications. Milling is a produced by the powder metallurgy method. After the experiments
commonly used process in the shaping of carbon fiber composite ma­ performed based on Taguchi design, it was determined that the material
terials and providing the desired quality surface. Surface roughness, structure is the most effective factor in the surface roughness and the
which has effects on different features of machined parts like friction, feed rate is the dominant factor affecting the cutting force. In Karabulut
wear, and heat conduction, is known as the main indicator of the surface [3], artificial neural networks (ANN) and regression analysis were used
quality of materials. The low value of surface roughness contributes to predict surface roughness and cutting force. Sreenivasulu [4] exam­
significantly to the product's tribological properties, fatigue strength, ined the effect of cutting speed, feed rate, and cutting depth on the
corrosion resistance, and aesthetic appearance [1]. The surface rough­ delamination damage and the surface roughness of glass fiber reinforced
ness value (Ra) is influenced by many factors such as machining pa­ polymeric composite material (GFRP) during end milling. From the
rameters, cutting phenomena, workpiece properties, and cutting tool findings of the analysis of variance (ANOVA), it was determined that the

* Corresponding author.
E-mail address: tkoroglu@atu.edu.tr (T. Koroglu).

https://doi.org/10.1016/j.jmapro.2021.08.062
Received 22 January 2021; Received in revised form 18 August 2021; Accepted 31 August 2021
Available online 15 September 2021
1526-6125/© 2021 The Society of Manufacturing Engineers. Published by Elsevier Ltd. All rights reserved.
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

most important contributing factors to the response are the cutting 0.1446 m/min for speed, cutting depth, and feed respectively. Kivak
speed and the depth of cut with 26.84% and 40.44%, respectively. ANN [14] applied the Taguchi method and regression analysis to evaluate the
was applied to estimate both the surface roughness and the delamination machinability of Hadfield steel with PVD TiAlN- and CVD TiCN/Al2O3
damage. The difference between experimental and the predicted values coated carbide inserts under dry milling conditions. The findings
were found in the range of 3.7%. Muthukrishnan and Davim [5] carried showed that the feed rate was the dominant factor over the other
out a study to investigate the effect of different cutting conditions, in machining parameters (cutting tool and cutting speed) affecting surface
which coarse grade polycrystalline diamond (PCD) insert was used, on roughness. Zhou et al. [15] developed a feed-forward multi-layered ANN
the surface roughness of Al-SiC metal matrix composites. The surface based prediction model to investigate the mathematical relationship
roughness data collected for various combinations of cutting speed, feed between cutting parameters and average surface roughness during the
rate, and depth of cut was tested with ANOVA and ANN techniques. milling process of Al/SiC particle composite material. Laot et al. [16]
Varol et al. [6] proposed an ANN approach to estimate the effect of performed parameter optimization in the end milling process of Carbon
reinforcement size and content on the physical and mechanical prop­ Fiber Reinforced Polymer (CFRP) material using back propagation ANN
erties of Al2024-B4C composite specimens. The density, hardness, and and GA to predict cutting force and surface roughness. Mohsin et al. [17]
tensile strength values of the composites were used as the outputs of the worked on circular samples using different polishing parameters.
developed ANN model and the effectiveness of the ANN was evaluated Taguchi method and analysis of variance (ANOVA) were used to analyze
by comparing these outputs with the measured real data. It was stated in the effects of these input parameters on robotic polishing efficiency and
Varol et al. [6] that the effect of reinforcement size and content on the torque. From the results obtained after the experiment, they determined
homogeneous distribution of B4C particles are as important as the effect that the most effective parameters for polishing efficiency were polish­
of the milling time. In Daniel et al. [7], the Taguchi design based grey ing force and tool speed. Zahaf and Benghersallah [18] studied the
relational analysis was presented for the optimization of the control mathematical modeling and optimization of surface roughness during
parameters such as mass ratio and the particle size of SiC as well as the milling of AISI 52100 bearing steel with TiAlN coated carbide tools.
depth of cut, feed, and cutting speed in the milling of Al5059/SiC/MoS2 Taguchi experimental design was used to investigate the effect of cutting
(aluminum hybrid metal matrix composite). The effect of each control parameters such as cutting speed, feed per tooth, axial depth of cut, and
factor on response variables was analyzed through the Taguchi signal- radial depth of cut on surface roughness. After the experiments, they
to-noise (S/N) ratio method and it was determined that the feed rate stated that the most effective parameters on the milled surface rough­
and depth of cut are the most effective factors. The optimized control ness were cutting speed and feed per tooth. Ramasamy et al., [19]
parameters were used as the inputs of ANN to predict the surface investigated the surface roughness of Al7050 metal matrix material
roughness. Karabulut et al. [8] conducted experiments based on Tagu­ reinforced with Titanium Carbide (TiC) and Boron Nitride (BN) as a
chi's L18 (21 × 32) with a mixed orthogonal array to investigate the result of milling process. After the ANOVA results, they determined that
influential machining parameters affecting the surface roughness in the TiC has a high significant effect on surface roughness, followed by cut­
milling of AA7039-based metal matrix composites produced by powder ting speed and feedrate. They emphasized that BN has less effect. Liu
metallurgy. It was emphasized that the most significant cutting param­ et al. [20] conducted a series of experiments on AISI 304 steel under
eters on the finished surface are the cutting speed and feed rate ac­ different cutting conditions to evaluate surface quality and tool vibra­
cording to the analysis results. Karabulut and Karakoç [9] put forward tion. As a result, they demonstrated a reduction in surface roughness and
the optimization of the machining parameters of aluminum alloy 7075 vibration under minimal quantity cooling lubrication conditions. They
(Al7075) based open cell silicon carbide (SiC) foam composite for stated that the parameter that significantly affects the surface roughness,
providing acceptable surface quality. Taguchi L27 orthogonal array was axial vibration and radial vibration is the feed rate. In addition, they
applied for the optimization of milling parameters while ANN and found that while the vibration and surface roughness decreased with the
regression methods were utilized for the prediction of surface roughness. increase in cutting speed, the depth of cut did not significantly affect the
The feed rate was found to be the major factor influencing the surface surface roughness.
roughness which was predicted with a mean squared error equal to 1.6 It is seen in the literature review that there is a lack of publications
and 0.24%. Öktem et al. [10] aimed to develop a Taguchi design opti­ both on the estimation of surface roughness of carbon fiber composites
mization method for low surface roughness of 7075-T6 aluminum ma­ using the hybrid ANN-GA algorithm and on the examination of the
terial by taking into account of processing parameters such as feed, cutting tool effect with the Taguchi design technique. With this paper,
cutting speed, axial-radial depth of cut, and machining tolerance. S/N the gap in the literature will be filled. The novelty of this study is to
ratio and ANOVA were harnessed to find the optimal levels and the ef­ analyze the effect of various milling parameters on the surface rough­
fect of the process parameters on the surface roughness. The effective­ ness of carbon fiber composite plate with a depth of cut below 0.5 mm
ness of the Taguchi method in solving the surface quality problem of and to develop a hybrid ANN-GA based prediction algorithm by using
mold surfaces was demonstrated via a verification experiment and it was experimental data. The learning rate, the training function, and the
concluded that the surface roughness was improved (from the value of number of neurons used in the hidden layers of the ANN have been
0.412 μm to 0.395 μm) by 5.1%. Moghri et al. [11] asserted an inte­ optimized with GA to increase the prediction performance. To minimize
grated ANN-genetic algorithm (GA) approach for modeling and opti­ the prediction error of the proposed algorithm, the GA is implemented
mizing surface roughness of polyamide-6/nanoclay (PA-6/NC) with higher population size and iteration number. ANOVA has been
nanocomposites. A surface roughness predictive model was developed applied to determine the most important cutting factors based on the
by considering milling parameters (spindle speed and feed rate) and Taguchi L32 method.
nano clay (NC) content. It was inferred that the minimum surface This paper is organized into six sections. Section 1 provides the
roughness could be achieved at the lowest level of feed rate and inter­ literature review on the analysis and optimization of machining pa­
mediate level of spindle speed. Sangwan et al. [12] came up with a rameters affecting the surface roughness of the materials and the esti­
hybrid ANN-GA technique to determine optimum machining parameters mation studies. In Section 2, the workpiece material, the milling
to acquire minimal surface roughness during the turning of Ti-6Al-4 V experiments, and the machining parameters to be analyzed are
titanium alloy. Gopan et al. [13] introduced a predictive and optimi­ explained. In Section 3, the Taguchi method and the design of experi­
zation model composed of ANN and GA to find the global optimal point ments are described in detail. In Section 4, the proposed hybrid ANN-GA
of the machining parameters ensuring the minimum surface roughness approach is introduced. In Section 5, experimental results and the esti­
value. In the experiment performed on a cylindrical grinding machine mation performance of the hybrid ANN-GA algorithm are presented and
with silicon carbide stone, the optimum surface roughness was attained discussed. Finally, in Section 6, the significant contributions of the paper
as 0.18017 μm with the adjusted values of 183.88 rpm, 0.0288 mm, and are highlighted as a conclusion.

561
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

2. Experimental methods Table 1

Factors and levels of milling.
2.1. Workpiece material Parameters Symbol Level 1 Level 2 Level 3 Level 4

Cutting tools A Tool 1 Tool 2

The workpiece material is obtained from a 340 × 90 × 10 mm carbon Spindle speed (rpm) B 4000 5000 6000 7000
fiber plate with high mechanical properties and a good surface finish. It Feed rate (mm/dev) C 250 500 1000 1500
has been manufactured using 100% carbon fiber reinforcement and
epoxy resin matrix, cured under pressure at a high temperature. To
create a more even distribution of strength, all sheets are produced in 20
ply with 0◦ /90◦ angle orientation known as semi-isotropic fiber orien­
tation. The matrix material used in the carbon fiber plate is a high-
performance epoxy (F-1564) resin that offers good mechanical proper­
ties, clarity, and improved UV stability. During the production phase of
the composite plate, curing time and temperature have been applied as
8 h and 80 ◦ C, respectively.

2.2. Milling experiments

In order to carry out the milling experiments under dry cutting

conditions, Sunmill JHV 710 model, a three-axis CNC milling machine
with 8000 rpm rotation speed, X (800 mm) Y (600 mm) Z (450 mm)
table features, was used.
Fig. 1 shows the setup and application of the test for the milling
process. In the experiments, Er 32 Veldon tool holder, two different tools
of Titanium-Aluminum-Nitride coated-TiAlN (Tool 1) and Mikrograin
Carbide-C10 (Tool 2) were used. For the milling operations in all ex­
periments, 0.5 mm depth of cut was applied. The milling parameters Fig. 2. The effects of machining parameters on surface roughness.
were determined taking into account the values recommended in the
tool manufacturer's catalog. The surface roughness (Ra) of the work­ 3. Experimental design and optimization
piece was measured with a Mitutoyo SJ-210 surface roughness tester
device according to ISO 4287 standard. 3.1. The Taguchi method and design of experiments
The cutting parameters used in the experiments are shown in
Table 1. The optimal selection of machining parameter values (spindle Taguchi method is an experimental design method widely used in
speed and feed rate) contributes to the increase of surface quality by engineering analysis. It provides a highly efficient and systematic
reducing the roughness. As can be seen from the surface plot depicted in approach to determine the optimum cutting parameters during pro­
Fig. 2 that the machining parameters have a significant effect on the duction. It significantly reduces the number of tests and minimizes the
surface roughness. effects of uncontrollable factors by using orthogonal arrays. It calculates
the deviation between experimental and desired values using a loss
function which is then converted into a signal-to-noise (S/N) ratio.
Generally, three types of quality characteristics are used in the analysis
of the S/N ratio: (i) the smaller the better, (ii) the larger the better, and

Fig. 1. Schematic of experimental setting.

562
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

(iii) the nominal the best state. Since the aim of this study is to minimize
the surface roughness, “the smaller the better” quality property is used
as shown in Eq. (1).
/ ( )
1∑n 2
S N = − 10log xi (1)
n i=1

where n denotes the number of observations in the experiment and xi,

the data observed in the experiment. In this study, cutting tools, spindle
speed and feed rate were examined to determine the effect of processing
parameters on surface roughness. The most suitable orthogonal array
L32 (21 × 42) was chosen to determine the optimal cutting parameters
and analyze the effects of processing parameters.
Surface roughness (Ra) was measured using the experimental design
Taguchi method for each combination of control factors. Optimization
of the measured control factors was achieved by signal-to-noise (S/N)
ratios. The lowest values of surface roughness are very important to
Fig. 3. S/N ratio graph of the factors.
improve the quality of the product and to reduce production costs.
Therefore, “the smaller the better” equation was used to calculate the S/
N ratio. Analysis of the effect of each control factor (cutting tools,
Table 3
spindle speed, feed rate) on surface roughness was carried out with a Response table for means.
“Response Table for Signal to Noise Ratios”. S/N reactions for Ra are
Level Cutting tools Spindle speed (rpm) Feed rate (mm/rev)
shown in Table 2.
The level that gives the optimum value is the result of A1B2C1 1 1.901 2.429 1.887
2 3.071 2.345 1.898
response value. Likewise, the option where S/N gives the biggest
3 2.608 3.409
response is also seen in Fig. 3. Each point marked for each parameter in 4 2.560 2.747
the graph shows different levels. The highest graphical value of these Delta 1.170 0.263 1.520
points gives the level that fulfills the optimum condition. Rank 2 3 1
Table 3 represents the responses for means taken from MINITAB
software. Since the lowest value of the surface roughness means good
surface quality, the smallest (the smaller is the better) response is
sought. The lowest value for each individual parameter indicates the
level that meets the optimum condition as shown in Fig. 4. In this case, it
is clearly seen that the optimum levels are A1B2C1.
The S/N values given in Table 2 were obtained as negative quanti­
ties, among which the largest graphical value gave the smallest response
in Table 3. It is an accurate approach to find the smallest response value
from Table 3 and then to determine the optimum S/N.

3.2. Analysis of variance

ANOVA was applied by examining the data which is calculated

through Taguchi analysis in MINITAB software. Here, “P” values for
each factor were examined.
As can be seen in Table 4, the “P” value less than 0.05 belongs to
cutting tools and feed rate. When looking at the Contribution Rate (CR)
Fig. 4. Main effects plot for means.
values calculated in Eq. (2)), it can be said that the cutting tools and feed
rate are the most effective factors in the surface roughness value of the
composite material after milling. With the optimization carried out by 4. Hybrid estimation algorithm
using the Taguchi design method, it was determined that the experiment
with the combination of “A1B2C1” provided the optimal result. 4.1. Artificial neural network

Contribution Rate (CR) (%) =

Factor sum of squares
× 100 (2) ANN, which is a computational model consisting of a large number of
Total sum of squares interconnecting processing elements called neurons, is known as a
powerful approach implemented for forecasting of data in a variety of
systems. In this study, a multilayer feed-forward backpropagation ANN
architecture is utilized for the prediction of surface roughness because of
Table 2
its wide application. As it is seen from Fig. 5 that the ANN has been
Response table for signal to noise ratios (smaller is better).
constructed with four layers including the input, output, and two hidden
Level Cutting tools Spindle speed (rpm) Feed rate (mm/rev)
layers. Three neurons (x1, x2, x3) representing the cutting tool, spindle
1 − 5.514 − 6.904 − 5.373 speed, and feed rate have been used in the input layer whereas one
2 − 8.771 − 6.881 − 5.414 neuron (y1) is employed in the output layer denoting for the surface
3 7.115 9.418
roughness.
− −
4 − 7.671 − 8.366
Delta 3.257 0.789 4.045 Neurons are connected to other neurons in adjacent layers by weight
Rank 2 3 1 factor according to their importance. The inputs of each artificial neuron

563
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

Table 4
Analysis of variance for means.
Source DF Seq SS Adj SS Adj MS F P CR (%)

Cutting tools 1 10.9571 10.9571 10.9571 11.83 0.002 23.55

Spindle speed (rpm) 3 0.3482 0.3482 0.1161 0.13 0.944 0.7484
Feed rate (mm/rev) 3 12.9835 12.9835 4.3278 4.67 0.010 27.9082
Residual error 24 22.2334 22.2334 0.9264 47.7910
Total 31 46.5221

Fig. 5. Feed-forward backpropagation ANN structure with two hidden layers.

are the weighted sum of each incoming connection with a bias which is first and the second hidden layer; wks is the connection weights between
added to make the following data meaningful in the transfer function. the output layer and the second hidden layer. f1, f2 and f3 denote the
Since there is a nonlinear relationship between input and output neu­ transfer function (hyperbolic tangent sigmoid as shown in Eq. (4)); φj,
rons, the hyperbolic tangent sigmoid (tansig) is preferred as the transfer φk, and φs representing the biases on the nodes of first hidden, second
function. In order to stabilize and enhance ANN training, the input and hidden and output layers respectively.
output data have been normalized in [− 1,1] range. The output of the
ex − e− x
network can be mathematically expressed by Eq. (3). f (x) = (4)
ex + e− x
( ( ( ) ) )
The number of neurons in hidden layers, learning rate, and the
p
∑ ∑
m ∑
n
∧
y = f3 wks x f2 wjk x f1 wij x(xi ) + φj + φk + φs (3)
k=1 j=1 i=1 training function are determined by optimization using GA for the ar­
chitecture in which the ANN predicts the best. As given in Table 5, the
where ŷ is the output values and xi the input values of the network; wij, optimal number of neurons used in the hidden layers from 1 to 50 and
the connection weights between the ith node of the input layer and the jth the best learning rate from 0.01 to 0.99 are determined by taking the
node of the first hidden layer; wjk, the connection weights between the prediction accuracy of the ANN as the fitness function of the GA.

564
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

Table 5 neurons and learning rate that will ensure a satisfactory result [21].
ANN parameters tuned by GAs. There is no clear and precise technique in the literature to set these
Parameter Ranges/functions parameters, even though they determine the success of the training
process of the ANN. Recently, researchers have pointed out the suc­
Number of neurons for the first hidden layer [1, 2, 3, ….., 48, 49, 50]
[n1] cessful performance of hybrid structures in which ANN is integrated
Number of neurons for the second hidden layer [1, 2, 3, ….., 48, 49, 50] with evolutionary metaheuristic algorithms. Particle swarm optimiza­
[n2] tion [22], ant colony optimization [23], simulated annealing [24],
Learning Rate [0.01–0.99] artificial bee colony [25], bat and firefly algorithm [26], grey wolf
Training function trainlm (Levenberg-Marquardt)
trainscg (Scaled conjugate
optimizer [27], cuckoo search [28] have been implemented to optimize
gradient) the parameters directly affecting the performance of ANN. Since GA has
trainrp (Resilient strong adaptability and global optimization ability; it has been exten­
backpropagation) sively preferred in ANN parameter optimization. Previous studies draw
trainbr (Bayesian regularization)
attention that the ANN model combined with GA (ANN-GA) is more
trainbfg (BFGS quasi-Newton)
effective in finding the optimal parameters of ANN than the trial-and-
error method and most of the hybrid ANN structures [29].
Similarly, the effect of 5 different training algorithms (Levenberg-Mar­ In this study, GA is applied to optimize the structure and learning
quardt, Scaled conjugate gradient, Resilient backpropagation, Bayesian parameters of the feed-forward backpropagation ANN including the
regularization, and Broyden–Fletcher–Goldfarb–Shanno (BFGS) quasi- number of neurons in the first and the second hidden layer, the learning
Newton) on the estimation performance of the ANN is evaluated with rate, and the training function. Fig. 6 demonstrates the flowchart of the
GA and the algorithm that gives the best result is determined. The initial proposed hybrid ANN-GA algorithm. As it is seen from Fig. 6, the hybrid
sets of weights and biases are generated randomly. These values are then algorithm starts with setting of GA parameters such as population size
updated and optimized by using GA to allow the outputs of the ANN to (n), crossover probability (pc), mutation probability (pm) and the gen­
accurately fit the actual data. eration size (gm) which are listed in Table 7. The initial population
Table 6 provides the design parameters of the ANN for the MATLAB comprising of n sets of chromosomes, which represents four ANN pa­
implementation. The neural network has been trained with 70% of the rameters to be optimized (called as genes), is generated randomly. Each
experimental data and has been subsequently validated and tested with of the chromosomes in the population is a candidate solution to the
the remaining 30% of the data. The maximum number of epochs, optimization problem and is evaluated by the fitness (objective) func­
training time to minimize the error between the estimated output and tion that how close to the optimum solution. Minimization of the MSE is
the actual data, is defined as 1000 whereas the minimum performance considered as the fitness function that is already given in Eq. (5). The
gradient is selected as 0.00001. In this paper, the mean square error initial population is evolved by using three genetic operators namely
(MSE) is preferred as the fitness function of the GA to evaluate the selection, crossover, and mutation within an iterative process.
performance of the ANN model. MSE, which is obtained by dividing the The selection operator chooses the strongest chromosomes (i.e. the
sum of the squared errors by the number of observations (as formulated fittest solutions) in the population for reproduction based on a proba­
in Eq. (5)), is probably the most commonly used error metric in the bility distribution. The roulette wheel technique, which is the simplest
literature. It penalizes larger errors because squaring larger numbers has and the most commonly preferred method, is used. Each chromosome of
a greater impact than squaring smaller numbers. the population is allocated a section of an imaginary roulette wheel,
∑n ∧
which is proportionate to its fitness. The fittest candidate has the biggest
MSE = i=1 (yi − xi )2
(5) section of the wheel and the weakest candidate has the smallest [30].
n The probability of the ith chromosome according to the roulette wheel is
defined by Altay et al. [21];
where n is the number of observations, yi and xi represent the predicted
∧

fi
and measured test data respectively. Pi = ∑ n (6)
i=0 fi

4.2. Genetic algorithm where Pi is the probability of the ith chromosome and, n is the population
size. n random numbers are generated and chromosomes are chosen
ANN is known to be effective in the estimation; nevertheless, it is according to their probability intervals.
difficult to set appropriate parameters such as the number of hidden The crossover operator is used to create new chromosomes for the
next generation by randomly combining two selected chromosomes
from the current generation through the selection process. Randomly
Table 6
paired chromosomes exchange genes after an assigned crossover point
ANN design parameters for MATLAB implementation.
to make offspring solutions for the next generation. Crossover is per­
Parameter Specification
formed according to crossover probability (pc), which is generally
Neural network type Feed-forward backpropagation assigned between 0.6 and 1.0 [21].
Number of neurons in input layer 3 The mutation is a process that increases the diversity of the popu­
Number of neurons in output layer 1
lation by changing the genes in a certain position with respect to mu­
Data division 70% - Training
30% - Testing tation probability (pm). Randomly chosen gene in the entire population
Number of hidden neurons in first layer [n1] [15] (tuned with GA) is mutated by adding or subtracting a small number (2% of the input
Number of hidden neurons in second layer [29] (tuned with GA) range).
[n2]
The genetic process of selection, crossover, and mutation is per­
Training function [trainbr] (tuned with GA)
Transfer function Hyperbolic tangent sigmoid
formed on the whole population until the number of iteration reaches
[tansig] 40. When the algorithm finally ends (the GA terminated after 40 itera­
Learning rate [0.42] (tuned with GA) tions), the optimized ANN parameters are attained and the best fitness
Performance function Mean squared error (MSE) function (MSE) value is obtained as the optimal solution.
Maximum epoch (training time) 1000
Performance goal 1 × 10− 5
Normalized range [− 1,1]

565
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

Fig. 6. Flowchart of the hybrid ANN-GA algorithm.

5. Results and discussion

Table 7
Parameters of GA in BP-ANN parameter tuning.
The experiments have been conducted using Taguchi's L32 mixed
Parameter Value orthogonal array with 3 levels, as shown in Table 8 which is taken as the
Genotype (coding) Chromosomes of four genes (each real values and dataset for this study for training and testing the proposed ANN model.
representing an ANN parameter to be optimized) 22 values were used for training and 10 values were used for the testing
Fitness function Minimization of MSE as it is specified in Table 8. As can be seen from Table 8, the minimum
Generation size 40
(iteration number)
surface roughness value was obtained as 1.397 μm during the 26th
Population size (n) 50 experiment when 6000 rpm spindle speed, 500 mm/rev feed rate, and
Selection Roulette wheel technique the micro-grain carbide-C10 cutting tool is used. Due to the structure of
Crossover Single point crossover the carbon fibers, only a roughness of 1.397 μm can be achieved in this
Crossover probability 0.95 (95%)
process.
(pc)
Mutation Adding (or subtracting) a small number The ANN parameters tuned by the GA are presented in Table 6. The
Mutation probability 0.01 (1%) optimum number of neurons in the first and second hidden layers has
(pm) been determined as 15 and 29 respectively. The learning rate is opti­
mized as 0.42 and the transfer function is set as “tansig”. The mean, the
minimum, and the best fitness values (MSE) with respect to the

566
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

Table 8
Experimental data set and predicted surface roughness.
Experiment Cutting Spindle Feed Actual Predicted
number tool speed rate surface surface
(rpm) (mm/ roughness roughness
rev) (μm) (μm)

1 1 4000 250 1.648 1.803

2 1 4000 500 1.749 1.759
3 1 4000 1000 1.765 1.762
4 1 4000 1500 2.410 2.411
5a 1 5000 250 1.957 1.936
6 1 5000 500 1.780 1.776
7 1 5000 1000 1.768 1.746
8 1 5000 1500 2.286 2.286
9 1 6000 250 1.741 2.187
10a 1 6000 500 1.981 1.821
11 1 6000 1000 1.750 1.757
12 1 6000 1500 1.923 1.920
Fig. 8. The minimum and best values of fitness with respect to the generation
13 1 7000 250 2.403 2.402 size of GA.
14a 1 7000 500 1.741 1.839
15a 1 7000 1000 1.765 1.773 type and amount of data. It is clearly seen in Fig. 10 that there is a
16 1 7000 1500 1.742 1.743
17 2 4000 250 1.499 1.491
healthy correlation between training loss and test loss. Both training and
18 2 4000 500 2.547 2.544 test losses reach their minimum values after 83th epoch and remain at a
19 2 4000 1000 5.614 5.613 constant value, which means that the model is well trained and a good fit
20a 2 4000 1500 2.197 2.652 is achieved.
21a 2 5000 250 1.510 1.654
Different materials were examined in studies that have previously
22 2 5000 500 1.856 1.869
23 2 5000 1000 4.165 6.487 applied ANN for prediction of surface roughness and various perfor­
24 2 5000 1500 3.442 3.445 mance metrics were used to evaluate the accuracies of the models. In
25a 2 6000 250 1.693 2.072 some studies, the prediction performance of the ANN model has not
26a 2 6000 500 1.397 1.749 been demonstrated with concrete data. In Zhou et al. [15], surface
27 2 6000 1000 6.371 6.368
28 2 6000 1500 4.010 4.004
roughness for the milling of Al/SiC metal matrix composites were pre­
29 2 7000 250 2.727 2.729 dicted by using a feed-forward multi-layered ANN model with three
30a 2 7000 500 2.064 1.857 different training algorithms. The mean relative error (MRE) obtained
31a 2 7000 1000 4.075 4.481 by Levenberg-Marquardt, Gradient Descent and Bayesian Regulation
32 2 7000 1500 3.967 3.973
training algorithms were 3.50%, 13.93% and 3.67% respectively. In
a
Used for testing of ANN model and not included in training. Karabulut [3], ANN was able to predict the surface roughness with a
MSE of 2.25% in the milling of AA7039 and Al2O3 reinforced composite.
generation size of GA are illustrated in Figs. 7 and 8. For the test data, In Moghri et al. [11], ANN-GA approach was presented for modeling and
the best MSE is obtained as 0.074 during the 33th iteration of the GA and optimization of surface roughness in the milling Polyamide-6/Nanoclay
the correlation is calculated as 0.96177. nanocomposites. The mean absolute percentage errors (MAPE) using
Fig. 9 shows the linear correlation between the target and predicted training and testing data were found to be 2.49% and 2.99%, respec­
results for the training and test data. According to the results, the tively. In Sangwan et al. [12], the MAPE was found as 4.13% in the
regression line of test and predicted data is given by y = 1.1 * T − 0.032. prediction of surface roughness of Ti-6Al-4V titanium alloy with an
The error between predicted and measured results and the comparison ANN-GA model. In Zain et al. [2], the lowest absolute MSE average value
of them are also presented graphically in Fig. 9. It can be concluded from was obtained as 0.023 among six developed ANN models for prediction
Fig. 9 that the training error (represented in the first 22 data) and the of surface roughness of Ti-6A1-4V titanium alloy. In this study, the
testing error (represented in the last 10 data) are low which means that proposed hybrid ANN-GA algorithm is able to make an accurate pre­
the proposed ANN-GA algorithm provides satisfactory performance. diction of surface roughness of the high-strength carbon fiber composite
To assess the effectiveness and the performance of the proposed GA- plate, which is more difficult to process, by utilizing a small-sized
ANN model, the loss function has been used as presented in Fig. 10. The training and testing dataset. The absolute MSE has been obtained as
loss function graph shows that the network settings are suitable for this 0.074 and the correlation of 96.177% has been achieved.

6. Conclusion

This paper proposes a combined study involving the experimental

investigation of the effects of different machining parameters such as
cutting tools, feed rate, and spindle speed on the surface roughness of
the high-strength carbon fiber composite plate and the estimation of
surface roughness implementing an improved hybrid algorithm. Tagu­
chi optimization method based on a mixed orthogonal array L32 (21 ×
42) has been used to design the experiments. ANOVA analysis is per­
formed to determine the effect of machining parameters on the surface
roughness by using MINITAB software. It has been found that the cutting
tool and the feed rate are the most effective factors on the surface
roughness of the composite material. The optimal parametric combi­
nation has been determined as “A1B2C1” (TiAlN coated cutting tool,
5000 rpm spindle speed, and 250 mm/rev feed rate). To estimate the
surface roughness, a feed-forward backpropagation ANN model with
Fig. 7. The mean fitness values with respect to the generation size of GA.

567
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

Fig. 9. Correlation between target and predicted results for the training and test data.

GA.
Further research studies will be focused on hybridization of the ANN
with different new metaheuristic approaches and comparing them with
ANN-GA approach for solving such a problem presented in the manu­
script in an effective manner.

Declaration of competing interest

The authors declare that they have no known competing financial

interests or personal relationships that could have appeared to influence
the work reported in this paper.

References
Fig. 10. Trend of loss versus the number of epochs. [1] Groover MP. Fundamentals of modern manufacturing. New Jersey: Prentice Hall;
1996.
[2] Zain AM, Haron H, Sharif S. Prediction of surface roughness in the end milling
parameter tuning assisted by GA is developed using MATLAB® 9.6
machining using Artificial Neural Network. Expert Syst Appl 2010;37:1755–68.
software. GA has provided a significant reduction in time and effort in https://doi.org/10.1016/j.eswa.2009.07.033.
the selection of optimum parameters and the structure of the ANN as [3] Karabulut Ş. Optimization of surface roughness and cutting force during AA7039/
Al2O3 metal matrix composites milling using neural networks and Taguchi method.
compared to the trial and error approach. As a result of the optimization
Measurement. 2015;66:139–49. https://doi.org/10.1016/j.
with GA, Bayesian regularization is found to be more suitable training measurement.2015.01.027.
algorithm for multi-layer ANN for better prediction performance. The [4] Sreenivasulu R. Optimization of surface roughness and delamination damage of
learning rate is optimized as 0.42 while the number of neurons in hidden GFRP composite material in end milling using Taguchi design method and artificial
neural network. Procedia Eng 2013;64:785–94. https://doi.org/10.1016/j.
layers are specified as 15 and 29 respectively. The proposed hybrid proeng.2013.09.154.
ANN-GA algorithm provides superior relationship between the esti­ [5] Muthukrishnan N, Dawim JP. Optimization of machining parameters of Al/SiC-
mated and measured values of surface roughness. The results have MMC with ANOVA and ANN analysis. J Mater Process Technol 2009;209:225–32.
https://doi.org/10.1016/j.jmatprotec.2008.01.041.
shown that the best coefficient of correlation is 0.96177 and the MSE is [6] Varol T, Canakci A, Ozsahin S. Artificial neural network modeling to effect of
0.074 obtained with the test data using ANN during 33th iteration of the reinforcement properties on the physical and mechanical properties of Al2024–B4C

568
C. Boga and T. Koroglu Journal of Manufacturing Processes 70 (2021) 560–569

composites produced by powder metallurgy. Compos Part B 2013;54:224–33. [18] Zahaf MZ, Benghersallah M. Surface roughness and vibration analysis in end
https://doi.org/10.1016/j.compositesb.2013.05.015. milling of annealed and hardened bearing steel. Measurement: Sensors 2021;13:
[7] Daniel SAA, Pugazhenthi R, Kumar E, Vijayananth S. Multi objective prediction 100035. https://doi.org/10.1016/j.measen.2020.100035.
and optimization control parameters in the milling of aluminum hybrid matrix [19] Ramasamy M, Daniel AA, Nithya M. Investigation on surface roughness of
composites using ANN and Taguchi-grey relations analysis. Def Technol 2019;15: aluminium (Al7050/TiC/BN) hybrid metal matrix. Mater Today 2021. https://doi.
545–56. https://doi.org/10.1016/j.dt.2019.01.001. org/10.1016/j.matpr.2020.12.852.
[8] Karabulut Ş, Gökmen U, Çinici H. Optimization of machining conditions for surface [20] Liu N, Liu B, Jiang H, Wu S, Yang C, Chen Y. Study on vibration and surface
quality in milling AA703-based metal matrix composites. Arab J Sci Eng 2018;43 roughness in MQCL turning of stainless steel. J Manuf Process 2021;65:343–53.
(3):1071–82. https://doi.org/10.1007/s13369-017-2691-z. https://doi.org/10.1016/j.jmapro.2021.03.041.
[9] Karabulut Ş, Karakoç H. Investigation of surface roughness in the milling of Al7075 [21] Altay A, Ozkan O, Kayakutlu G. Prediction of aircraft failure times using artificial
and open-cell SiC foam composite and optimization of machining parameters. neural networks and genetic algorithms. J Aircr 2014;51:47–53. https://doi.org/
Neural Comput & Applic 2017;28:313–27. https://doi.org/10.1007/s00521-015- 10.2514/1.C031793.
2058-x. [22] Gambhir S, Malik SK, Kumar Y. PSO-ANN based diagnostic model for the early
[10] Öktem H, Erzurumlu T, Çöl M. A study of the Taguchi optimization method for detection of dengue disease. New Horiz Transl Med 2017;4(1–4):1–8. https://doi.
surface roughness in finish milling of mold surfaces. Int J Adv Manuf Technol org/10.1016/j.nhtm.2017.10.001.
2006;28:694–700. https://doi.org/10.1007/s00170-004-2435-6. [23] Mohammadhassani J, Dadvand A, Khalilarya S, Solimanpur M. Prediction and
[11] Moghri M, Madic M, Omidi M, Farahnakian M. Surface roughness optimization of reduction of diesel engine emissions using a combined ANN–ACO method. Appl
polyamide-6/nanoclay nanocomposites using artificial neural network: genetic Soft Comput 2015;34:139–50. https://doi.org/10.1016/j.asoc.2015.04.059.
algorithm approach [Hindawi Publishing Corporation. Article ID 485205] Sci [24] Manoochehri M, Kolahan F. Integration of artificial neural network and simulated
World J 2014:1–7. https://doi.org/10.1155/2014/485205. annealing algorithm to optimize deep drawing process. Int J Adv Manuf Technol
[12] Sangwan KS, Saxena S, Kant G. Optimization of machining parameters to minimize 2014;73:241–9. https://doi.org/10.1007/s00170-014-5788-5.
surface roughness using integrated ANN-GA approach. Proc CIRP 2015;29:305–10. [25] Taheri K, Hasanipanah M, Golzar SB, Majid MZA. A hybrid artificial bee colony
https://doi.org/10.1016/j.procir.2015.02.002. algorithm-artificial neural network for forecasting the blast-produced ground
[13] Goopan V, Wins LDK, Surendran A. Integrated ANN-GA approach for predictive vibration. Eng Comput 2017;33:689–700. https://doi.org/10.1007/s00366-016-
modeling and optimization of grinding parameters with surface roughness as the 0497-3.
response. Mater Today 2018;5(5):12133–41. https://doi.org/10.1016/j. [26] Mostafaeipour A, Goli A, Qolipour M. Prediction of air travel demand using a
matpr.2018.02.191. hybrid artificial neural network (ANN) with bat and firefly algorithms: a case
[14] Kıvak T. Optimization of surface roughness and flank wear using the Taguchi study. J Supercomput 2018;74:5461–84. https://doi.org/10.1007/s11227-018-
method in milling of Hadfield steel with PVD and CVD coated inserts. 2452-0.
Measurement. 2014;50:19–28. https://doi.org/10.1016/j. [27] Golafshani EM, Behnood A, Arashpour M. Predicting the compressive strength of
measurement.2013.12.017. normal and High-Performance Concretes using ANN and ANFIS hybridized with
[15] Zhou G, Xu C, Ma Y, Wang XH, Feng PF, Zhang M. Prediction and control of surface Grey Wolf Optimizer. Constr Build Mater 2020;232:1–14. https://doi.org/
roughness for the milling of Al/SiC metal matrix composites based on neural 10.1016/j.conbuildmat.2019.117266.
networks. Adv Manuf 2020;8:486–507. https://doi.org/10.1007/s40436-020- [28] Valian E, Mohanna S, Tavakoli S. Improved cuckoo search algorithm for feed
00326-x. forward neural network training. Int J Artif Intell Applic 2011;2(3):36–43. https://
[16] Laot PAL, Suhardjono Sutikno, Sampurno. Multi-response optimization of cutting doi.org/10.5121/ijaia.2011.2304.
force and surface roughness in carbon fiber reinforced polymer end milling using [29] Wahono RS, Herman NS, Ahmad S. Neural network parameter optimization based
back propagation neural network and genetic algorithm. AIP Conf Proc 2019;2114: on genetic algorithm for software defect prediction. Adv Sci Lett 2014;20:1951–5.
020001. https://doi.org/10.1063/1.5112385. https://doi.org/10.1166/asl.2014.5641.
[17] Mohsin I, He K, Li Z, Zhang F, Du R. Optimization of the polishing efficiency and [30] Yadav SL, Sohal A. Study of the various selection techniques in genetic algorithms.
torque by using Taguchi method and ANOVA in robotic polishing. Appl Sci 2020; Int J Eng Sci Math 2017;6(3):198–204.
10(3):824. https://doi.org/10.3390/app10030824.

569