International Journal of Current Engineering and Technology

E-ISSN 2277 – 4106, P-ISSN 2347 - 5161

®
©2014 INPRESSCO , All Rights Reserved
Available at http://inpressco.com/category/ijcet

Research Article
Artificial Neural Network Modeling of the Effect of Cutting Conditions on
Cutting Force Components during Orthogonal Turning
Sanjib BanerjeeȦ*, Angshuman DekaȦ, Gitartha Dev SarmahȦ, Nitish BhardwajȦ
Ȧ
Department of Mechanical Engineering, Tezpur University, 784028, Tezpur, India

Accepted 16 January 2014, Available online 01 February 2014, Special Issue-2, (February 2014)

Abstract

Variation in cutting force components (FX, FY and FZ in three mutually perpendicular directions X, Y and Z) with cutting
conditions viz. speed (v), feed (f) and depth of cut (d) during orthogonal turning of mild steel specimen using a HSS
cutting tool was investigated in an automatic lathe machine. During the turning process, the cutting forces experienced
insignificant variations with change in speed. Cutting force values were however observed to considerably increase with
increase in feed and depth of cut. Subsequently these cutting forces could be modelled as function of the cutting
conditions viz. v, f and d, by artificial neural network (ANN). The cutting force values modelled and subsequently
predicted at various cutting conditions within the specified domain have been successfully correlated with the
experimental results and literature review, with fairly good accuracy. During the validation process, it was possible to
predict 100%, 94% and 100% of the cutting force values for F X, FY and FZ, respectively within a percentage deviation of
±10%. This observation highlights the superior prediction capability of ANN technique in the current research area.

Keywords: Cutting Forces; Cutting Conditions; Artificial neural network; Orthogonal Turning; Mild Steel.

1. Introduction work piece, the cutting force experienced by the tool may
be considered as functions of the cutting conditions viz. v,
1
It has been realized by most of the manufacturing f and d. These force values of a particular material can be
engineers and researchers that in order to optimize the modeled from the experimental results, by using artificial
efficiency of metal cutting operations, efficient neural network (ANN) (Suksawat, 2010). The cutting
quantitative and predictive models that establish a force components can then be successfully predicted for
relationship between a big group of independent input any combination of the input variables within the specific
parameters and output variables are required for the wide domain. Such prediction provides immense industrial help
spectrum of manufacturing processes, cutting tools and in formulating the cutting conditions on the basis of
conditions, and engineering materials currently used in the cutting forces developed and thus deciding the subsequent
industries (Wong, et al, 1999). Furthermore, it has been manufacturing parameters required during orthogonal
observed that proper optimization of the output variables, turning of mild steel. Out of these three forces, FY shows
such as tool life, cutting forces, surface roughness, etc., the most considerable effect during orthogonal turning
through the optimization of input parameters like cutting operation as power required for machining is determined
speed, feed rate and depth of cut, results in a significant by the cutting force component, FY.
improvement in the economic performance of machining A neural network is similar to the biological nervous
operations (Li and Li, 2002). One of these output variables system, which is basically a connectionist system, in
that may either directly or indirectly affect the which various nodes called neurons are interconnected. An
performance of other variables such as tool wear rate, artificial neural network (ANN) can be defined as a model
machined surface characteristics and machining costs is of reasoning similar to the human brain, where a large
cutting force (Kadirgama, et al, 2008). Cutting forces amount of complex information can be stored and
directly relate with the tool wear as higher the forces processed simultaneously by each neuron along the entire
experienced by the tool; higher will be the tool wear rate. domain. A typical neuron receives one or more input
Subsequently the cutting forces also have an impact on the signals and provides an output signal depending on the
tool life, cutting cost, machine vibrations, and dimensional processing function of the neuron. This output is
accuracies, cutting power, and machined surface transferred to connected neurons in varying intensities, the
characteristics. The same fact reveals the significance of signal intensity being decided by the weights assigned.
investigating and modeling of cutting forces (Yun and ANN has the advantages of (i) modelling the data where
Cho, 2001). During orthogonal turning of a cylindrical the input-output relation is either unknown or nonlinear,
(ii) adaptive learning during training (iii) real time
*Corresponding author: Sanjib Banerjee
DOI: http://dx.doi.org/10.14741/ijcet/spl.2.2014.23 127 | International Conference on Advances in Mechanical Sciences 2014
Sanjib Banerjee et al International Journal of Current Engineering and Technology, Special Issue-2 (Feb 2014)

applications aided by a very fast computational speed. The not be taken from a single rod because this would have
conventional computational techniques follow an required a rod of substantial length and resulted in
algorithmic approach, where a set of instructions in the considerable vibration. So, it was decided to take 16
specified order is followed to solve a problem. The readings in a rod of particular depth of cut. The details
relationship between each stage is required for solving the regarding the cutting conditions investigated have been
problem. On the other hand, neural network technique is shown in Table 1. The turning operation was carried out in
data driven and the solution can be obtained even if the an automatic lathe machine (MAKE: HMT MACHINE
exact relationship is unknown. This can be obtained if a TOOLS LIMITED, MODEL: NH26). High speed steel
number of input-output data sets are available. Once the (HSS) was used as the cutting tool material. HSS are
ANN architecture is fixed, the output for any combination basically high carbon steels with significant quantities of
of input variables can be predicted. W, Mo, Cr, V and Co.
The most popular neural networks are feed forward During each turning operation, the cutting forces
networks. Feed forward network architecture consists of viz. FX, FY and FZ were measured in three mutually
three distinct layers: the input layer, the hidden layer(s), perpendicular directions, X, Y and Z, as indicated:
and the output layer. Each layer consists of a number of FX = Acts in the direction of the feed (axial direction) and
neurons. The output from the neurons of one layer is is called the feed force. If unchecked it can lead to
transferred as input to neurons of the succeeding layer. dimensional inaccuracy and vibration.
The first layer, called an input layer, receives data from FY = Acts in the tangential direction and is called the
the outside world. The second layer, called the hidden cutting force or power component.
layer, does not have any direct contact with the outside FZ = It acts in the radial direction and is called the thrust
world and is used to help in extracting higher-level force (Jain, 2007; Lal 2007).
features and to facilitate generalization of outputs. The last Out of these three force components experienced
layer is the output layer, which sends information out to during an orthogonal turning operation, the maximum
users. For a given input vector, it generates the output value is for cutting force followed by thrust force and feed
vector by a forward pass. The data are fed to the network force. Thus cutting force has the maximum effect on the
at the input layer and propagated with weights and tool wear and tool life (Altintas, 2000; Gokkaya 2010).
activation functions to the output layer to provide the Hence for optimization purpose, this force, FY, is most
response. Once the data at the output neuron is reached, significant to be analyzed.
the mean squared error (MSE), which is the difference The cutting forces were measured by a piezoelectric
between the network output vector and the known target tool dynamometer (MAKE: IEICOS LATHE TOOL
vector, is computed and back-propagated to modify the DYNAMOMETER, MODEL: 620B) along with a digital
weights for the entire network. multi component force indicator (MODEL: 652). In order
For orthogonal turning of mild steel, the input layer of to eliminate experimental errors, five consecutive readings
ANN consists of three neurons representing the three of each parameter were taken at a particular set of cutting
cutting conditions viz., v, f and d, as shown in Fig. 1. The condition and the average of these five was considered for
output layer on the other hand consists of one neuron subsequent modeling.
representing FX, FY or FZ. Once the proper network After evaluating cutting force values for different
architecture is determined, FX, FY and FZ can be cutting conditions, the same were modeled by artificial
successfully predicted for any combination of the input neural network (ANN). The ANN modeling was carried
variables within the specified domain (Banerjee, et al, out by the multiple layer perception (MLP) feed forward
2011). back propagation network. The cutting conditions viz. v, f
and d representing three neurons in the input layer and
v cutting force components representing the neuron in the
output layer formed the data sets for training the network.
‘Neural Network’ tool box available with the MATLAB
f
Fx, Fy, Fz (Release 7) software package was used for the present
modeling. Training of the neural network was done using
d the ANN tool kit of MATLAB software, using
‘TRAINLM’ function. ‘TRAINLM’ is a network training
function that updates weights and bias values in a back
propagation algorithm according to Levenberg–Marquardt
optimization. Levenberg–Marquardt algorithm is a highly
Fig. 1 A typical neural network architectre for modeling efficient method for solving non-linear optimization
cutting force problems (Robi and Dixit, 2003). Single layer hidden
neurons were used in the network architecture. The
2. Experimental procedures number of neurons in the hidden layer, the transfer
functions at the input-to-hidden layer and hidden-to-output
Four numbers of mild steel rods of 24 mm diameter were layer were optimized by trial and error method during the
taken up for orthogonal turning operation at four different network training and testing stages. The mean square error
depths of cuts. The total of 64 readings (the combination (MSE) during the training and testing was determined for
of four speeds, four feeds and four depths of cuts) could each trial. The network architecture was finally frozen

128 | International Conference on Advances in Mechanical Sciences 2014

Sanjib Banerjee et al International Journal of Current Engineering and Technology, Special Issue-2 (Feb 2014)

based on the minimum MSE value obtained during both Table 2 Best fit network architectures for the investigated
the training and testing stages. Out of the total 64 number cutting force components
of data sets, 32 and 20 data sets were used for respectively
training and testing operations. 12 data sets from the Force Hidden 1st Transfer 2nd Transfer
remaining were used for validating the trained network Component Neurons Function Function
FX 5 logsig purelin
architecture. Once the proper network architecture was FY 6 logsig purelin
arrived at, the cutting force components could be FZ 4 logsig purelin
successfully predicted for any combination of input
variables (v, f and d), within the specified domain.

Table 1 Different Cutting Conditions for prediction of

Cutting Forces
Depth of Cut(mm) Feed(mm/rev) Speed(RPM)
0.24, 0.50, 0.64 and 0.05, 0.20, 0.52 and 114, 325, 550 and
0.80 0.80 930

3. Results and Discussion

3.1 Experimental and Predicted Curves for Cutting Forces

(a)
The experimental values were obtained from the turning
operation of mild steel. The results revealed that the
cutting force almost remains constant with change in
cutting speed while it increases with increase in feed and
depth of cut, as shown in Figure 2. These trends are well
correlated with the available open literature and earlier
studies (Satyanarayana, 2011; Malagi and Rajesh, 2012;
Kanthraj, et al, 2012).

3.2 Modeling and Prediction of Cutting forces by Artificial

Neural Network
(b)
In order to obtain the prediction of cutting forces with
variation in cutting parameters, artificial neural network
(ANN) modeling was carried out. The cutting force
components being functions of the cutting conditions can
be expressed as:
FX, FY, FZ = F (v, f, d) (1)

The ANN architecture used in present modeling consists

of three layers, viz. an input layer of three neurons
representing the cutting conditions of v, f and d, a hidden
layer of n numbers of neurons and an output layer with
one neuron representing the cutting force components. (c)
Appropriate values for the weight and bias of the network
architecture were arrived at by training the network using Fig. 2 Variation of FY with (a) Speed and (b) Feed at
several data sets. Out of total number of 64 input-output d=0.64 mm and (c) Depth of Cut at V=550 rpm
data sets obtained from experimentations, it was decided
to use 32, 20 and 12 data sets for the network training, neural network architecture was frozen for which (i)
testing and validation purposes respectively. The functional RMS error was minimum, (ii) minimum
validation data sets were not used earlier for the testing or number of data sets has a deviation error of 10 %, (iii)
training purposes. Both training and testing of the network maximum deviation during testing and training is within
were carried our independently. A number of numerical 20 %, and (iv) minimum variation in RMS functional error
trials were carried out with single hidden layer neural during training and testing. Table 2 shows the details of
network. The tansig, logsig and purelin transfer functions the best fit ANN architecture for cutting force
were tried with while simultaneously varying the number components. Once the network was frozen, it was required
of neurons in the hidden layer, in order to arrive at the best to establish the confidence in the network architecture for
network architecture. The sum squared training error goal the prediction of FX, FY or FZ. Therefore, the trained
for FX, FY and FZ was fixed at 0.00001. After a number of network was used for prediction of FX, FY or FZ using the
trials with various initial weights and biases, the best 12 validation data sets.

129 | International Conference on Advances in Mechanical Sciences 2014

Sanjib Banerjee et al International Journal of Current Engineering and Technology, Special Issue-2 (Feb 2014)

Table 3 Errors in Prediction of Cutting Force Component the specified domain. The above ANN modeling may thus
be successfully used in formulating the cutting conditions
Force Training (speed, feed and depth of cut) on the basis of Cutting
Component Max. Error Max. % Error RMSD Error Forces as demanded by a specific application.
FX 3.053 23.138 1.281
FY 5.240 10.322 2.915
FZ 4.567 10.953 1.150 Conclusions
Testing
FX 3.287 17.611 1.451 1) The cutting force components were investigated
FY 4.709 10.328 2.795 during orthogonal turning operation for various
FZ 0.876 7.614 1.183
cutting conditions of speed, feed and depth of cut.
Validation
FX 2.810 17.745 1.065 2) The experimental results revealed that the cutting
FY 4.333 8.597 3.160 force components almost remain constant with
FZ 1.078 7.704 1.170 increase in cutting speed while they considerably
increase with increase in feed and depth of cut.
3) The cutting force components were modeled as
function of three independent cutting parameters
viz. speed, feed and depth of cut, by artificial
neural network (ANN).
4) The cutting forces subsequently predicted at various
cutting conditions within the specified domain
have been correlated with the experimental results
with fairly good accuracy. During the validation
process, it was possible to predict 100%, 94% and
100% of the cutting force values for FX, FY and FZ,
respectively within a percentage deviation of ±10%.
This observation highlights the superior prediction
capability of ANN technique in the present
Fig. 3 Variation of predicted cutting force parameters with research area.
experimental force parameters for the Validation data set
of FY References

X.P. Li, A.Y.C. Nee, Y.S. Wong and H.Q. Zheng, (1999) Theoretical modeling and
Table 3 gives the values of maximum absolute error, simulation of milling force, Journal of Material processing Technology, 89, 266.
maximum percentage error and RMS error calculated for H.Z. Li and X.P. Li, (2002), Milling force prediction using a dynamic shear length
model, International, Journal of Machine Tools Manufacturing, 42, 277.
FX, FY and FZ during training, testing and validation stages K. Kadirgama, M.M. Noor, N.M. Zuki, M.R.M. Rejab and B. Mohammad, (2008),
separately. Figure 3 shows the plots of experimental vs. Prediction Of Cutting Force By Numerical Solution And Statistic Method In
predicted values of FY during validation. The FX and FZ End-Milling Operation, KLCC, Kuala lampur, 466.
W.S. Yun and D.W. Cho, (2001) , Accurate 3-D cutting force prediction using
values showed similar prediction accuracy as that of FY. cutting condition independent coefficients in end milling, International Journal
For perfect prediction, all the points should lie on a line of Machine Tools and Manufacturing, 41, 463.
B. Suksawat, (2010), Chip Form Classification and Main Cutting Force Prediction
inclined at 45° to the x-axis shown in the figure. Dashed of Cast Nylon in Turning Operation Using Artificial Neural Network,
lines representing the boundaries of ±10 % deviations are International Conference on Control, Automation and Systems, Gyeonggi-do,
172.
also shown in the figures. Figures reveal that most of the S. Banerjee, P.S. Robi and A. Srinivasan, (2011), Prediction of High Temperature
points lie very close to the line of prefect prediction. The Flow Behavior of Microalloyed Al-Cu-Mg Alloys by Using Artificial Neural
RMS error in the prediction of FY is 3.16. The maximum Networks, International Conference on Computational Methods in
Manufacturing, IIT Guwahati, 355.
error in the prediction is 4.333, where the percentage error R.K. Jain, (2007), Production Technology, Khanna Publishers, 13, 92.
is 8.597%. The RMS error in the prediction of FX and FZ G.K. Lal, (2007), Introduction to Machining Science, New Age International
Publishers, 2, 95.
are 1.065 and 1.170 respectively. The maximum errors in Y. Altintas, (2000), Manufacturing Automation, Cambridge University Press, 60,
the prediction are 2.81 and 1.078, whereas the percentage 67.
Gokkaya, (2010), The Effects of Machining Parameters on Cutting Forces, Surface
errors are 17.745 % and 7.704 % respectively. The above Roughness, Built-Up Edge (BUE) and Built-Up Layer (BUL) During Machining
observation shows very good prediction of FX, FY or FZ by AA2014 (T4) Alloy, Journal Of Mechanical Engineering, 56, 589.
the present ANN modeling. This observation together with P.S. Robi and U.S. Dixit, (2003), Application of neural networks in generating
processing map for hot working, Journal of Materials Processing Technology,
the low RMS errors registered, highlight the superior 142, 289.
prediction capability of this technique. The prediction also Machine Tool Design Handbook, (1982), Tata McGraw-Hill Publishing Company
Limited, 12.
indicates that for most of the cases, the predicted values of K. Satyanarayana, A. Venugopal and G. Venkateshwara Rao, (2011), Effect Of
the cutting force components hold similar trend with the Rake Angle And Feed Rate On Cutting Forces In An Orthogonal Turning
Process, ICTMIE, Bangkok.
experimental results. For all the three force components R.R. Malagi and B.C. Rajesh, (2012), Factors Influencing Cutting Forces in
FX, FY and FZ, the predicted cutting force values almost Turning and Development of Software to estimate Cutting Forces in Turning,
remains constant with increase in cutting speed but IJEIT, 2, 37.
L.L.R. Rodrigues, A.N. Kantharaj, B. Kanthraj, W.R.C. Freitas, B.R.N Murthy,
increases with increase in feed and depth of cut. Figure 2 (2012), Effect of Cutting Parameters on Surface Roughness and Cutting
shows the close comparison between experimental and Force in Turning Mild Steel, Research Journal of Recent Sciences, 1, 19.
predicted values of FY for different cutting conditions,
revealing the accuracy of the present ANN model within
130 | International Conference on Advances in Mechanical Sciences 2014