Proceedings of the 7th

World Congress on Intelligent Control and Automation

June 25 - 27, 2008, Chongqing, China

Modeling of Dimensional Errors in Slender Bar Turning

Using Artificial Neural Networks
Bodi Cui and Rongdi Han
Department of Mechanical Manufacture and Automation
Harbin Institute of Technology
Harbin 150001, Heilongjiang Province, P.R. China
Mailbox: 422#; Telephone: 86-451-86413257; Email: [email protected]

Abstract - Dimensional error is one of the most important were more accurate than regression analysis models that were
product quality characteristics during slender bar turning developed for comparison purposes. Suneel et al. developed a
operations. In this study, an artificial neural network model was neural network predictive model of dimensional errors and
developed to describe the relationship between cutting form accuracy to correct the CNC code [8]. Significant
parameters and dimensional errors in slender bar turning
improvement in dimensional accuracy of parts was seen in
process. For including the effect of part stiffness on dimensional
errors, the diameter and the ratio of length to diameter of the their study.
machined parts were considered as independent input However, rather less attention has been paid to the
parameters of the model. Experimental results showed that the prediction of dimensional errors during the turning process of
developed model had the ability to predict efficiently dimensional slender bar. Risbood et al. developed a neural network model
errors within the range in which they had been trained. Finally, to predict dimensional deviation by measuring cutting forces
the obtained model was used to investigate the influence of and vibrations during the slender bar turning [9]. Nevertheless,
cutting parameters on dimensional errors. the error in prediction was quite higher and the range of part
dimension was not wide.
Index Terms - slender bar; artificial neural network; modeling; In this study, dimensional errors in slender bar turning
dimensional errors; turning
were investigated by the use of artificial neural networks.
I. INTRODUCTION Further, the obtained model was used to study the influences
of the individual cutting parameters on dimensional errors.
Slender bar play a crucial part in mechanical industry. It Experiments were carried out to support the developed model.
is also well known that slender bar turning operation is much
more difficult than ordinary shafts turning operation duo to its II. ARTIFICIAL NEURAL NETWORKS MODELING
low stiffness. During turning process, the deflection of slender Artificial neural network is one of the most powerful
part is the main source of dimensional errors, while the computer modeling techniques, currently being used in many
deflections of part holder and tool holder have important fields of engineering for modeling complex relationships
contribution to dimensional errors in the non-slender bar which are difficult to describe with physical models. It mimics
turning. Thus the approaches about the turning of non-slender the ability of biological neural systems to perform complex
bar are not well suited for the case of slender bar. For decision-making tasks without prior programming, and has
predicting the dimensional errors accurately, the development shown powerful capacity in solving high nonlinearity and
of a special model for slender bar turning is very important. uncertainty problems. For machining operations, the process
Investigations on dimensional errors in turning have has high nonlinearity, time variability and uncertainty. Thus,
received a great deal of attention in the literatures. Kops et al. neural network technique is an ideal approach for modeling
dealt with the dimensional errors due to workpiece deflection the machining process. There has been continual increase in
by an analytical model, which accounts for the existence of the applications of neural network to model machining
two diameters during the bar turning, and then built a operations [10-15]. In this research, this technique was
relationship between deflection and depth of cut in order to employed to perform the modeling of the slender bar turning.
get the predictive compensation [1]. Mayer et al. developed a
finite element model with closed-form solutions to predict A. Experiments
dimensional errors in bar turning [2]. Phan et al. developed an The training data and testing data were collected from
improved finite element model with closed-form solutions to experiments. For minimizing the number of experiments,
predict dimensional errors in turning cantilever multi-diameter five-factor three-level design was used here for the collection
bars [3]. Qiang developed a finite difference analysis method of training data. Table I listed the values of the five-factor
for calculating the deformation of multi-diameter workpieces three-level design. The ranges of machining condition were
during turning [4]. Carrino et al. proposed a model to evaluate selected from the recommendations given by tool
dimensional errors in bar turning based on the unified manufacturers and machining data handbook.
generalized mechanics of cutting approach [5, 6]. Azouzi et al. Machining experiments were conducted on CA6140 lathe.
proposed a neural network model based sensor fusion to C45 steel workpiece was turned with P10 inserts. The material
evaluate surface roughness and dimensional deviation during characteristics of the workpiece used in this work were listed
turning [7]. The results showed that neural network models in Table II. The cutting tools employed had the following

978-1-4244-2114-5/08/$25.00 © 2008 IEEE. 6053

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on February 09,2023 at 04:56:18 UTC from IEEE Xplore. Restrictions apply.
 TABLE I algorithm with Bayesian regularization was used to train the
EXPERIMENTAL FACTORS AND LEVELS FOR TRAINING DATA
neural network model. For all networks tangent sigmoid
Parameters Symbol Unit Values of levels transfer function and linear transfer function were adopted in
Cutting speed v m/min 30 45 60 the hidden and output layer, respectively.
Feed f mm/rev 0.1 0.15 0.2 A number of different network topologies, consisting of
Depth of cut ap mm 0.1 0.2 0.3
one or two hidden layers and different numbers of neurons in
hidden layers, were considered. Results showed that the
Diameter D mm 20 25 30
network with one hidden layer and four neurons in hidden
Slenderness ratio L/D 10 20 30 layer gave the best performance for prediction of dimensional
errors. The topology of selected networks was shown in Fig.1.
geometrical parameters: tool cutting edge angle, r = 75° ; D. Testing of Neural Network
tool orthogonal rake angle,  0 = 18° ; tool orthogonal In order to verify the established neural network model,
clearance angle,  0 = 3° and tool cutting edge inclination
experiments were performed to collect testing data. The
angle,  = 5° . Coolant was present in all the experiments.
testing data consisted of two parts. The first part of the testing
The dimensional errors of machined parts were measured
data was collected with different machining conditions shown
by micrometer. The measurements were taken at 12
in table III. The other testing data were obtained randomly
predetermined points along the machined parts. For a chosen
within the scope in which the model had been trained. The
axial location, three readings were recorded by orienting the
performance of the developed model was shown in Fig. 2. The
instrument at intervals of about 120 degrees. The average of
correlation coefficient of 0.9925 was obtained between all the
three readings was used as the dimensional error of the
testing dataset and the model predictions. The maximum of
measured location. The final dimensional error of the part was
the percentage error was 13.39%, and 80% of the testing
determined by the maximum of the values measured at all
dataset had the percentage error ranging between ±10%. These
measured points.
demonstrated that the obtained model had high accuracy for
B. Preprocessing of Dataset predicting the dimensional errors.
Before training the neural networks, the input and output
dataset were normalized within the range of ±1. The III. SIMULATION ANALYSIS
normalized value (xi) for each raw input/output dataset (di) In the previous section, neural network model was
was calculated as developed to describe the relationship between machining
2(d i − d min ) conditions and dimensional errors. This section used the
xi= −1, (1) obtained neural network model to study the influence of the
d max − d min
individual cutting parameters on dimensional errors within the
where dmax and dmin are the maximum and minimum values of range in which the model had been trained. Simulations were
the raw data. performed with fixed parts diameter D=24mm and slenderness
C. Neural Network Design and Training
During slender bar turning processes, the deflection of Hidden layer
Input layer Output layer
part is the main source of dimensional errors. Thus, the v
characteristics of part stiffness need to be considered into the
model of slender bar turning. In order to include the effect of f
part stiffness on dimensional errors, the diameter and the ratio
Dimensional error
of length to diameter of the machined parts were chosen as ap
independent input parameters of the developed models in this
study. The cutting parameters considered in this study were D
cutting speed, feed and depth of cut.
The network architecture or features, such as number of L/D
layers and neurons, are very important factors that determine Fig.1 Neural network used in predicting
the functionality and generalization capability of the trained dimensional errors
network. For this model, standard multilayer feedforward
neural network was trained with MATLAB. In the design of
TABLE III
neural networks, major concern was to obtain a good EXPERIMENTAL FACTORS AND LEVELS FOR TESTING DATA
generalization capability. Thus, Levenberg–Marquardt training
Parameters Unit Values of levels
TABLE II Cutting speed m/min 35 40 50 55
MATERIAL CHARACTERISTICS OF THE WORKPIECE
Feed mm/rev 0.12 0.18
Young’s Modulus of elasticity E=2.13 × 10 MPa
5
Depth of cut mm 0.15 0.25
Tensile strength  =598MPa
Diameter mm 23 28
Density  =7.85 × 103Kg/m3
Slenderness ratio 15 25

6054

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on February 09,2023 at 04:56:18 UTC from IEEE Xplore. Restrictions apply.
 50 50
ap=0.3 mm
ap=0.2 mm
40 40 ap=0.1 mm

Dimensional errors (μm)

Predictions ( μm)

30 30

20 20

10 10

0 0
0 10 20 30 40 50
0.1 0.12 0.14 0.16 0.18 0.2
Experimental data ( μm)
Feed (mm/rev)
Fig.2 Comparison between experimental results and Fig.4 Relation between feed and dimensional errors
predictions by neural network model (v=45m/min)

ratio L/D=22. The predicted results are illustrated in Fig. 3-5. 50

Fig. 3 shows the influence of the cutting speed v on
dimensional errors with three different feed levels f and a 40

Dimensional errors (μm)

fixed depth of cut ap=0.24mm. It can be found that as cutting
speed increases, dimensional errors decreases slightly. 30
The predicted relationship between dimensional errors
and feed is shown in Fig. 4 with three different values of the
20
cutting depth and a fixed cutting speed v=45m/min. As feed
increases, dimensional errors increase gradually.
Fig. 5 shows the influence of the depth of cut on 10 f =0.2 mm/rev
f =0.15 mm/rev
dimensional errors with three different feed levels and a fixed f =0.1 mm/rev
cutting speed v=45m/min. It can be observed that depth of cut 0
0.1 0.15 0.2 0.25 0.3
has a significant effect on dimensional errors. As depth of cut Depth of cut (mm)
increases, dimensional errors increase significantly. Fig.5 Relation between depth of cut and dimensional errors
At the same time, it can also be found that the cutting (v=45m/min)
parameters mostly affecting dimensional errors are the depth
of cut and feed by comparing the three figures. scope of cutting conditions investigated in this study, the
These results can be explained as follows: cutting increased depth of cut or feed result in the increase of cutting
parameters have different influences on cutting force, which force, whereas the increased cutting speed result in
are recognized to have a significant effect on dimensional the decrease of cutting force. Further, the cutting forces are
errors of the part. The increase of cutting force can result in more sensitive to changes in the feed and the depth of cut than
larger deflection of slender bar. Hence, dimensional errors will to variations of the cutting speed.
increase. The decrease of cutting force can result in the
IV. CONCLUSIONS
decrease of dimensional error in the same way. Within the
An artificial neural network model was developed to
describe the relationship between machining parameters and
50
dimensional errors in slender bar turning operations. For
including the effect of part stiffness on dimensional errors, the
40
diameter and the ratio of length to diameter of the machined
Dimensional errors ( μm)

parts were considered as independent input parameters of the

30 model. The model was trained by Levenberg–Marquardt
training algorithm with Bayesian regularization. Experiments
20 were performed to verify the developed model. Further, the
model was used to investigate the influence of cutting
10 f =0.2 mm/rev parameters on dimensional errors. From the foregoing
f =0.15 mm/rev analyses, the following conclusions can be drawn within the
f =0.1 mm/rev scope of cutting conditions investigated in this study:
0
30 35 40 45 50 55 60 1. The developed neural network model is able to model
Cutting speed (m/min)
accurately dimensional errors in slender bar turning within
Fig.3 Relation between cutting speed and dimensional errors
the range in which it had been trained.
(ap=0.24mm)
2. Increased cutting speed helps in improving the dimensional

6055

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on February 09,2023 at 04:56:18 UTC from IEEE Xplore. Restrictions apply.
 errors, whereas increased feed or depth of cut deteriorate [7] R. Azouzi and M. Guillot, “On-line prediction of surface finish and
dimensional deviation in turning using neural network based sensor
the dimensional errors.
fusion,” International Journal of Machine Tools & Manufacture, vol. 37,
3. The cutting parameters mostly affecting dimensional errors no.9, pp.1201-1217, 1997.
are the depth of cut and feed in slender bar turning [8] T. S. Suneel and S. S. Pande, “Intelligent tool path correction for
operations. The effect of the cutting speed is slight. improving profile accuracy in CNC turning,” International Journal of
Production Research, vol. 38, no. 14, pp. 3181-3202, 2000.
REFERENCES [9] K. A. Risbood, U. S. Dixit, and A. D. Sahasrabudhe, “Prediction of
surface roughness and dimensional deviation by measuring cutting forces
[1] L. Kops, M. Gould, and M. Mizrach, “Improved analysis of the accuracy and vibrations in turning process,” Journal of Materials Processing
in turning, based on emerging diameter,” Journal of Engineering for Technology, vol. 132, pp. 203-214, 2003.
Industry, vol. 115, pp. 253-257, 1993. [10] László Monostori, “AI and machine learning techniques for managing
[2] J.R.R. Mayer, A.-V. Phan, and G. Cloutier, “Prediction of diameter errors complexity, changes and uncertainties in manufacturing,” Engineering
in bar turning: a computationally effective model,” Applied Application of Artificial Intelligence, vol. 16, pp. 277-291, 2003.
Mathematical Modeling, vol. 24, pp. 943-956, 2000. [11] H. C. Zhang and S. H. Huang, “Artificial neural networks in
[3] A. –V. Phan, L. Baron, J. R. R. Mayer, and G. Cloutier, “Finite element manufacturing-a state of the art survey,” International Journal of
and experimental studies of diametral errors in cantilever bar turning,” Production Research, vol. 33, no. 3, pp. 705-728, 1995.
Applied Mathematical Modelling, vol. 27, pp. 221-232, 2003. [12] E. O. Ezugwu, D. A. Fadare, and J. Bonney, “Modeling the correlation
[4] L.Z. Qiang, “Finite difference calculation of the deformation of between cutting and process parameters in high-speed machining of
multi-diameter workpiece during turning,” Journal of Material Inconel 718 alloy using an artificial neural network,” International
Processing Technology, vol. 98, pp. 310-316, 2000. Journal of Machine Tools & Manufacture, vol. 45, pp. 1375-1385, 2005.
[5] L.Carrino, G. Giorleo, W. Polini, and U. Prisco, “Dimensional errors in [13] T. S. Suneel, S. S. Pande, and P. P. Date, “A technical note on integrated
longitudinal turning based on the unified generalized mechanics of product quality model using artificial neural networks,” Journal of
cutting approach. Part 1: Three-dimensional theory,” International Materials Processing Technology, vol. 121, pp. 77-86, 2002.
Journal of Machine Tools & Manufacture, vol. 42, pp. 1509-1515, 2002. [14] C. Chungchoo and D. Saini, “On-line tool wear estimation in CNC
[6] L.Carrino, G. Giorleo, W. Polini, and U. Prisco, “Dimensional errors in turning operations using fuzzy neural network model,” International
longitudinal turning based on the unified generalized mechanics of Journal of Machine Tools & Manufacture, vol. 42, pp. 29-40, 2002.
cutting approach. Part 2: Machining process analysis and dimensional [15] T. Özel and Y. Karpat, “Predictive modeling of surface roughness and
error estimate,” International Journal of Machine Tools & Manufacture, tool wear in hard turning using regression and neural networks,”
vol. 42, pp. 1517-1525, 2002. International Journal of Machine Tools & Manufacture, vol. 45, pp.
467-479, 2005.

6056

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on February 09,2023 at 04:56:18 UTC from IEEE Xplore. Restrictions apply.