Available online at [Link].
com
ScienceDirect
Procedia CIRP 12 (2013) 49 – 54
8th CIRP Conference on Intelligent Computation in Manufacturing Engineering
Neural models for predicting hole diameters in drilling processes
F.C. Netoa, T.M. Gerônimoa, C.E.D. Cruza, P.R. Aguiara,*, E.E.C. Bianchia
a
UNESP - Universidade Estadual Paulista - Faculdade de Engenharia, Av. Eng. Luiz Edmundo C. Coube 14-01, Bauru 17033-360, Brazil
* Corresponding author. Tel.:+55-14-3234-7811 ; fax: +55-14-3103-6101 .E-mail address: aguiarpr@[Link] .
Abstract
The control of industrial manufacturing processes is of great economic importance due to the ongoing search to reduce raw
materials and labor wastage. Indirect manufacturing operations such as dimensional quality control generate indirect costs that can
be avoided or reduced through the use of control systems. The use of intelligent manufacturing systems, which is the next step in
the monitoring of manufacturing processes, has been researched through the application of artificial neural networks in the last two
decades. In this work, artificial intelligence systems were trained to estimate the diameter of holes in precision drilling processes.
The methodology involved the use of an acoustic emission sensor, a three-dimensional dynamometer, an accelerometer, and a Hall
effect sensor to monitor the drilling process. The method was applied to test specimens composed of packages of Ti6Al4V titanium
alloy and 2024-T3 aluminum alloy sheets, which are widely employed in the aerospace industry. The collected signals were
processed and the data were organized and fed into artificial intelligence systems, which consisted of an artificial multilayer
perceptron (MLP) neural network and the adaptive neuro-fuzzy inference system (ANFIS). The results indicated that the MLP
network was the most efficient of the two artificial intelligence techniques. The results also demonstrated a strong potential for the
industrial application of the models.
© 2012
© 2013 The
The Authors.
Authors. Published
Published by
byElsevier [Link]
ElsevierB.V. access under CC BY-NC-ND license.
Selection and/or peer-review under responsibility of Professor Roberto Teti
Selection and peer review under responsibility of Professor Roberto Teti
Keywords: Drilling; neural networks; ANFIS;
1. Introduction function of energy variation as a function of time, was
used to solve problems of travelling salesmen [4].
Machine drilling is one of the most common industrial Acoustic emission signals are used to feed intelligent
processes [1]. Control of the drilling process has become systems due to the simple setup of the apparatus, easy
essential, particularly in sectors such as the aerospace signal collection, and low cost of sensors, and the electric
industry, where precision drilled holes must be strictly power signal is greatly used because it provides highly
controlled. This control ensures uninterrupted series accurate data [5, 6]. Like acoustic emission, vibration or
production and high quality products while avoiding, for acceleration signal is widely accepted as a tool to
example, material wastage [2]. The use of intelligent monitor the operating conditions of a machine as it is
manufacturing systems (IMS) is the next step in nondestructive, reliable and permits continuous
monitoring manufacturing processes and has been monitoring without intervening with the process.
researched through the application of artificial neural The objective of this work was to use acoustic
networks (ANN) since the 1980s [3]. emission signals, electric power, cutting force and
Artificial neural networks are a technology derived vibration as input data to train two types of ANNs (MLP
from artificial intelligence (AI), which seeks to simulate and ANFIS) to estimate the diameter of machined holes,
the behavior and capacity of the human brain. The use of and then to compare the performance of the two tools
these intelligent systems has proved efficient in based on the analysis of the errors obtained.
nontrivial situations to be solved, for example, by a
mathematical model. A Hopfield neural network, with a
2212-8271 © 2013 The Authors. Published by Elsevier B.V. Open access under CC BY-NC-ND license.
Selection and peer review under responsibility of Professor Roberto Teti
doi:10.1016/[Link].2013.09.010
50 F.C. Neto et al. / Procedia CIRP 12 (2013) 49 – 54
2. Literature Review 2.2. Artificial Neural Networks (ANNs)
2.1. Monitoring of the drilling process According to Nandi and Banerjee [16] artificial
neural networks (NNs) are extensively used for
Metal cutting operations such as turning, milling and modeling of complex processes because of its many
drilling are used extensively in industry to produce the attractive properties, such as universal function
most varied mechanical components. Drilling, the approximation capability, resistance to noisy or missing
process of producing holes, is especially important data, accommodation of multiple non-linear variables
because it represents a large part of machining processes with unknown interaction among them and good
in general. Drilling problems may result in production generalization ability. On the other hand, fuzzy logic, a
losses, since this is one of the last processes involved in powerful tool for dealing with imprecision and
the production of a workpiece [7]. Although it has been uncertainty, is attracted to many researchers after
realizing its ability in solving real-world complex
less studied, the drilling process is responsible for the
problems. Neural networks and fuzzy logic are
removal of 25% of all material removed in industrial
considered the good alternative approaches where no
processes [2].
satisfactory analytical or empirical models exist.
Drilling forces depend on drill geometry (diameter, An adaptive-network-based fuzzy inference system
point angle, length), on the cutting conditions (feed, (ANFIS) is a fuzzy inference system whose parameters
rotation, lubrication) and on the properties of the are trained by means of neural-network algorithms. The
materials (characterized principally by hardness or system can be viewed as a particular neural-network that
coeff is functionally equivalent to a fuzzy inference system.
shear strength). Controlling these forces has to do with As long as the fuzzy system is represented as a neural-
drilling quality, which depends on the presence of network, it is straightforward to train the system by
geometric errors and dimensions, as well as burrs, and means of any of the well-known neural-network learning
on the surface integrity of the machined workpiece [8]. algorithm [17]
The main parameters that determine chip generation
3. Methodology
feed and torque [9]. Tool wear has a strong effect on the
quality of holes and precision of the diameter during 3.1. Data acquisition and drilling parameters
drilling, and is associated with high cutting forces that
can damage the workpiece, the workpiece holder or even Tests were performed on nine test specimens, whose
the machine tool itself. lower portion was composed of aluminum alloy
Monitoring machining processes using intelligent (2024T3) and upper portion of titanium alloy (Ti6Al4V).
systems and machining signals has been the subject of In each of these specimens, 162 holes were drilled
extensive research in recent decades, since the efficiency except in one test, during which the drill bit (cemented
of this process has improved over time. carbide) broke while drilling hole number 55. The
In the area of grinding, the combination of acoustic titanium specimen was drilled initially, since this is the
emission signals and cutting force, when handled model used in aircraft titanium sheets. This resulted in a
correctly, allows for real-time control of grinding wheel large databank (1143 holes effectively used) to train the
burn [10]. The root mean square (RMS) of the acoustic neural networks.
emission signal has been the main parameter studied in The machine tool used in this work was a Zocca U-30
the area of grinding, since this signal is rich in sound universal milling machine. No lubrication system was
waves, thus containing a great deal of useful information used, in order to reproduce the conditions employed in
[11]. the aerospace industry.
The application of neural networks also stands out in The electric power signals were collected with a
the control of surfaces in machining processes. Chen transducer, which measured the electric power
[12] used a neuro-fuzzy approximation in the consumed by the motor while guiding the tool holder
recognition of surfaces for milling operations, while during drilling. This transducer measured the values of
Tang [13] used polynomial networks to calculate the electric current and voltage at the terminals of the motor
approximate value of surface roughness. by means of sensors, and processed these signals in its
For turning operations, Rangwala [14] used a module, which in turn emitted a voltage signal (used as
multilayer neural network and acoustic emission and input to the neural network) proportional to the electric
cutting force signals to estimate tool wear in this power consumed by the motor. The voltage was
operation, while Purushothaman [15] used a multilayer measured using a model LV25-P Hall-effect sensor
perceptron (MLP) network to classify the wear condition produced by LEM, USA, while the electric current was
of the tool.
F.C. Neto et al. / Procedia CIRP 12 (2013) 49 – 54 51
measured with a Hall-effect sensor manufactured by
Nana Electronics Co. Ltd.
The acoustic emission signals were collected using a
Sensis DM-42 sensor attached by clamps to the test
bench; the cutting-force signals were collected with a
Kistler 9257BA piezoelectric sensor placed on the
workbench and positioned over the workpiece holder;
and the acceleration signals were collected using a PCB
482811 sensor, with full scale range at 100 times.
All the collected signals were sent to a computer
equipped with a data acquisition board (National
Instrument PCI-6035E). LabVIEW software was used to
acquire and store the data (in binary mode).
Since the machining conditions used for each test
specimen varied, Table (1) describes the cutting
parameters applied to each of these specimens.
Table 1. Cutting parameters used in the test
Fig. 1. Characteristic signals collected during the drilling process
Condition ID Rotation Feed Cutting Feed 3.3. Determination of the MLP
(rpm) Velocity Velocity (mm/rev)
(mm/min) (m/min) The determination of an ANN consists of the
1 1A 1000 90.0 15.52 0.090 optimization of its principal parameters, such as the
2 1B 1000 22.4 15.50 0.022 number of hidden layers and of neurons per layer, as
3 1C 1000 250.0 15.49 0.250 well as the learning and momentum rates, which affect
4 2A 500 90.0 7.75 0.180 the readjustment of the synaptic weights and the
5 2B 500 22.4 7.75 0.045 stabilization of the readjustment process, respectively.
6 2C 500 250.0 7.76 0.500 This optimization is performed experimentally by
7 3A 2000 90.0 31.03 0.045 attributing values to the parameters in order to find an
8 3B 2000 22.4 31.02 0.011 optimal configuration for the network, i.e., the one that
9 3C 2000 250.0 31.01 0.125 results in the lowest final mean error of the estimated
diameter. Thus, a routine was created and used to
The passes consisted of a simple drilling movement simulate approximately 7000 different networks for each
under a given condition along the workpiece, with the alloy to identify the optimal configuration. The input
signals (acoustic emission, power, vibration and force) and output data were divided randomly into two vectors
measured in real time at a rate of 2500 samples per and the vector comprising 80% of the data was used to
second. train the networks, while the vector composed of 20% of
Figure (1) shows the collected signals, where the first the data used for simulating and analyzing the results.
part of the signal (highlighted in yellow) refers to the Table (2) describes the two networks that obtained the
titanium alloy and the other part to the aluminum alloy. lowest mean error in the prediction of the hole diameter
in each type of alloy. All signals collected were
3.2. Measurement of the diameters considered as inputs to the MLP model.
In the drilling process, machined holes do not have a Table 2. Principal parameters for the construction of the MLP
perfectly circular shape but instead, always present an
oval geometry, particularly when they are analyzed in a Titanium Aluminum
high precision process. Therefore, the minimum and Layers 3 3
maximum diameters of each hole were measured using a Number of Neurons 10-5-10 10-5-10
MAHR MarCator 1087 B precision dial indicator, Learning Rate 0.3 0.15
which, when inserted into the holes, made a 360º turn to Momentum Rate 0.2 0.5
detect the aforementioned diameters. Each of these two Training Function learnlm learnlm
measures was used as a databank for the simulation of Transfer Function (between layers) tansig tansig
the neural networks. This procedure was performed for Transfer Function (output) purelin purelin
both the titanium alloy and the aluminum alloy. Number of Epochs 200 200
Error 1.00E-07 1.00E-07
52 F.C. Neto et al. / Procedia CIRP 12 (2013) 49 – 54
3.4. Determination of the ANFIS
The determination of the ANFIS was performed
analogously to that of the MLP, i.e., values were
attributed to its main parameters in order to identify the
network with the lowest mean error in the prediction of
the diameter. In addition, for the ANFIS an algorithm
that already exists in MATLAB was used, which tests
various combinations of inputs in order to select the best
group of inputs; this option does not exist for the MLP.
Table (3) describes these parameters, where PW is the
power, Fx the force in X-axis, Fz the force in Z-axis, AE
the acoustic emission.
4. Results Fig. 2. Prediction and errors for aluminum using the MLP
4.1. MLP results The prediction for titanium presented better results,
since the behavior of the measured curve for this alloy is
Based on the parameters listed in Table (2), an MLP smoother than for the aluminum alloy. Thus, a mean
network was designed to estimate the diameters of the
holes machined in the two alloys. Using the quadratic can be
means of the sensor signals as input and the minimum considered an excellent result.
and maximum measured diameters as output, the
network was then trained with 80% of the total data
(1143 holes) and simulated with the rest of the data for
the analysis and presentation of results. The analysis was
performed graphically, comparing the diameter
estimated by the network with the real diameter
measured previously in order to obtain the mean error
for each material. The prediction for aluminum is shown
in Fig. (2).
A careful observation of the curves plotted for
aluminum indicates that the prediction was successful,
since the behavior of the estimated curve follows the
measured curve. Despite some inevitable peaks, the
mean error achieved for aluminum (i.e., the average of
the mean error of the minimum and maximum diameter)
Fig. 3. Prediction and errors for titanium using the MLP
satisfactory result in view of the highly oscillating 4.2. ANFIS results
behavior of the measured curve. The curves for titanium
are depicted in Fig. (3). Like the MLP, the ANFIS was designed with the
parameters listed in Table (3) to obtain results in a
Table 3. Principal parameters for the construction of the ANFIS random sample, but with the same size as that applied to
the MLP. The prediction for aluminum is illustrated in
Titanium Aluminum Fig. (4).
Best Inputs PW-Fx-Fz AE-Fx-Fz An analysis of the curves indicates that the prediction
Influence of a cluster [0.8; 0.7; 0.1; 0] [0.8; 0.7; 0.1; 0] for aluminum by the ANFIS, like the one produced by
Initial Step 0.01 0.01 the MLP, presented excellent results, since the behavior
Relaxation Step 0.9 0.9 of the estimated diameter curve follows the profile of the
Incremental Step 1.1 1.1 measured diameter curve, despite some undesirable but
Training Method hybrid hybrid inevitable oscillations. The mean error was slightly
Number of Epochs 200 200
higher than the error obtained by the MLP, reaching
Error 1.00E-07 1.00E-07
slightly exceeded the value obtained by the MLP,
F.C. Neto et al. / Procedia CIRP 12 (2013) 49 – 54 53
considering the Brazilian ABNT NBR 6158:1995
very good result since the variation was very minor in standard.
relation to the MLP. Figure (5) presents the curves for
titanium obtained by the ANFIS. Table 4. Mean and maximum errors found for each network
MLP ANFIS
Aluminum Titanium Aluminum Titanium
Mean error
Maximum error
Based on this standard, four error ranges can be
defined to analyze the efficiency of each network:
This range is considered
100% success of the network, since this is the
maximum precision of the diameter measuring
device.
The range in which
errors are considered excellent, since they fall within
the IT7 range, which is most widely used in industry.
The range in which
Fig. 4. Prediction and errors for aluminum using the ANFIS errors exceed the IT7 limit, but are within the class of
precision mechanics.
The range that contains the
errors to be analyzed carefully, since they already lie
within the range of a rough mechanics, which is
invalid for precision drilling work.
Figure (6) illustrates the error classification results
obtained for aluminum.
For aluminum, the percentage in each error class is
not very different in the two networks, with the MLP
presenting a higher range of samples in the higher
precision classes. The ANFIS presented a sample with
should be noted since, in this tolerance range, the hole is
discarded. Figure (7), below, illustrates the results found
for titanium.
For titanium, the networks showed a behavior similar
to that presented for aluminum, i.e., there was a higher
Fig. 5. Prediction and errors for titanium using the ANFIS percentage of samples in the higher precision classes
using the MLP. However, in general, the number of
Like
i the MLP, the prediction for titanium presented correct estimates for titanium (errors lower than
was approximately two-fold higher than for aluminum.
and the maximum error was 0.0404mm (40.4 Nevertheless, no sample was classified in the range of
is slightly higher than the error found by the MLP for coarse mechanics, since the maximum error found by the
this alloy, but even so represents an excellent result.
The mean and maximum values of error for each
alloy and network are summarized in Table (4).
4.3. Comparative analysis of the two ANNs
To gain a better understanding of the potential of each
ANN, a comparative analysis was made of the results
presented by MLP and ANFIS. For this comparison, the
errors found for each sample in the simulation (a total of
229 samples in the simulation) were used, instead of the Fig. 6. Error classification for aluminium
mean error, in order to separate them into four ranges,
54 F.C. Neto et al. / Procedia CIRP 12 (2013) 49 – 54
and effective electrical power signal, CIRP Journal of
Manufacturing Systems, Paris, 31, p. 253.
[7] Strenkowski a, J.S. Hsieh a, C.C., Shih, A.J., 2004. An analytical
finite element technique for predicting thrust force and torque in
drilling, International Journal of Machine Tools & Manufacture
44, p. 1413.
[8] Drozda, T. J., Wick, C., (Ed.)., 1983. Milling. In: Tool and
manufacturing engineers handbook: machining (4 ed.)
Michigan: SME, p. 10.
[9] Batzer et al, S. A., 1998. Chip morphology and hole surface
texture in the drilling of cast Aluminum alloys, Journal r of
Materials Processing Technology, v. 79, p. 72.
Fig. 7. Error classification for titanium [10] Dotto, F. R. L., Aguiar, P.R., Bianchi, E. C., Serni, P. J.A.,
Thomazella, R., 2006. Automatic system for thermal damage
detection in manufacturing process with internet monitoring,
5. Conclusions Journal of Brazilian Society of Mechanical Science &
Engineering, Vol.28, No.2, p. 153.
[11] Spadotto, M. M., Aguiar, P. R., Souza, C. C. P., Bianchi, E. C.,
This research confirmed the efficiency of artificial Nunes, A. N., 2008. Classification of burn degrees in grinding
intelligence systems in predicting the diameter of high by neural nets The IASTED International Conference on
precision machined holes, since, based on the analysis of Artificial Intelligence and Applications, AIA 2008, p. 175, ISBN
the results, the errors were found to be sufficiently low 978-0-88986-709-3, Innsbruck, Austria, February 11-13.
[12] Chen, J. C., 2001. Neural networks and neural fuzzy approaches
to produce holes within the desired tolerances. in process surface roughness recognition system for end milling,
In the case of the MLP, it was found that the errors Iowa State University.
were, in general, lower than those obtained with the [13] Tang, W.S., Wang, J., 2001. Neural networks applications in
intelligent manufacturing an updated survev, The Chinese
ANFIS, making this the ideal tool for the prediction of University of Hong Kong.
hole diameters in both aluminum and titanium alloys. [14] Rangwala, S., Donfeld, D., 1990. Sensor integration using
The mean and maximum errors obtained with the MLP neural networks for intelligent tool condition monitoring,
g J. Eng.
Ind, Trans. ASME112 219-228.
were always slightly lower than those obtained with the [15] Purushotham, S., Srinivasa, Y. G., 1994. A back-propagation
ANFIS. In addition, a larger percentage of the errors algorithm applied to tool wear monitoring Int. J. Mach. Tools
estimated by the MLP fell within the higher precision Manuf. 34 (5), p. 625.
[16] Nandi, A. K., Banerjee, M. K., 2005. FBF-NN-based modelling
ranges, and no sample was classified in the range of of cylindrical plunge grinding process using GA. Journal of
rough mechanics. Materials Processing Technology, 162-163, p. 655.
Nevertheless, the ANFIS proved to produce very [17] Echanobe, J., Campo, I., Bosque, G., 2008. An adaptive neuro-
precise predictions, since its results did not differ fuzzy system for efficient implementations. Information Sciences,
178, p. 2150.
significantly from those presented by the MLP.
Acknowledgements
The authors are grateful to the National Council for
Scientific and Technological Development (CNPq) and
the Brazilian Federal Agency for Support and Evaluation
of Graduate Education (CAPES) for supporting this
research.
References
[1] Kamen, E. W., 1999. Industrialt Controls and Manufacturing
San Diego: Academic Press, p. 230.
[2] Konig, W., Klocke, F., 2002. Fertigungsverfahren: drehen,
frasen, bohren (7 ed.), Berlin: Springer-Verlag, p. 409.
[3] Huang, S.H., Zhang, H-C., 1994. Artificial Neural Networks in
Manufacturing: Concepts, Applications, and Perspectives, IEEE
Trans. Comp. Pack. Manuf. Tech. Part A, Vol. 17, No. 2.
[4] Sanjay, C., Jyothi, C., 2006. A study of surface roughness in
drilling using mathematical analysis and neural networks.
[5] Hundt, W., Kuster, F. & Rehsteiner, F., 1997. Model-based AE
monitoring of the grinding process, Annals of the CIRP, 45, p.
243.
[6] Aguiar, P. R., Bianchi, E. C., Oliveira, J. F. G., 2002. A method
for burning detection in grinding process using acoustic emission
