Artificial Neural Network Modelling

Advances in Science and Technology
Research Journal
Volume 11, Issue 3, September 2017, pages 261–269 Research Article
DOI: 10.12913/22998624/76546
ARTIFICIAL NEURAL NETWORK MODELLING OF VIBRATION

IN THE MILLING OF AZ91D ALLOY
Ireneusz Zagórski1, Monika Kulisz2, Aleksandra Semeniuk3, Anna Malec3
1
Department of Production Engineering, Faculty of Mechanical Engineering, Lublin University of Technology,
Nadbystrzycka 36, 20-618 Lublin, Poland, e-mail: i.zagorski@pollub.pl
2
Department of Organisation of Enterprises, Faculty of Management, Lublin University of Technology,
Nadbystrzycka 38, 20-618 Lublin, Poland, e-mail: m.kulisz@pollub.pl
3
Student of Production Engineering, Faculty of Mechanical Engineering, Lublin University of Technology,
Nadbystrzycka 36, 20-618 Lublin, Poland, e-mail: aleksandra.semeniuk@pollub.edu.pl, anna.malec@pollub.
edu.pl
Received: 2017.05.15 ABSTRACT

Accepted: 2017.08.01 The paper reports the results of artificial neural network modelling of vibration in a
Published: 2017.09.03 milling process of magnesium alloy AZ91D by a TiAlN-coated carbide tool. Vibra-
tions in machining processes are regarded as an additional, absolute machinability
index. The modelling was performed using the so-called “black box” model. The best
fit was determined for the input and output data obtained from the machining process.
The simulations were performed by the Statistica software using two types of neural
networks: RBF (Radial Basis Function) and MLP (Multi-Layered Perceptron).
Keywords: magnesium alloys, high-speed dry milling, vibration, chatter in milling,

simulation, artificial neural networks.
INTRODUCTION enabling the users to apply parameters which en-

sure machining stability. In terms of process sta-
The knowledge of phenomena in machining bility, the most important parameters include the
processes is of vital importance. One particularly cutting depth ap and the rotational speed n. Given
undesired phenomenon is vibration, particularly the advancement of cutting tools, machine tools
self-excited vibration known as chatter which oc- and materials, special attention should be paid
curs between the tool and the workpiece due to to the control and modelling of the cutting pro-
temporary surface imbalance. Vibration results in cess. New methods are sought that could ensure
a lower quality of the machined surface, dynamic reduced vibration of the cutting tools and work-
tool wear and even damage of machine tool com- piece [21, 25]. Vibration is a typical by-product
ponents. The occurrence of vibration in the mill- of technological operations and treatment, and is
ing process can also lead to decreased machining difficult to eliminate completely. However, there
efficiency [1, 25]. What is more, besides exerting are many techniques for preventing self-excited
a negative effect on the human body, excessive vibration. Such techniques include numerical
vibration can significantly contribute to the desta- simulations and process dynamics modelling, the
bilization of the production process [16]. use of actuators, changing rotational speed fre-
Taking the above into consideration, the key quency or increasing feed per tooth rate [15]. Ac-
problem is to determine such conditions of the cording to Quintan and Ciuran [14], the methods
milling process which will prevent vibration and for preventing vibration can be divided into two
thus ensure stability of this process. As a result, main groups. The first group includes techniques
different types of specialist software are applied, that ensure machining stability via the selection
261
Advances in Science and Technology Research Journal Vol. 11 (3), 2017
of parameters from the stability lobe diagram “wave” notched in the previous tool pass (inter-
and those that utilize the so-called lobbing effect. nal modulation) becomes an external modulation,
This group includes both out-of-process methods leading to an additional variation in the machined
(methods focusing on the prediction, estimation layer thickness. The above external modulation is
and identification of SLDs by process modelling; a source of external excitation (variations in the
analytical and experimental methods; machin- cutting force), to which the machine tool-chuck-
ing processes modelling) and in-process methods workpiece-tool system responds with vibration of
(process and signal monitoring via the use of e.g. the same frequency, phase-delayed [3]. To ensure
sensors). the stability of the milling process and high-speed
The other group of vibration-preventing tech- machining, the feed rate should always corre-
niques includes all methods that involve changes spond to the total numerical ratio of chatter and
in the behaviour of a system and modification of tool rigidity [4]. In this way, one can minimize the
the stability limit. Here, one can distinguish ac- regenerative effect and thus ensure the stability of
tive methods that focus on active modification the machining process.
of system behaviour, e.g., by changing the feed One can distinguish many factors affecting
rate or rotational speed of the spindle, and pas- the stability of a milling process. The most impor-
sive methods which require the implementation tant include: machining strategy, type of machine
of changes in the design of the cutting tools or tool, chucks, tool and its mounting, workpiece,
the use of additional devices mounted on the ma- cooling and lubrication of the tool-workpiece
chine components. The work [12] overviews a contact zone, and technological parameters. The
wide spectrum of techniques for damping chatter study [25] investigated the stability of a 5-axis
vibration; it also describes the way of selecting milling process for a rotor made of aluminium al-
a given vibration damping method depending on loy AlMgSiO5. The tool rigidity and the rotor ele-
various aspects of the machining process such as ment were examined to determine the frequency
material workability and stiffness of the machine of free vibration. Using ShopPro, a stability lobe
tool-chuck-workpiece-tool system. The work also diagram for the analysed process was produced
describes how to design a stable process using and experimental tests were conducted of the
stability lobe diagrams. The diagram was made milling process for a rotor combined with simul-
using a stability model based on the cutting force taneous measurement of vibration. The results
components, dynamic and technological param- demonstrate that under unstable conditions the
eters of the process, and tool geometry. amplitude of vibration increases even by 5 times,
which affects the surface quality after machin-
THE STATE OF KNOWLEDGE ing. Stability analyses are often performed for
thin-walled parts made of aluminium alloys such
Vibrations occurring in the machine tool- as EN AW-7075 T6 [17, 22], EN AW-2024-T351
chuck-workpiece-tool system can be divided into [2] and EN AW-6061-T6 [23, 24]. The works
two groups: vibrations that are not related to the [17, 18] investigate acceleration and cutting
machining process and vibrations that are caused force components to determine the stability limit
by the machining process. Taking into account the for parts with a varying (decreasing) thickness
source of vibration, vibrations can be divided into of the wall. Recurrence plots were made and re-
free, forced and self-excited. The main causes of currence quantification analysis (RQA) was per-
self-excited vibration in the machining process formed. Variations in the modal parameters of
include: envelope feedback between the param- the cutting process for thin-walled parts are sig-
eters of working movements and the mass-dis- nificant for predicting stability diagrams. Modal
sipation-elastic machine tool-chuck-workpiece- weight, rigidity and damping ratio are vital, too.
tool system, feedback by trace machining (trace Stability indices such as recurrence rate (RR),
regeneration or regenerative effect) as well as the determinism (DET), laminarity (LAM), diver-
friction between the workpiece and the tool and gence (DIV), length of the longest diagonal Line
chip [18]. To account for the phenomenon of vi- (Lmax), diagonal line segment distribution (Lent)
bration reproduction, one can resort to the exter- can be used to determine limit stability. The pa-
nal and internal modulation of a machined layer per [18] employs the classical modal analysis
thickness. After one revolution of the workpiece and recurrence plots to investigate, using the
or an incomplete revolution of the end mill, the CutPro software, the stability of a milling pro-
262
cess for aluminium alloy EN AW-7075 by a PCD using the PD controller, the level of vibration was
end mill (GuhringWNR-05492-16) with 16 mm significantly reduced.
in diameter. RP is a method for locating recurrent The stability of a milling process can also be
patterns and structural changes in a system. This investigated with the following methods [22]:
technique ensures qualitative interpretation of chase plane method, Poincare’s method and spec-
hidden patterns in the dynamic systems based on tra analysis. This paper discusses the relationship
the reconstruction of phase space. The following between the maximum Lyapunov exponent and
were used: RQA, stability indices: DET, LAM, the spindle speed and milling depth. Finally, tak-
DIV, LENT, VENT, and SLD plots. The study dem- ing the maximum Lyapunov exponent as a cri-
onstrated that, compared to the classical modal terion, the study determines the chatter stability
analysis, the RP method yields better results that domain of milling by the contour method. The
can be shown in a simplified form as stability in- maximum Lyapunov exponent increases with
dices. In addition, it is suggested that the classi- increasing the milling depth. The maximum Ly-
cal modal analysis is not sufficiently effective in apunov exponent of vibration signal equal 0.61
assessing stable regions. is taken as a milling chatter nonlinear criterion.
The analysis was performed using a nonlin- In addition, the maximum Lyapunov exponent of
ear model of milling described by two degrees the workpiece vibration signal processing can be
of freedom, taking account of the nonlinear phe- a means of optimizing the processing parameters.
nomena and susceptibility of the tool and work- The stability analysis was also performed for
piece moving in the same direction. To select the thin-walled curvilinear elements made of mate-
technological parameters ensuring stable machin- rials such as EN AW-6061-T6.The modelling of
ing for the depth of cut ap=0.5 mm and the ro- vibration in the cutting zone can also be done
tational speed n= 3000 revolutions per minute, by FEM simulation techniques. A new dynamic
stability lobe diagrams were made. After that, the model of tool and workpiece system is proposed
effect of the external force was investigated for to consider the dynamic behaviour of tool and
the above parameters. The results of previous re- workpiece as well as the influences of engage-
search demonstrate that such parameters are typi- ment and tool feed direction. The numerical and
cal of unstable machining. The closed loop con- experimental results demonstrate that chatter can
trol involved the use of a proportionate-integral be accurately predicted for the peripheral milling
PD controller. Due to the use of external force, of a thin-walled workpiece [23]. The milling pro-
suitable values of the technological parameters cess of EN AW-6061-T6 alloy was performed by
(based on stability diagrams) and the PD control- a cutter with a varying angle of helix inclination.
ler, the vibration clearly decreased. Vibration (displacement) was analysed, the Poin-
The paper [20] describes the use of an ex- caré section was designed, and Fast Fourier trans-
ternal force to control the milling process in an form (FFT) analysis was performed in frequency
open and closed loop. The numerical analysis spectrum. Another parameter that may be help-
was performed in Matlab-Simulink by the fourth- ful in stability analysis is the so-called root mean
order Runge-Kutta method with variable step in- square (RMS) of displacement. The irregular tool
tegration. A nonlinear model with two degrees of geometry (varying angle of inclination of the he-
freedom was employed, taking into account the lix) may affect the modulation of regenerative ef-
phenomena occurring during the process as well fect and may have a positive impact on vibration
as the susceptibility to nonlinearity of the tool damping [24]. The application of FEM to simu-
and the workpiece rotated in the same direction. late the milling of EN AW-2024-T351 alloy can
In order to select process parameters that would help determine the effect of tool vibration on chip
ensure stable machining for the cutting depth morphology, cutting force, and surface topology.
ap of 0.5 mm and the rotational speed n of 3000 A hybrid dynamic cutting model (HDC) is pro-
revolutions per minute, stability diagrams were posed to predict chip morphology under dynamic
determined and the impact of the external force cutting conditions. The numerical results show
was investigated. The results demonstrate that the that higher values of cutting speed and uncut
parameters match an unstable machining process. chip thickness are associated with higher vibra-
With the closed-loop control a proportional inte- tion amplitudes [2]. The limit stability of a mill-
grating controller PD is used. By applying the ex- ing process can also be determined by methods
ternal force, proper technological parameters and containing so-called constant training algorithms.
263
The paper [8] offers a new approach to assessing by serrated cutters, the machining efficiency can
process stability by measuring acceleration sig- be higher (by increasing the depth of cut) due to
nals. A multidimensional stability lobe diagram a smaller cutting torque when compared to mill
(MSDL) is used with two new algorithms of cutters with a straight cutting edge. The effect of
constant training. A new criterion in introduced tool geometry on stability of the milling process
which provides information about the quality of is also discussed in [5]. The experimental find-
prediction for a given input region. The study was ings and numerical results served for devising an
performed using a support vector machine and ar- optimal design of varying geometry milling tools.
tificial neural networks. Pocket components are The simulations were performed for a four-blade
often manufactured for the aircraft industry. The milling cutter with 12mm in diameter and made
stability of the milling process for such elements, of aluminium alloy EN AW-7075. The results
often made of EN AW-7075 T6, has been inves- demonstrate that the use of optimal helix inclina-
tigated using a genetic search algorithm. A wide tion angle leads to higher stability of the machin-
range of spindle speeds, feeds, depth and width ing process.
of cuts are processed through a genetic search It is important that machining can be per-
algorithm. The most optimal cutting condition formed in the full range of process parameters
set, which gives the shortest machining time without decreasing the cutting speed, feed rate
while respecting the chatter, torque and power or depth of cut, while at the same time main-
constraints, is determined [9]. It is also possible taining the required machining accuracy. To this
to produce three-dimensional stability lobe dia- end, we often use active control of tool vibra-
grams which provide vibration frequencies (ex- tion. Vibration monitoring involves the control
cept the basic parameters ap and n). The solid line of physical quantities and the creation of tempo-
(in three-dimensional stability lobe diagrams) can rary values of the control signal in accordance
indicate the threshold value of stability limit [7]. with the adopted monitoring strategy. This type
An important factor affecting the stability of control entails not only passive observation of
of the milling process and production efficiency a negative phenomenon but also taking steps to
is suitable design of special milling tools using prevent this phenomenon. The supervision and
stability models. Stability analyses are often actions taken on its basis do not necessarily oc-
performed with respect to shank tools with the cur simultaneously [10].
so-called “classical geometry” (straight cutting Despite its many advantages, the machining
edge) or with a wave-shaped cutting edge at the of magnesium alloys poses several risks. Dur-
tool perimeter (e.g. Kordell design geometry) ing the process there may occur undesired phe-
[11]. Using the general dynamic model of the nomena such as built-up edge (on the tool flank
cutting force [6], numerical simulations were or rake face), resulting from intensive adhesion.
performed to investigate the stability of a milling These phenomena can strongly affect variations
process for aluminium alloy EN AW-7075-T6 us- in the cutting force components, which may, in
ing carbide mills with a wave-shaped (serrated) turn, lead to decreased surface quality, reduced
cutting edge and with a regular (straight) cutting accuracy of size and shape, increased vibration
edge. The diameter of the serrated tool was 20 and temperature during the machining process-
mm and that of the regular tool was 16mm. the es. There is also a risk of self-ignition of mag-
modelling was performed by the semi-discreti- nesium, which can occur when the temperature
zation method which consists in transforming suddenly increases in the cutting zone. Finally,
the models and equations of continuous func- magnesium dust produced during the machining
tions into their discrete counterparts. In addition, has a negative effect on the health of the ma-
a harmonic analysis in the Frequency Response chine tool operating staff and the state of the ma-
Function (FRF) was performed. Based on these chine tools alike [11, 13].
results, it was found that the regular cutters, in Therefore, the analysis of vibration and com-
contrast to the serrated ones, generate periodi- puter simulations can lead to enhanced stability,
cally varying and uniform (repeatable in terms of security and efficiency of magnesium alloy ma-
value) cutting forces. chining. The model enables choosing techno-
The above may be caused by an uneven dis- logical parameters in such a way as to obtain the
tribution of loads between the grooves of the ser- lowest vibration, without causing any machining
rated tool. During the removal of material layer errors. During the milling process there may oc-
264
cur temporary surface irregularities caused by vi-

brations between the tool and the workpiece. That
is why it is important to define the conditions that
will ensure a stable course of the treatment and
help prevent vibration, especially self-excited vi-
bration. Vibration modelling can be helpful in the
design of a engineering decision support model in Fig. 1. General plan of the study – model
terms of technological parameters selection. The of the milling process
use of artificial neural networks allows us to pre-
determine the values of individual technologi-
cal parameters and predicted values of vibration The technical specifications of the accelerometer
components. The numerical results provide posi- were as follows: measuring range ± 18g, dynamic
tive evidence for the possible use of artificial neu- range - from 85 dB to 10 Hz, resolution – from 1
ral networks in the design of nonlinear processes. mg to 10 Hz.
The variable parameters included: cutting
METHODOLOGY, AIM AND OBJECT OF speed vc, feed per tooth fz, cutting depth ap, cutting
THE STUDY width ae, tool blade geometry. It can be assumed
that the milling process is a kind of controllable
The modelling of vibration was performed for object, wherein output parameters are the follow-
a milling process of magnesium alloy AZ91D. The ing vibration components: ax component along
process was performed using a TiAlN-coated car- the X axis and ay component along the Y axis,
bide milling cutter with a smooth cylindrical grip and az vibration component along the Z axis.
used for the machining of Al and Mg alloys. The The input variables included: cutting speed vc,
cutter had the following dimensions: 16x25x100 feed per tooth fz and cutting depth ap, the remain-
mm W-Z2, λs=30°. The parameters were applied ing parameters were maintained constant during
in the following ranges: vc=400÷1200 m/min, machining. A schematic model of this process is
fz=0,05÷0,3 mm/tooth, depth of cut ap=0,5÷6 mm. shown in Figure 1. The design of this process con-
the width of milling was maintained constant at sists in the use of a so-called “black box” model
ae=14 mm. The cutting process was performed on to determine the best fit between input and output
the vertical machining centre AVIA VMC800HS data obtained from the process. This solution was
provided with the Heidenhain iTNC 530 control applied due to the complicated nature of the pro-
system, with the maximum rotational speed of the cess. Such a model is used whenever it is difficult
spindle n set to 24 000 revolutions per minute and to determine mathematical equations describing a
the maximum feed rate of 40 m/min. given process [11].
The tools were mounted in the HSK63A SFD Assuming that the machining process for a
16x120 holder from SECO. Before mounting the given part is repeatable, it can be assumed that
holders in the spindle cone, the holders were bal- forced acceleration and cutting conditions at se-
anced using the CIMAT CMT-15V2N balancer, lected points of the tool path are repeatable, too.
for the class G 2.5 at 25 000 revolutions per min- The accuracy of such an assumption requires that
ute according to ISO 1940: 2003. This standard the cutting is done with a sharp instrument, at
allows for residual unbalance of as low as 1 g mm constant (for subsequent machining cycles) tech-
depending on the tool weight. For the TiAlN-coat- nological parameters of the cutting process.
ed carbide tool mounted in the thermo-shrinkable
holder, the residual unbalance was 1 g mm. TEST RESULTS AND ANALYSIS – NEURAL
Vibration was measured using the SETAC NETWORK MODELLING
(Triaxial Acceleration Computer) accelerometer
manufactured by SEQUOIA, based on MEMS The aim of modelling is to predict nonlinear
(Micro-Electro-Mechanical Systems). Installed processes (in this case, a milling process) using
on the machine tool, this device creates an au- neural networks. These analyses can be used to
tonomous system for triaxial monitoring of vibra- create a system that would aid decision-making
tion during machining processes such as milling. processes in a business company. For example,
The accelerometer was mounted on the vice in via selection of appropriate machining param-
which the magnesium alloy specimen was fixed. eters, it is possible to ensure stable machining
265
was performed using the following activation

functions: linear, exponential, logistic, tanh
and sine. The training of the RBF network was
performed by the RBFT method. The activa-
tion function of the RGB networks for hidden
neurons had Gaussian distribution and for out-
put neurons – linear function.
The modelling of vibration components mill-
Fig. 2. Schematic design of an artificial neural net- ing AZ91D alloy using artificial neural networks
work for vibration component along the X – ax axis performed using Statistica Neural Networks. The
study analysed two networks: RBF (Radial Basis
Function) and MLP (Multi-Layered Perceptron).
and to eliminate vibration between the tool and Each vibration component was modelled sepa-
the workpiece. This can result in, among others, rately. These values were calculated as an average
enhanced efficiency of the machining process, a of the maximum values of 10 separate compart-
higher quality of the machined surface and lower ments a stable machining area, representing an
tool wear. output value for each model. The MLP network
The modelling of vibration components in training was done by the Broyden-Fletcher-Gold-
the milling of AZ91D alloy using artificial neu- farb-Shanno (BFGS) method. In the simulation,
ral networks was performed by the Statistica MLP uses the following activation functions:
Neural Networks software. Two network types linear, exponential, logistic, tanh and sinus. Net-
were modelled: Radial Basis Function (RBF) work RBF taught using methods RBFT. Activa-
and Multi-Layered Perceptron (MLP). Every tion function network RBG for hidden neurons is
vibration component was modelled separately. a Gaussian distribution, and the output neurons
Their values were calculated as an average of - linear function.
the maximum values from 10 intervals isolat- In order to ensure the simplest network de-
ed from a region of stable machining, which sign, the networks were built with one hidden lay-
was considered the output value of each mod- er. The input layer consists of three neurons (the
el. The MLP network training was done by the cutting speed vc, the feed per tooth rate fz and the
Broyden-Fletcher-Goldfarb-Shanno (BFGS) depth of cut ap) while the output layer had only
method. The modelling of the MLP network one neuron (a relevant vibration component). The
Table 1. Parameters of the best MLP and RBF networks for the tested vibration components
Quality
Network Network Quality (validation) Error Activation Activation
Error (validation)
no. name (training) [%] [%] (training) (hidden) (output)
Vibration component ax
1 MLP 3-4-1 96,97% 96,90% 0,38 1,21 Sine Sine
2 MLP 3-4-1 95,53% 95,45% 0,55 0,73 Linear Linear
3 RBF 3-4-1 94,76% 91,67% 0,64 1,49 Gaussian Linear
Vibration component ay
5 MLP 3-4-1 99,70% 91,34% 0,02 1,16 Logistic Logistic
6 MLP 3-6-1 99,70% 91,32% 0,02 1,15 Logistic Logistic
Vibration component az
9 MLP 3-4-1 99,03% 99,23% 0,09 1,28 Linear Linear
10 MLP 3-4-1 99,24% 99,54% 0,08 1,61 Linear Logistic
266
Table 2. Comparison of the actual values of the vibration component ax with the predicted values for selected cases
of MLP 3-4-1
vc fz ap ax (experimental) ax (numerical) Δ δ
400 0,15 6 11,15 11,22 0,08 0,68%
600 0,15 6 10,85 11,39 0,54 5,02%
800 0,15 6 11,86 11,45 0,41 3,46%
800 0,15 4,5 11,22 11,35 0,13 1,15%
800 0,15 3 10,41 10,32 0,08 0,82%
800 0,2 6 14,45 14,01 0,45 3,10%
800 0,25 6 15,95 16,25 0,30 1,90%
800 0,3 6 16,98 16,90 0,08 0,46%
1200 0,15 6 10,74 11,32 0,58 5,41%
number of training epochs (200) and the number component ay was obtained in 78 iterations. For
of neurons in the hidden layer (3÷7) were select- the modelled components, the quality of the MLP
ed experimentally. To simulate all three vibration network is higher than that of the RBF network,
components, we built 3 models of artificial neural therefore it is recommended to use the MLP net-
networks. A schematic design of one of these network.
works for a vibration component in the direction To illustrate the accuracy of the modelled net-
X – ax is shown in Figure 2. Schematic designs of works, Table 2 gives both the actual values of the
other components are the same, each model out- vibration components and their predicted values
put produces a relevant vibration component. for selected cases of the best network (MPL 3-4-
The experiments were conducted for 16 da- 1) and the component ax along with the values of
tasets of machining parameters, out of which relative and absolute error. As regards the actual
13 were used for network training. The train- values of the vibration components, the relative
ing group was made of 80% of the measurement error does not exceed 7 %.
results while the validation group comprised of By the modelling of vibration components oc-
20% of the results. As mentioned in [19], with curring during the AZ91D alloy milling by neural
small numbers of data sets it is possible to aban- networks, it was possible to design models for
don the use of the so-called test group. calculating the values of individual components.
The most important indices of correct network Taking into consideration the afore-mentioned in-
selection are: training quality, validation quality, dices of correct network selection, below are the
training error and validation error. The training results of the modelling of vibration components
and validation errors were determined by the versus cutting speed vc and feed per tooth fz: ax
least squares method. Every simulation involved – Network 1: MLP 3-4-1 (Fig. 3), ay – Network
experimental determination of 100 networks, out 6: MLP 3-6-1 (Fig. 4) and az – Network 10: MLP
of which – based on the above indices -- 2 best 3-4-1 (Fig. 5). Based on the results, it is possible
networks per each tested network type (MLP and to determine the values of individual vibration
RBF). Their parameters for individual vibration components (ax, ay, az) with respect to given val-
components (ax, ay and az) describing the milling ues of cutting speed vc and feed per tooth fz. After
process for AZ91D alloy are listed in Table 1. The importing the values of vc, fz, ap into Statistica, the
results of the analysis of the neural network mod- programme returns the value of a relevant vibra-
els demonstrate that among the tested vibration tion component..
components the best results were obtained for the The designed models of neural networks are a
MLP network (ax – Network 1: MLP 3-4-1 with tool enabling the determination of the maximum
four neurons, ay – Network 6: MLP 3-6-1 with six values of vibration components. The numerical
neurons, az – Network 10: MLP 3-4-1 with four results may be used to determine the technologi-
neurons). The MLP 3-4-1 network for the compo- cal parameters of machining processes for ma-
nent ax was obtained in 26 iterations, for az – in 15 chine parts, especially in the case of non-linear
iterations, while the network MLP 3-6-1 for the processes. The simulation of the cutting force
267
Fig. 3. Vibration component ax versus cutting speed vc Fig. 5. Vibration component az versus cutting speed vc
and feed per tooth fz – MLP 3-4-1 and feed per tooth fz – MLP 3-4-1
tion of suitable technological parameters leads

to the optimization of machining conditions,
reduced vibration in the machine tool-chuck-
workpiece-tool system and more efficient ma-
chining.
3. The simulation of vibration components en-
ables the optimization of machining condi-
tions without the necessity of conducting
technological tests (time and material saving);
they can also prevent tool or workpiece dam-
age (caused by size and shape inaccuracies or
high surface roughness) due to preliminary
selection of machining parameters for a pre-
production batch.
Fig. 4. Vibration component ay versus cutting speed
vc and feed per tooth fz – MLP 3-6-1 4. Vibration can be regarded as an interference
with the process, so simulations are a kind of
„regulator” of the process; interference com-
pensation can be omitted in process control
components can be used to create a decision sup-
owing to the prediction of output values.
port computer system. In this study, the model
5. The simulations do not eradicate the root cause
variables are: cutting speed vc, feed per tooth fz,
of the problem because vibration is caused not
cutting depth ap. By selecting suitable machining
only by technological parameters of the pro-
parameters, it is possible to ensure a stable ma-
cess; they can, however, eliminate the cause
chining range, which leads to enhanced efficiency
of vibration occurring due to incorrect techno-
of the machining process, higher surface quality
logical parameters.
and decreased tool wear.
6. In high-speed machining it is vital to consider
CONCLUSIONS dynamic phenomena such as vibration, as this
ensures a shorter machining time and process
The results of the conducted analytical and stability at the same time.
experimental research lead to the following gen-
eral conclusions: REFERENCES
1. Artificial neural networks are advanced tech-
1. Adamski W.: Manufacturing development strate-
niques for the modelling of machining dynam- gies in aviation industry. Advances In Manufactur-
ics. They can be used for, i.a., predicting non- ing Science and Technology, 2010, 34(3), 73–84.
linear machining processes such as milling. 2. Asad M., Mabrouki T., Rigal J-F.: Finite-element-
2. Given that vibration has a significant effect based hybrid dynamic cutting model for alumini-
on the machining process stability, the selec- um alloy milling. Proceedings of The Institution of
268
Mechanical Engineers - Part B - Journal Of Engi- 14. Quintana G., Ciurana J.: Chatter in machining pro-
neering Manufacture, 2010, 224, B1, 1-13. cesses: A review. International Journal of Machine
3. Bąk P.A., Jemielniak K.: Symulacja numeryczna Tools &Manufacture, 2011, 51, 363–376.
drgań samowzbudnych przy frezowaniu ze zmien- 15. Rusinek R., Kęcik K., Warmiński J., Weremczuk
nymi w czasie parametrami skrawania. Mechanik, A.: Dynamic model of cutting process with modu-
2015, 12, 10-13. lated spindle speed. AIP Conference Proceedings,
4. Burek J., Płodzień M.: Wysoko wydajna obrób- 2012, 1493(1), 805-809.
ka części ze stopów aluminium o złożonych 16. Rusinek R., Weremczuk A., Warmiński J.: Wpływ
kształtach. Mechanik, 2012, 7, 542-549. nieliniowości w układzie skrawania na skuteczność
5. Comak A., Budak E.: Modeling dynamics and redukcji drgań samowzbudnych. Mechanik, 2014,
stability of variable pitch and helix milling tools 8-9, 81-88.
for development of a design method to maxi- 17. Rusinek R., Zaleski K.: Dynamics of thin-walled
mize chatter stability. Precision Engineering, element milling expressed by recurrence analysis.
2017, 47, 459-468. Meccanica, 2016, 51, 1275–1286.
6. Dombovari Z., Altintas Y., Stepan G.: The effect 18. Rusinek R.: Stability criterion for aluminium al-
of serration on mechanics and stability of milling loy milling expressed by recurrence plot measures.
cutters. International Journal of Machine Tools & Proceedings of the Institution of Mechanical Engi-
Manufacture, 2010, 50, 511–520. neers Part B-Journal of Engineering Manufacture,
7. Eynian M.: Vibration frequencies instable and un- 2012, 226, A12, 1976-1985.
stable milling. International Journal of Machine 19. Szaleniec, M.: Sieci Neuronowe i regresja wielora-
Tools & Manufacture, 2015, 90, 44–49. ka - czyli jak okiełznać złożoność w badaniach nau-
8. Friedrich J., Hinze Ch., Renner A., Verl A., Lechler kowych. Retrieved from (2016, October 11) http://
A.: Estimation of stability lobe diagrams in mill- www.statsoft.pl/Portals/0/Downloads/Sieci%20
ing with continuous learning algorithms. Robotics neuronowe.pdf
and Computer-Integrated Manufacturing, 2017, 20. Weremczuk A., Rudzik M., Rusinek R., Warmiński
43, 124–134. J.: The concept of active elimination of vibrations
9. Heo E.Y., Merdol D., Altintas Y.: High speed pock- in milling process. Procedia CIRP, 2015, 31, 82–87.
eting strategy. CIRP Journal of Manufacturing Sci- 21. Weremczuk A., Rusinek R., Warmiński J.: Kon-
ence and Technology, 2010, 3, 1–7. cepcja aktywnej eliminacji drgań w procesie fr-
10. Kasprowiak M., Parus A.: Weryfikacja ezowania. Mechanik, 2014, 7-8, 99-105.
skuteczności działania aktywnego narzędzia do 22. Wu S., Li R., Liu X., Yang L., Zhu M.: Experimental
obniżenia poziomu drgań w trakcie obróbki podat- study of thin wall milling chatter stability nonlinear
nym narzędziem. Mechanik,2014, 8–9, 39-46. criterion. Procedia CIRP, 2016, 56, 422 – 427.
11. Kulisz M., Zagórski I., Semeniuk A.: Artificial 23. Yang Y., Zhang W-H., Ma Y-Ch., Wan M.: Chatter
neural network modelling of cutting force compo- prediction for the peripheral milling of thin-walled
nents during AZ91HP alloy milling. Applied Com- workpieces with curved surfaces. International
puter Science, 12, 4, 49–58. Journal of Machine Tools &Manufacture, 2016,
12. Munoa J., Beudaert X., Dombovari Z., Altintas 109, 36–48.
Y., Budak E., Brecher C., Stepan G.: Chatter sup- 24. Yusoff A. R.: Identifying bifurcation behavior dur-
pression techniques in metal cutting. CIRP Jour- ing machining process for an irregular milling tool
nal of Manufacturing Science and Technology, geometry. Measurement,2016, 93, 57–66.
2016, 65, 785-808. 25. Żyłka Ł., Płodzień M.: Stabilność 5-osiowego
13. Oczoś K.E.: Rozszerzenie granic stosowalności frezowania stopów aluminium. Mechanik, 2013,
stopów magnezu. Mechanik, 2009, 5-6, 386–400. 2, 1-12.
269

Artificial Neural Network Modelling

Uploaded by

Artificial Neural Network Modelling

Uploaded by

Advances in Science and Technology

ARTIFICIAL NEURAL NETWORK MODELLING OF VIBRATION

Ireneusz Zagórski1, Monika Kulisz2, Aleksandra Semeniuk3, Anna Malec3

Received: 2017.05.15 ABSTRACT

Keywords: magnesium alloys, high-speed dry milling, vibration, chatter in milling,

INTRODUCTION enabling the users to apply parameters which en-

cur temporary surface irregularities caused by vi-

was performed using the following activation

tion of suitable technological parameters leads

You might also like