You are on page 1of 13

Mechatronics 23 (2013) 581–593

Contents lists available at SciVerse ScienceDirect

Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics

Dynamic modeling of a horizontal washing machine and optimization


of vibration characteristics using Genetic Algorithms
Pınar Boyraz ⇑, Mutlu Gündüz
Mechanical Engineering Department, Istanbul Technical University, Inonu Cd, No 65, Gumussuyu, 34437, Istanbul, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: In this work, a 2D dynamic model of a horizontal axis washing machine is derived regarding the rotation
Received 17 November 2012 plane in order to examine the vibration characteristics of the spin-cycle and improve the design propos-
Accepted 17 May 2013 ing a new optimization scheme based on Genetic Algorithms (GA). The dynamic model is numerically
Available online 17 June 2013
simulated and the outputs are validated using experimental vibration data acquired from a test-rig
including the drum and the motor of a horizontal-axis washing machine. The measurements are per-
Keywords: formed using piezo-transducers and a novel measurement scheme is used to obtain displacement values
Dynamic modeling
from acceleration data as well as estimating the instantaneous frequency of the rotation with appropriate
Horizontal washing machine
GA optimization
signal processing. This study has two main contributions: (i) a new method for design improvement
Vibration characteristics applying GA to optimization of vibration characteristics for the horizontal-axis washing machines, and
(ii) a novel measurement method yielding the displacement in 2D and instantaneous frequency of vibra-
tion from acceleration data. While the GA is contributing to passive improvement methods in the field,
the novel measurement method opens the way for low-cost diagnosis and active-vibration control of
washing machines.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction tics of the machine. Deriving a dynamic model is the first step of
the design improvement cycle and further work is required to val-
As the modern life requirements and customer expectations are idate the model and parameterize the system variables so that it
upgraded, the washing machine design is also being affected by can be recast as a well-defined optimization problem. There are
this trend. There have been a general tendency toward light- several studies on the optimization of structural parameters [6–
weight, portable and high-capacity washing machines with re- 8] offering a short design cycle because the models are parametric
duced noise and minimized vibration during operation. However, and allows the designer to test a new design option. In some stud-
these customer expectations make the washing machine more ies the passive suspension idea was replaced by an active system
prone to exhibit poor vibration characteristics, such as high ampli- using either moving, controlled balance mass(es) [9] or active-sus-
tudes of vibration as well as noise, even tendency to stepping and pension elements such as magneto-rheological dampers [10]. It is
tipping motions. Therefore, great effort have been taken to improve an expected improvement in washing machines to use active
the vibration characteristics of the washing machine while still vibration control as long as the addition of actuation and control
meeting the customer defined-market applied criteria on the oper- does not affect the manufacturing cost of these appliances.
ation of these devices. The first reported studies [1–3] focused on The efforts on the dynamic modeling and optimization are intri-
modeling and experimental assessment of suspension-group of cately connected. For example a detailed model of washing ma-
the washing machine which comprises of springs and dampers in chine using multi-body system formalism [11] can be constructed
a certain arrangement that can be decided during design cycle. to represent the functional parts of the machine in terms of optimi-
Then, in some of the more recent work [4] the dynamic model zation parameters in the model. Such a model can be constructed in
was derived to include more degree-of-freedoms and complex multi-body dynamic simulation programs such as Adams/View and
coordinate space to examine the whirling motion of the tub more powerful optimization algorithms can be run on a cluster using par-
realistically. In some of those models [5] the focus of the attention allel computing [12]. Furthermore, a multistep approach in optimi-
was details such as flexible components and the noise characteris- zation process can be employed in order to solve the multi-
objective optimization of washing machines taking into account
several cost functions including kinematic, dynamic, noise level
⇑ Corresponding author. Tel.: +90 212 2931300. and walking avoidance. This type of optimization is very beneficial
E-mail address: pboyraz@itu.edu.tr (P. Boyraz).
since it takes into account all dimensions of the optimization and

0957-4158/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.mechatronics.2013.05.006
582 P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593

Nomenclature

Variable Meaning fver vertical forces on machine cabinet


m mass of the tub fhor horizontal forces on machine cabinet
k stiffness coefficients of springs Fn normal force component caused by unbalanced mass
c viscous damping coefficient for shock absorbers Ft tangential force component caused by unbalanced mass
xs geometric place of spring on x-axis Fd damping force of shock absorbers
xd geometric place of damper on x-axis Fs force supplied by springs
zs geometric place of spring on z-axis
zd geometric place of damper on z-axis
fc friction coefficient

allows the researcher to find the right cost function definition to and rotations are restricted in the model. Washing machine cabi-
find better washing machine designs [13]. net is assumed to be rigid and fixed to the ground. No gyroscopic
In this work, a simple 2D dynamic model of the washing ma- effects are taken into consideration on drum movement and the
chine is derived in a parametric way so that it could be used in unbalanced mass is assumed to be in the middle plane of the oscil-
optimization process using a single CPU. The model is validated lation group, hence not moving along y-axis. Stiffness effects of the
using experimental vibration data collected from a horizontal-axis flexible parts such as gasket and water inlet tub are neglected. The
washing machine. As the model was seen to approach to the real spin speed of the drum is realistic and sweeps spin-speed from 0 to
vibration characteristics closely, it was decided to be used in opti- 900 rpm as in the experiments. The model is derived based on the
mization. The parameters of the washing machine used in the physical model given in Fig. 1 showing the washing machine oscil-
model are converted into string formats to be used in GA and three lation group in equilibrium state and when moved on x and z axes
different fitness functions were defined representing the transient positive directions.
and/or steady-state properties of the washing machine vibration. Equation of motion can be written as in (1) and (2) for x and z
The study has two main contributions in the field: (1) a new design axes based on the physical model given in Fig. 1.
improvement method was introduced applying GA in optimization
m€x þ F s1x  F s2x  F d1x þ F d2x  F x ext ¼ 0 ð1Þ
of vibration characteristics of the appliance and (2) a novel and
practical method for obtaining the displacement and instanta-
m€z þ F s1z þ F s2z þ F d1z þ F d2z  F z ext ¼ 0 ð2Þ
neous frequency from the measured acceleration data was applied.
The novel measurement scheme can be used as a part of active Now, each term in the equations of motion is explained in detail
vibration control making it feasible and low-cost. starting from the external force components Fx ext, Fz ext and contin-
uing with spring force terms F snx ; F snz and damper forces denoted by
F dnx ; F dnz , n being the number of the component. The external forces
2. Dynamic model derivation for vibration behavior on the washing machine oscillation group are forces caused by the
unbalanced mass rotating in the drum. This force depends on the
Dynamic model derivation is perhaps the most important part drum rotation/spin speed which follows an exponential curve given
of the work since the optimization algorithm uses the parameters in [3]. The ramp-up characteristic of the spin speed of the drum
of the model directly to improve the final design. In this section, which changes from 0 to 900 rpm given by (3) versus the real values
first the model assumptions and general equations of motion are given in experimental measurement can be observed in Fig. 2.
given, and then the modeling details of each term in the equation
are explained. b0d ¼ Nð1  e1=1;8t Þ ð3Þ
In order to obtain a model which is simple enough to compre-
Forces caused by unbalanced mass have tangential and normal
hend and use in optimization or control design, however, detailed
components given by (4) and (5) and shown in Fig. 3.
enough to represent the real dynamics of the washing machine,
some assumptions are included in the model. The oscillation group
is assumed to move on xz-plane only and any movements on y-axis F n ¼ mu rðb0d Þ2 ð4Þ

Fig. 1. Washing machine physical model showing the equilibrium and moved states.
P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593 583

Fig. 2. Spin speed ramp-up characteristics in simulation (left) and in real measurements (right).

can be performed using Eqs. (8)–(11). The same procedure for the
spring on the left of the suspension block is given in Appendix for
the sake of completeness.
xs1dyn ðtÞ ¼ xs1  xðtÞ ð8Þ

zs1dyn ðtÞ ¼ zs1  zðtÞ ð9Þ


qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ls1 ¼ xs12 þ zs12 ð10Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ls1dynamic ðtÞ ¼ xs1 dynðtÞ2 þ zs1 dynðtÞ2 ð11Þ
Fig. 3. Tangential and normal components of the force caused by unbalanced mass.
F s1 ðtÞ ¼ ½ls1  ls1dynamic ðtÞk; with the components

F t ¼ mu rb00d ð5Þ xs1 dynðtÞ zs1 dynðtÞ


F s1x ðtÞ ¼ F s1 ðtÞ ; F s1z ðtÞ ¼ F s1 ðtÞ ð12Þ
In order to see the effects of these forces on x and z-axes (4) and ls1 dynamic ðtÞ ls1 dynamic ðtÞ
(5) are used in xz-plane decomposition to give (6) and (7). Finally, the third main term in the force equation comes from the
F x ext ¼ F n cosðbd Þ  F t sinðbd Þ ð6Þ dampers/shock absorbers. The shock absorbers are modeled as lin-
ear elements with a coefficient to represent the viscous damping
F z ext ¼ F n sinðbd Þ þ F t cosðbd Þ ð7Þ and it is calculated from the force–velocity curve then taking the
average of the values found. The shock absorber on the right of sus-
The second important term in the equation of motion is the spring pension block is shown in Fig. 5 with the positive displacement vec-
force calculated by taking the dynamic length change of spring and tor, force components and the relevant equations are given in Eqs.
multiplying it with spring constant, thus the spring is assumed to be (13)–(18).
linear; however, because of the angular attachment of the springs to
xd1dyn ðtÞ ¼ xd1  xðtÞ ð13Þ
the tub the spring force term is nonlinear. The spring force is formu-
lated using Fig. 4 as it shows the dynamic length of the spring in
zd1dyn ðtÞ ¼ zd1 þ zðtÞ ð14Þ
two axes for right spring. Derivation of the nonlinear spring force

Fig. 4. Dynamic length definition, displacement vector r and nonlinear force components for springs.
584 P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593

the parameters to be used in 3D model upgrade in future work.


The constant matrix is updated at each simulation with the new
parameters except the mass. Those parameters are mainly stiffness
coefficient of springs, damping coefficients of dampers and the geo-
metric locations of suspension elements in x and z axes. All the rela-
tionships between forces are coded in the function block as an m-
file.

3. Experimental measurements, estimations and model


validation

For the validation of the dynamic system model derived, the


outputs of the real system to the inputs used in the simulation
should be measured. Vibration characteristics of a horizontal load
washing machine are measured using a low-cost set-up which is
shown in Fig. 7. The experimental set-up includes variable auto-
Fig. 5. Dynamic length definition, displacement vector r and force components for
dampers. transformers to adjust the voltage input for the electric motor in
the washing machine assembly, a simple 3-axis accelerometer
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi arrangement is used for measuring the acceleration in 3-axes and
finally a data acquisition card is employed to log the data on com-
ld1 ¼ xd12 þ zd12 ð15Þ
puter. Table 1 enlists the components used in the set-up.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi In the experimental set-up, only 3 accelerometers are used as
ld1 dynamic ðtÞ ¼ xd1 dynðtÞ2 þ zd1 dynðtÞ2 ð16Þ vibration sensors and arranged perpendicular to each other as
shown in Fig. 7. Data acquisition board can measure up to
m1 ¼ ld1 dynamic ðtÞ ð17Þ 350 Hz and the maximum frequency corresponding to maximum
spin speed is 15 Hz. Considering the Nyquist criterion F s P 2f max
for sampling rate selection, the signal is sampled at 300 Hz without
F d1 ðtÞ ¼ m1 c; with the components; F d1x ðtÞ
any aliasing. Obtaining displacement data and instantaneous fre-
xd1 dynðtÞ zd1 dynðtÞ quency from acceleration measurements required appropriate sig-
¼ F d1 ðtÞ ; F d1z ðtÞ ¼ F d1 ðtÞ ð18Þ
ld1 dynamic ðtÞ ld1dynamic ðtÞ nal processing in order to estimate them without any distortion,
phase shift and integration errors. Therefore, the signal processing
Damper forces for the left shock-absorber are given in Appendix.
algorithms will be detailed here to clarify the measurement-esti-
The full model of the suspension group under the effect of unbal-
mation methodology.
anced mass when the washing machine is ramping up the full speed
range from 0 to 900 rpm is formed in MATLAB/SIMULINK environ-
ment. The full model is given in Fig. 6 taking the external force com- 3.1. Displacement data using acceleration measurements
ponents as input and translations as the output. The constant matrix
includes the mass values and coefficients for the parameters that In obtaining the displacement data from acceleration measure-
are included in the optimization whereas variant matrix includes ments a double integration should take place theoretically.

Fig. 6. Full model for the suspension block of a washing machine gathering all optimization parameters in one matrix.
P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593 585

Fig. 7. Experimental set-up components and 3-axis accelerometer arrangement.

advantage of realizing the filtering operations with lower orders


Table 1 requiring less processing time. Although IIR is the right choice for
Technical specifications of experimental set-up. the application given here it has a drawback of introducing phase
Component Technical specifications
distortions in the signal. To eliminate the phase distortion, the time
signal is reversed and filtered again. The change in the amplitude
Accelerometer Capacitive, measuring range: ±10 g, frequency range
0–350 Hz, sensitivity: 30 mV/g
due to this double filtering is corrected as a final step. As the
DAQ card: spider8 of Sampling rate: [1–9600], 8-channels, 9600 baud numerical integration method, trapezoidal approximation is em-
HBM (serial connection) ployed. The signal processing flow includes a low-pass filter first
Autotransformer Voltage range: [0–250] V (Fig. 8) since the high-frequency noise in the signal must be elim-
inated. The high-frequency noise is caused by oversampling since
the DAQ device has a sampling rate of 300 Hz whereas the maxi-
mum frequency for the washing machine vibration is around
However, due to the noise in the signal and the effect of initial con- 15 Hz. Therefore, the low-pass filter at start is a fifth order Butter-
ditions on the integration a sequence of low-pass and high-pass fil- worth filter with a cut-off frequency of 20 Hz eliminating all the
ters are employed as in Fig. 8. higher frequency terms. The second filter is a high-pass filter with
Since the harmonics with close frequencies to the noise fre- cut-off frequency of 1.5 Hz to eliminate the DC term in the acceler-
quency have the risk of elimination, a digital filter with sharp ation caused by the capacitive nature of accelerometer. High-pass
cut-off frequency is required. Infinite Impulse Response (IIR) filters filtering does not lead to serious frequency information loss since
have sharp cut-off frequency property and they also have the the frequency range we are interested in is not around 1.5 Hz.

Fig. 8. Signal processing flow for obtaining the displacement values from acceleration.

Fig. 9. Raw acceleration data (left) and filtered acceleration data without DC term (right).
586 P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593

Fig. 10. Displacement data calculated without high-pass (left) and with high-pass (right).

Fig. 11. Phase angle of the signal in [p, +p] (left) and unwrapped phase angle (right).

Fig. 12. Calculated instantaneous frequency (left) and smoothed curve (right).

The elimination of DC term from the raw acceleration data can be Hilbert transform converts a real-valued signal into its complex
observed in Fig. 9. sequence with the same length. The resulting analytic signal is
After the integration, DC term problem arises again and an addi- used in calculation of instantaneous attributes of the signal. The
tional high-pass filter is used to correct this. The displacement data instantaneous phase angle of the input sequence is the angle of
found by integration with and without high-pass filtering is given the analytic signal, so taking its rate change yields the instanta-
in Fig. 10. neous frequency. In implementation, DC term is removed with a
high-pass filter of very low cut-off frequency as it was performed
3.2. Instantaneous frequency estimation in displacement data estimation. Then the Hilbert transform is ap-
plied to calculate the phase angle of the signal. The calculated
During the experiments, the input voltage to electric motor is phase angle of the signal and its unwrapped version is given in
supplied using a variable autotransformer to adjust the spin speed Fig. 11.
manually. However, the direct measurement of spin speed by an Next, the time derivative of the unwrapped phase angle is cal-
encoder is not always possible due to shaft accessibility problems. culated to give the instantaneous frequency. In the measurement
Therefore a practical way of indirect measurement is necessary. set-up to obtain this data, the spin speed of the washing machine
Among many frequency estimation methods [14–17], method used starts from 0 rpm ramps-up to 900 rpm and then falls down to
in [17] is selected. Hilbert transform is given in (19) in proper form. 0 rpm again. Calculated instantaneous frequency from this experi-
ment using the method here can be seen in Fig. 12. There are two
Z Z !
1 t
gðsÞ tþ1=
gðsÞ important observations about the calculated values: (i) when the
H½gðtÞ ¼ lim!0þ ds þ ds ð19Þ frequency is about 1.5 Hz or below the frequency estimates are
p t1= t  s tþ ts
not correct since these terms were removed from the signal
P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593 587

Fig. 13. Spin speed measurement by encoder and estimation by instantaneous frequency method.

Fig. 14. Dynamic system model predictions and real measurements comparison: x-axis (left), z-axis (right).

previously, and (ii) the signal has to be smoothed in the mid-region The absolute errors between the model and the measurements
as well. After the removal of the noise by Savitsy–Golay filtering are caused by the omitted parts of washing machine assembly.
[18] we obtain the second curve given in Fig. 12. Considering the simplification applied to reduce the model com-
After the smoothing and excluding the incorrectly estimated plexity and the model flexibility in terms of parametric representa-
lower frequency terms the ramp-up of spin speed can be obtained tion, the model is considered to be sufficient for the optimization
as it was given in Fig. 2. In order to validate the instantaneous fre- purposes. Although the model is validated experimentally and is
quency measurements, an experiment was run using the same adequate for the optimization purposes, it is necessary to state
ramp-up and deceleration inputs and the frequency was measured the full capability of the validated 2D model especially compared
by encoder at first run and on the second run it was estimated by to 3D dynamic models. Modeling the vibration dynamics of a
the method detailed here. The spin speeds measured by encoder washing machine using a 3D model would give the possibility to
and by the instantaneous frequency measurement here are com- examine the rotational motion around the x-axis, which might give
pared in Fig. 13. As it can be seen they are in good agreement ex- the opportunity to examine the rotational slip as in [9]. In addition
cept the low-frequency values. to this, any side effects caused by the motion of the laundry mass
along the y-axis may cause a wobbling motion and this could be
explored if a 3D model is formed. Including the y-axis in the model,
3.3. Validation of the model we could also optimize an additional design parameter such as the
angle of the tub to prevent the wobbling between the front and
In order to validate the dynamic model, two main magnitudes rear planes of the machine or the rotation around x-axis. However,
representing the steady-state behavior and transient response this type of in-depth exploration is out of the scope for this paper.
are compared between the model prediction and real measure- In the 2D simple model validated here, it is assumed that the seg-
ments. These magnitudes are (i) maximum amplitudes of transient ment along the y-axis where the laundry mass is located can be as-
vibrations and (ii) maximum amplitudes of steady-state vibrations. sumed stationary. In other words, the laundry mass does not move
One of the comparisons can be seen in Fig. 14 and the relevant along the y-axis during spinning cycle it sticks to a region in the
metrics can be seen in Table 2. inner surface of the tub. Furthermore, the critical speed for
588 P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593

Table 2 4. Optimization with Genetic Algorithms


Comparison of transient and steady-state key magnitudes between model and
measurements.
Genetic Algorithms are selected for the optimization because of
# Simulation Measurement Absolute error their attributes such as being independent from the function eval-
x Transient uation, continuity, and differentiability. In other words, GA can be
max (mm) 10.93 8.9 2.03 applied to all kinds of problems including discrete, continuous,
min (mm) 10.46 9 1.46
non-differentiable and nonlinear systems because the emphasis
z Transient of the algorithm is not on the exploitation but on the exploration
max (mm) 7.75 7.3 0.45
of the solution space. However, important part of GA implementa-
min (mm) 7.13 7.5 0.37
tion is in recasting the optimization problem in GA terminology. In
x Steadt-state (mm) ±4.45 ±5.8 1.35
order to do that the problem parameters must be refined as chro-
z Steady-state (mm) ±4.6 ±5.1 0.5
mosomes and the performance of each chromosome should be
evaluated in a fitness function defined based on the original prob-
lem parameters. In addition, any optimization constraint should be
translational speed that is examined in our paper along x-axis is encoded as an inhibition mechanism in GA. In a generic GA, the
much higher than the rotational slip condition. Thus, the 2D model application of the algorithm includes following steps: initial popu-
can account for the worst conditions although neglecting less crit- lation generation, evaluation of the members in initial population
ical slip conditions that happen before the max speed is reached. according to fitness function, selection of members based on their

Table 3
Fitness function definitions, optimization criteria and relevant system dynamics.

Fitness function Optimization criteria Relevant system dynamics

f1 ¼ p1ffiffiffiffiffiffiffiffiffiffiffi Minimize the wobbling movement of the oscillation group


maxð 2
x þz Þ2

xst
span ¼ xmax  xmin Minimize steady state vibration span
zst
span ¼ zmax  zmin
Spanst ¼ xst st
span þ zspan
1
f 2 ¼ Span
st

xst
span ¼ xmax  xmin Minimize both steady state and transient region vibration
zst span
span ¼ zmax  zmin
Spanst ¼ xst st
span þ zspan
xtr
span ¼ xmax  xmin
ztr
span ¼ zmax  zmin
Spantr ¼ xtr tr
span þ zspan
1 1
f 3 ¼ Span þ Span
st tr
P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593 589

Table 4
GA optimization runs and vibration characteristics of the given solutions for all fitness criteria.

Fitness function Fitness values/termination condition Vibration characteristics of solution

f1 ¼ p1ffiffiffiffiffiffiffiffiffiffiffi Kept running for 100 generations. The highest value was
maxð x2 þz2 Þ
obtained at 92th generation

xst
span ¼ xmax  xmin Terminated on first stage at 71th generation
zst
span ¼ zmax  zmin
Spanst ¼ xst st
span þ zspan
1
f 2 ¼ Span
st

xst
span ¼ xmax  xmin
Terminated on first stage at 14th generation
zst
span ¼ zmax  zmin
Spanst ¼ xst st
span þ zspan
xtr
span ¼ xmax  xmin
ztr
span ¼ zmax  zmin
Spantr ¼ xtr tr
span þ zspan
1 1
f 3 ¼ Span þ Span
st tr

Table 5
Comparison of simulation results and improvement by GA.

Original Fitness f1 Fitness f2 Fitness f3


Design vector k = 6155 N/m k = 6071.4 N/m k = 7142.9 N/m k = 5000 N/m
c = 308.6 Ns/m c = 400 Ns/m c = 150 Ns/m c = 364.29 Ns/m
xs = 0.075 xs = 0.0871 m xs = 0.0729 m xs = 0.0729 m
xd = 0.075 xd = 0.0871 m xd = 0.0871 xd = 0.0871 m
zs = 0.19 zs = 0.1513 m zs = 0.2126 m zs = 0.1717 m
zd = 0.205 zd = 0.1153 m zd = 0.1 m zd = 0.1459 m
x Transient (mm)
Max 10.93 6.74 6.93 6.6
Min 10.46 5.2 6.3 4.8
z Transient (mm)
Max 7.75 7.47 11.89 7.01
Min 7.13 6.3 9.45 6.2
x Steady state ±4.45 ±4.43 ±4.4 ±4.43
z Steady state ±4.6 ±4.5 ±4.5 ±4.5
590 P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593

fitness, and reproduction to generate new members to enter into conditions may occur when horizontal forces acting on the wash-
the GA cycle again. There are two separate algorithms running in ing machine exceeds the friction force between the floor and the
evaluation and reproduction parts. During evaluation, the solution washing machine. In order to avoid the solution candidates violat-
candidates are converted into their phenotypes and evaluated in ing the stepping condition, the constraint given in (21) is used,
objective functions to determine the fitness values. In reproduc- where fver and fhor are vertical and horizontal forces acting on
tion, the genotypes are used in mutation and crossover processes washing machine cabinet respectively. fc is the friction coefficient
to generate new members. and it is assumed to be 0.2 which is the coefficient of plastic foots
In minimizing the vibrations of washing machine, the problem and a smooth surface such as a bathroom surface.
is recast in terms of GA. First, the design vector to be used in opti-
ðfv er þ MgÞ  fc  fhor > 0 ð21Þ
mization is defined to comprise of stiffness and damping coeffi-
cients of elements and the geometric places at which those The members of the population violating (21) constraint are as-
elements are assembled to the tub. The suspension block is sym- sumed to be dead-members in GA and they are assigned zero fit-
metrical; therefore there are six independent parameters in design ness value to inhibit their reproduction.
vector given in (20). Here, k is stiffness, c is damping coefficient, xs Another important consideration in GA optimization is to deter-
and zs are the coordinates of the springs whereas xd and zd are mine the criteria for termination. GA does not guarantee a global
coordinates of dampers. optimization result but it can always converge to a sub-optimal
solution given the appropriate mutation and cross-over probabili-
d ¼ ðk; c; xs; zs; xd; zÞ ð20Þ
ties. In our study, a two-stage termination condition is applied:
The fitness functions are designed to represent three different First, total fitness value of generations are checked and compared
optimization criteria as given in Table 3. to each other and if it did not change significantly (i.e. less than
During GA optimization a well-known constraint in washing a) for a pre-determined number of generations (i.e. 5) then the
machine vibration called stepping condition is used. Stepping GA evaluation is terminated. However, in the probabilistic nature

Fig. 15. Sensitivity analysis including all fitness values and all variables: the % change in fitness values versus % change in the variables from the optimum.
P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593 591

of GA this condition may not occur, so there is an upper bound optimization in the loop for obtaining adaptive parameters under
determined to halt the evaluation, selected as 100 here. The solu- different conditions and employing the measurement set-up
tion is chosen as the best individual amongst the individuals of developed in this work.
the last generation if the evaluation stops at the first stage. If the
evaluation continues until the upper bound, then the individual Appendix A. Models for spring 2 and damper 2
with the max fitness value in all populations is declared as the
optimum solution. Table 4 shows the termination conditions for
each definition of fitness value criteria and the resultant vibration
characteristics with the given solution (i.e. the design vector). Spring 2 model:
In order to observe the improvement gained by GA optimization
a comparison is shown in Table 5 including the results obtained
from the simulation using original design parameters of the wash-
ing machine, and the design vectors obtained from GA for each fit-
ness criterion.
In order to observe the effects of the parameters on the
optimization results a sensitivity analysis is performed. The results
showing the variation around the optimum values for all the vari-
ables changing one variable at a time are given in Appendix. How-
ever, from the sensitivity analysis it can be briefly said that fitness
functions f2 and f3 are the most robust functions since the change xs2dyn(t) = xs2 + x(t)
in the optimum variable affects them the least. In addition to that if zs2dyn(t) = zs2 - z(t)
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the fitness functions f2 and f3 are taken as a comparison basis, the ls2 ¼ xs22 þ zs22
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
most effected variable is the z position of the damper and the dam- ls2dynamic ðtÞ ¼ xs22dynðtÞ þ zs22dynðtÞ
per coefficient itself. These effects can be seen in Fig. 15 clearly.
Fs2(t) = [ls2  ls2dynamic(t)]k
F s2x ðtÞ ¼ F s2 ðtÞ xs2 dynðtÞ
ls2dynamic ðtÞ
5. Conclusions and future work
F s2z ðtÞ ¼ F s2 ðtÞ zs2
l
dynðtÞ
s2dynamic ðtÞ

In this work, a washing machine suspension block is dynami- Damper 2 Model: xd2dyn(t) = xd2 + x(t)
cally modeled in 2D with the spin-speed as the input and the trans- zd2dyn(t) = zd2 + z(t)
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
lations in x–z plane as the outputs. The model is validated using the ld2 ¼ xd22 þ zd22
vibration characteristics measurements performed in a test-rig ld2dynamic ðtÞ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
including a real washing machine, variable autotransformer as
¼ xd2 dynðtÞ2 þ zd2 dynðtÞ2
the voltage source, the accelerometers as main sensors, encoders
as redundant sensors, a DAQ card and a PC for recording the data. m2 = ld2dynamic(t)
To obtain displacement data in x–z plane from the measured accel- Fd2(t) = v2c
eration values, a filtering sequence is employed to eliminate the F d2x ðtÞ ¼ F d2 ðtÞ lxd2 dynðtÞ
d2dynamic ðtÞ
oversampling, DC term and initial condition based problems in
F d2z ðtÞ ¼ F d2 ðtÞ zd2 dynðtÞ
ld2dynamic ðtÞ
integration. Instantaneous frequency of the rotation is also esti-
mated using Hilbert transform and validated by encoder readings.
After the dynamic model was found to be sufficient to represent
the vibration characteristics the optimization was performed using
GA. Three different fitness criteria is used in simple GA using rou-
lette wheel as selection mechanism and a user defined two-stage Appendix B. Sensitivity analysis full results
termination criterion. All fitness functions performed well, how-
ever the third fitness function yielded the most successful solution
improving both the transient and steady-state vibration character-
istics. The framework presented in this report can offer a fast
Deviation from K_spr f1 f2 f3
development/improvement method for washing machine design
optimum (%)
cycle as the previous work in the area offered. However, the contri-
bution of this work is in its flexibility to define the objectives of the 50 2500 61.6085 54.2238 18.3965
optimization without considering any linearity or differentiability 40 3000 65.5312 44.9791 17.8138
constraints. In addition, the measurement methods developed in 30 3500 65.8311 45.3975 17.8123
data acquisition phase can offer a low-cost method to obtain the 20 4000 73.2542 55.0170 17.8332
inputs for active vibration control applications in washing ma- 10 4500 76.2343 55.0276 18.8074
chines. As seen in the sensitivity analysis the damper coefficient 5 4750 77.1554 55.0714 18.5669
and its position are the most sensitive parameters affecting the Kopt (%) 5000 79.5182 55.1677 18.9945
cost or fitness functions. Therefore, it will be meaningful to design 5 5250 80.6161 55.2641 19.1193
a separate system including active-suspension control changing 10 5500 82.8090 55.5328 18.8669
the damper coefficient dynamically to suit the particular condition. 20 6000 87.5260 56.3109 18.6149
In future work, it is considered to extend the model in 3D and in- 30 6500 84.0599 50.1199 17.2157
clude the neglected parts. In addition, a new active vibration con- 40 7000 80.2365 48.9195 16.7296
trol scheme using the automotive-based semi-active suspension or 50 7500 81.1408 47.6735 16.5506
active suspension control [19] will be applied still keeping the GA
592 P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593

Sensitivity Analysis Full Resultsanalysis full results (continued)


Deviation from C_damp f1 f2 f3
optimum (%) Deviation from X_damp f1 f2 f3
optimum (%)
50 182.1450 69.6326 52.3279 15.2329
40 218.5740 71.4901 50.3224 16.1906 30 0.0610 72.3324 49.8935 17.3665
30 255.0030 73.1351 48.8887 17.1351 20 0.0697 76.0906 49.3524 17.5102
20 291.4320 75.4889 53.8359 17.8320 10 0.0784 76.3755 55.9765 18.7070
10 327.8610 77.4635 54.8938 18.1210 5 0.0827 78.0843 55.5268 18.8373
5 346.0755 78.1096 54.9781 18.2772 X_damp (%) 0.0871 79.5182 55.1677 18.9945
Copt (%) 364.29 79.5182 55.1677 18.9945 5 0.0915 80.0792 56.1022 19.1193
5 382.5045 79.7344 53.5699 18.9523 10 0.0958 80.9425 55.9034 18.8329
10 400.7190 82.2146 56.1880 19.5854 20 0.1045 85.6951 54.6803 18.9495
20 437.1480 81.3614 52.9521 19.4387 30 0.1132 88.2477 52.8166 18.9153
30 473.5770 83.1674 45.2792 18.5484 40 0.1219 88.8615 48.0593 18.3304
40 510.0060 86.0851 46.1224 18.7601 50 0.1307 92.5167 48.1945 18.4371
50 546.4350 88.0307 46.1898 19.1336

Deviation from Z_damp f1 f2 f3


Deviation from X_spr f1 f2 f3 optimum (%)
optimum (%)
50 0.0730 91.2087 53.3683 18.7540
50 0.0365 53.2141 54.6499 16.7260 40 0.0875 87.5323 48.6241 18.6074
40 0.0437 58.5445 55.1280 17.3341 30 0.1021 87.4362 47.4453 18.2248
30 0.0510 63.0254 55.2713 17.8567 20 0.1167 85.1749 50.4640 18.4956
20 0.0583 67.6372 55.2022 18.4031 10 0.1313 81.6661 55.7756 18.8315
10 0.0656 72.9139 55.1888 18.7622 5 0.1386 80.2932 56.0651 18.7183
5 0.0693 76.3199 55.1913 18.7153
Z_damp (%) 0.1459 79.5182 55.1677 18.9945
Xs_opt (%) 0.0729 79.5182 55.1677 18.9945 5 0.1532 78.1869 55.6216 18.8546
5 0.0765 82.7353 55.1941 18.7796 10 0.1605 77.4393 56.2400 18.7616
10 0.0802 86.0480 55.1420 18.9328 20 0.1751 75.8834 53.7766 18.4606
20 0.0875 89.6216 55.1033 18.8448 30 0.1897 74.6872 54.3648 18.1445
30 0.0948 77.7404 55.1408 18.9618 40 0.2043 73.0979 50.8689 17.4311
40 0.1021 70.8614 55.0443 18.6221 50 0.2189 73.6208 54.9471 17.6554
50 0.1094 63.5231 55.1184 17.6940

References

[1] Bagepalli BS. Dynamic modeling of washing machine suspension systems. In:
Deviation from Z_spr f1 f2 f3 ASME 11th biannual conference of mechanical vibrations and noise.
Massachusets, USA; September 27–30, 1987. p. 13–8.
optimum (%) [2] Conrad DC, Soedel W. On the problem of oscillatory walk of automatic washing
50 0.0859 67.5764 47.6900 18.3874 machines. J Sound Vib 1995;188(3):301–14.
[3] Türkay OS, Sümer IT,_ Tuğcu AK. Modeling and experimental assessment of
40 0.1030 68.5217 54.7428 18.1314 suspension dynamics of a horizontal-axis washing machine. Trans ASME: J Vib
30 0.1202 70.2937 55.1483 18.5122 Acoust 1998;120:534–43. April.
20 0.1374 84.5210 55.3227 18.8724 [4] Lim HT, Jeong WB, Kim KJ. Dynamic modeling and analysis of drum-type
washing machine. Int J Precision Eng Manuf 2010;11(3):407–17. June.
10 0.1545 86.6495 55.2343 18.9295 [5] Öztürk E, Erol H. On the dynamics of a washing machine with flexible
5 0.1631 82.7201 55.2249 18.7955 components. Noise Control Eng J 2010;58(6). November.
[6] Zuoxin P. A study of vibration of a fully automatic washer and optimizing of
Zs_opt (%) 0.1717 79.5182 55.1677 18.9945 structure parameters. In: Proc 42nd annual international appliance technical
5 0.1803 76.7890 55.1920 18.7257 conference. Uni. of Wisconsin, Madison, USA; May 21–22, 1991. p. 545–56.
10 0.1889 74.2533 55.1827 18.4682 [7] Kiray B. Parametric analysis and suspension design optimization of horizontal-
axis washing machines, MSc Thesis. Bogazici Uni., Istanbul, Turkey; 1991.
20 0.2060 70.2846 55.1906 18.5474 [8] Turkay OS, Sumer IT, Tugcu AK, Kiray B. Suspension design optimization of a
30 0.2232 65.9450 54.7624 18.2078 washing machine: Part I – modeling and validation results. Part II –
40 0.2404 64.0877 55.1761 18.0041 formulation and implementation of parametric optimization. In: Proceedings
of the ASME advances in design automation conference. Albuquerque, New
50 0.2576 62.1459 55.1477 17.7662 Mexico; September 19–22, 1993.
[9] Papadopoulos E, Papadimitriou I. Modeling and control of a portable washing
machine during spinning cycle. In: Proc. of the 2001 IEEE/ASME international
conference on advanced intelligent mechatronics systems (AIM 2001). Como,
Italy; 8–11 July 2001. p. 899–904.
[10] Spelta C, Previdi F, Savaresi SM, Fraternale G, Gaudiano N. ‘Control of
Deviation from X_damp f1 f2 f3 magnetorheological dampers for vibration reduction in a washing machine’.
Mechatronics 2009;19:410–21.
optimum (%)
[11] Nygårds T, Berbyuk V. Multibody modeling and vibration dynamics analysis of
50 0.0436 66.8306 45.8639 15.8951 washing machines. Multibody System Dynamics, 27: 197–238, first published
on: 30 December 2011 (iFirst); 2012. http://dx.doi.org/10.1007/s11044-011-
40 0.0523 68.8640 46.1163 16.1166 9292-5.
P. Boyraz, M. Gündüz / Mechatronics 23 (2013) 581–593 593

[12] Nygårds T, Berbyuk V. Pareto optimization of a washing machine suspension instrumentation and measurement technology conference, vol. 2. Budapest;
system. In: Proc of the 2nd international conference on engineering 2001. p. 1123–29.
optimization. Lisbon, Portugal; September 6–9, 2010. p. 1–10. [17] Matsuoka N, Hayakawa N, Yashiba Y, Ishida Y, Honda T, Ogawa Y. Pitch
[13] Nygårds T, Berbyuk V. Optimization of washing machine kinematics, estimation using discete analytic signals. In: The Australasian Speech Science
dynamics, and stability during spinning using a multistep approach. and Technology Association SST conference, vol. 2. Perth; 1994. p. 238–43.
Optimization and engineering, first published on: 05 December 2012 (iFirst); [18] Schafer RW. On the frequency domain properties of Savitzky–Golay filters. In:
2012. http://dx.doi.org/10.1007/s11081-012-9206-2. Digital signal processing workshop and IEEE signal processing education
[14] Christos Y, Gryllias K, Antoniadis I. Instantaneous frequency estimation in workshop. Sedona; 2011. p. 54–9.
rotatiing machinery using a harmonic signal decomposition parametric [19] Poussot-Vassal C, Spelta C, Sename O, Savaresi SM, Dugard L. Survey and
method. In: ASME International engineering technical conferences & performance evaluation on some automotive semi-active suspension control
computers and information in engineering conference. California, USA; 2009. methods: a comparative study on a single-corner model. Annu Rev Control
[15] Akan A, Yalçın M, Chaparro LF. An iterative method for instantaneous (IFAC – Elsevier) 2012;36(1):148–60.
frequency estimation. IEEE Electron Circ Syst ICECS 2001;3:1335–8.
[16] Angrisani L, D’Arco M. A measurement method based on an improved version
of the chirplet transform for instantaneous frequency estimation. In: IEEE

You might also like