Advances in Engineering Software 98 (2016) 112–122

Contents lists available at ScienceDirect

Advances in Engineering Software

journal homepage: www.elsevier.com/locate/advengsoft

Optimization design of a disc brake system with hybrid uncertainties

Hui Lü a,∗, Dejie Yu b
a
School of Mechanical and Automotive Engineering, South China University of Technology, Guangzhou City 510641, P R China
b
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan University, Changsha City 410082, P R China

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

Article history: Squeal reduction of disc brake systems have been extensively investigated for both academic and in-
Received 27 January 2016 dustrial purposes. However, most of the existing optimization designs of squeal reduction are based on
Revised 17 March 2016
deterministic approaches which have not considered the uncertainties of material properties, loading con-
Accepted 17 April 2016
ditions, geometric dimensions, etc. In this paper, a hybrid probabilistic and interval model is introduced
to deal with the uncertainties existing in a disc brake system for squeal reduction. The uncertain param-
Keywords: eters of the brake system with enough information are treated as probabilistic variables, while the pa-
Optimization design rameters with limited information are treated as interval variables. To improve computational efficiency,
Brake squeal the response surface methodology (RSM) is introduced to replace the time-consuming finite element (FE)
Hybrid uncertainties
simulations. By the hybrid uncertain model, an optimization design based on reliability and confidence
Probabilistic variable
interval is proposed to explore the optimal design for squeal reduction. In the proposed optimization,
Interval variable
both the design objective and the design constraint are interval probabilistic functions due to the effects
of hybrid uncertainties. In this case, the maximum of the upper bound of confidence interval of design
objective is selected as the objective function, while the minimal value of the probabilistic constraint is
selected as the constraint function. The combinational algorithm of Genetic Algorithm and Monte-Carlo
method is employed to perform the optimization. The results of a numerical example demonstrate the
effectiveness of the proposed optimization on reducing squeal propensity of the disc brake systems with
hybrid uncertainties.
© 2016 Elsevier Ltd. All rights reserved.

1. Introduction methods to determine the dominant modal parameters of sub-

The friction-induced vibrations existing in disc brake systems
can induce a dynamic instability, and causes an inconvenient
squealing noise. Disc brake squeal has become one of the most dif-
ficult concerns associated with vehicle disc brake systems, which
have poor performances on system stability [1]. Brake squeal, es-
pecially with the frequency range from 1 to 16 kHz, frequently
leads to customer complaints and results in enormous warranty
costs. Therefore, extensive efforts have been undertaken by indus-
trial corporations as well as by the scientific communities to pre-
dict and remove the squeal noise, and some interesting review pa-
pers have been presented on this subject [2–10]. So far, however,
there is not yet a comprehensive understanding of the root cause
of this phenomenon due to its immense complexity.
Optimization designs of disc brake systems for squeal reduction
have been extensively studied in the field of automotive engineer-
ing. For example, Guan et al. [11] suggested sensitivity analysis
∗
Corresponding author. Tel.: + 86 73188821915/+8613574884972; fax: +86
73188823946.
E-mail address: lvhui588@126.com (H. Lü).
structures of a disc brake system for squeal suppression, and the
dominant modal parameters were selected as optimization target
to explore modifications of disc and bracket to eliminate squeal
modes. Spelsberg-Korspeter [12,13] performed a structural opti-
mization of brake rotor, and the mathematical difficulties of such
an optimization were discussed. Lakkam and Koetniyom [14] pro-
posed an optimization study for squeal reduction, which aimed
to minimize the strain energy of vibrating pads with constrained
layer damping. The results of their research could guide to specify
the position of the constrained layer damping patch under pressure
conditions. Shintani and Azegami [15] presented a solution to a
non-parametric shape optimization problem of a disc brake model
to suppress squeal noise, the optimum shape of brake pad was
found and the real part of the complex eigenvalue representing the
cause of brake squeal was minimized. The above-mentioned stud-
ies on optimization designs of disc brake systems are all restricted
to deterministic optimizations, in which all design variables and
parameters involved are regarded as certain values. However, due
to the effects of manufacturing/measuring errors, aggressive en-
vironment factors and unpredictable external excitations, uncer-
tainties associated with loading, material properties, geometries,

http://dx.doi.org/10.1016/j.advengsoft.2016.04.009
0965-9978/© 2016 Elsevier Ltd. All rights reserved.
 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122
and environmental conditions are unavoidable. Without consider-
ing the uncertainties, the optimality yielded by deterministic op-
timization approaches will be destroyed and the constraint condi-
tions will be violated [16–18]. In order to take into account various
uncertainties, the reliability-based design optimization (RBDO) is
introduced and has been intensively studied both in the method-
ology and in applications [19–21]. In RBDO, the performances and
reliabilities of uncertain systems can be considered simultaneously.
Compared with the deterministic optimization, RBDO aims to seek
a reliable optimum by converting deterministic constraints into
probabilistic constraints. Therefore, RBDO can be considered as a
potential method to improve the dynamic performance of disc
brake systems considering uncertainties of practical engineering.
Even though great success has been achieved in the stability
analysis and optimization design of disc brake systems for squeal
reduction, there are only a few papers that investigate brake squeal
problem with uncertainties at present. Of those papers, Grange
et al. [22] have used the random decrement technique and the
Ibrahim time domain method to identify the modal parameters of
an equivalent linear brake system for squeal analysis. Sarrouy et
al. [23] have carried out an uncertain simulation of brake squeal
based on polynomial chaos expansions and a simplified disc brake,
in which the friction coefficient and the contact stiffness of the
brake are modeled as random parameters. Tison et al. [24] have
proposed a complete strategy to improve the prediction of squeal
simulations by introducing random uncertainty and robustness
concepts. Lü and Yu [25] have presented an uncertain optimiza-
tion method for brake squeal reduction of vehicle disc brake sys-
tems with interval parameters. In this research, the uncertain sys-
tem parameters are all treated as interval variables due to limited
data of system parameters. The interval method seems somewhat
conservative. Based on the work of [25], both random variables
and interval variables are used to deal with the hybrid uncertain-
ties of brake system and the stability of such a brake system has
been investigated by the same authors recently [26]. Nevertheless,
the stability optimization design and the numerical optimization
method of such a hybrid uncertain brake system have not been ex-
plored in [26]. Another recent work on uncertain squeal problem
can be found in [27]. In this research, Nobari et al. [27] presented
a random process and a Kriging surrogate model to overcome the
high computational workloads of uncertain analysis of brake sys-
tems, and three benefits gained from the surrogate model were
discussed, namely coming up with some design recommendations
to reduce brake noise, quantifying uncertainty and variability ex-
isting in a brake system and conducting a reliability analysis in
terms of noise. This research mainly focused on the construction
and application of the surrogate model with uncertainties, but the
uncertainty analysis and optimization method for squeal reduction
was not explored, just as in [26].
Probabilistic methods are the traditional approaches to cope
with the uncertainties arising in practical engineering problems,
just as we can see in the above researches [22–27]. In the prob-
abilistic methods, the uncertain parameters are treated as proba-
bilistic variables whose probability distributions are defined unam-
biguously [28]. To construct the precise probability distributions of
probabilistic variables, a large number of statistical information or
experimental data is required. Unfortunately, in the design stage of
disc brake systems, the information to construct the precise prob-
ability distributions of some probabilistic variables (e.g. the fric-
tion coefficient) is not always sufficient due to the immeasurability
or assumptions. For this case, the hybrid probabilistic and interval
model has been proposed for overcoming the deficiencies of proba-
bilistic methods. In the hybrid probabilistic and interval model, the
uncertain parameters with sufficient data to construct probabilistic
distributions are treated as probabilistic variables, while the uncer-
tain parameters without sufficient data to construct the distribu-
113
tions are treated as interval variables. The hybrid probabilistic and
interval model was firstly proposed by Elishakoff et al. [29,30], and
subsequently applied to the response analysis of hybrid uncertain
systems [31,32]. As previously mentioned, the disc brake systems
are treated as either deterministic systems or almost probabilistic
uncertain systems in the existing researches. From the overall per-
spective, researches on the hybrid probabilistic and interval model
are still in its preliminary stage and some important issues are still
unsolved. For example, the application of the hybrid probabilistic
and interval model in the optimization of brake squeal is not yet
explored.
The purpose of this paper is to take into account the hybrid
uncertainties existing in a disc brake system, and develop an op-
timization approach for improving system stability and reducing
squeal propensity. By using hybrid probabilistic and interval model,
the uncertainties of the thickness of the back plate, the Elastic
modulus of component materials and the brake pressure are rep-
resented by probabilistic variables, whose distribution parameters
are well-defined; whereas the uncertainties of the friction coef-
ficient, the densities of component materials and the assumptive
Elastic modulus of the friction material are modeled by interval
variables, whose lower and upper bounds are well-defined. Based
on the hybrid uncertain model, the optimization design based on
reliability and confidence interval for a disc brake system is pro-
posed. To improve the computational efficiency of the proposed
optimization, RSM is employed to establish the surrogate model of
the real part of the domain unstable eigenvalue, which is adopted
as the design objective. The mass of the design component is taken
as the design constraint. Due to the effects of hybrid uncertainties,
both the design objective and design constraint are interval proba-
bilistic functions. In this case, the maximum of the upper bound of
confidence interval of design objective is selected as the objective
function, while the minimal value of the probabilistic constraint is
selected as the constraint function. The combinational algorithm
of Genetic Algorithm and Monte-Carlo method is employed to per-
form the optimization. The effectiveness of the proposed approach
is demonstrated by a numerical example.
2. Complex eigenvalue analysis of disc brake systems
2.1. A simplified disc brake system
Disc brake systems are one of the most important safety and
performance components in automobiles. There are several major
components in a car disc brake system: brake disc, brake pad as-
semblies, carrier bracket, calliper and a hydraulic actuation system.
The brake disc is rigidly mounted on the axle hub and rotates with
the wheel. The pair of brake pad assemblies generally consist of
friction material and back plates. When the hydraulic pressure is
applied, the piston of actuation system is pushed forward to press
one brake pad against the brake disc and simultaneously the other
brake pad is pressed by the calliper against the disc. Then, a fric-
tional torque is generated to slow the disc rotation. For the pur-
pose of simulating the vibration characteristics of a disc brake sys-
tem reasonably with the acceptable computational burden, a sim-
plified model of disc brake system is taken for investigation. The
simplified disc brake consists of a disc and a pair of brake pads, as
shown in Fig. 1. The similar models have been previously consid-
ered and successfully used by some studies, such as [23,33–35].
2.2. Complex eigenvalue analysis
The CEA can be carried out by two stages. In the first stage, the
steady state of the brake system is found and the actual CEA is
performed in the second stage.
114 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122
Brake disc
Brake pad
Friction material
Back plate
Fig. 1. The simplified model of a disc brake system.
In the steady state of the brake system, the equation of motion
for the brake system is as follows
 
M y +D y +Ky = F
where M, D and K are mass, damping and stiffness matrices, re-
spectively; y is the steady-state displacement vector; The force F
is mainly resulted from the brake pressure and the friction forces
caused by the relative motion at the friction interface. The friction
interface can be modeled by using linear elements to account for
the nonlinear behavior of the contact [8], and then the force vector
F becomes linear
F=K f y
where Kf is the friction stiffness matrix. Substituting Eq. (2) into
Eq. (1) gives
  mines the response α (x); a0 , ai , aij and aii are the estimated re-
M y +D y + ( K − K f )y = 0
By the influence of friction, the new stiffness matrix (K − K f ) is
asymmetric, which may leads to complex eigenvalue problem. The
complex eigenvalue problem can be formulated as
(Mλ2 + Dλ + K∗ )φ = 0
where K∗ is the new stiffness matrix; λ is the complex eigenvalue
and φ is the complex eigenmode. The complex eigenvalues of the
system can be obtained by solving the complex eigenvalue prob-
lem of Eq. (4). The complex eigenvalue λ can be expressed as
λ = α + jω
where α and ω are the real and imaginary parts of the complex
eigenvalueλ, respectively.
It is widely known that a vibration system is unstable when the
real part of a complex eigenvalue of the system is positive. There-
fore, the positive real parts of the eigenvalues can be taken as in-
dicators for system stability and brake squeal. The main aim of this
study is to minimize the positive real part of the dominant unsta-
ble eigenvalue which can be determined according to the practical
situation.
It is worth pointing out that the main limitation of CEA is the
use of a linear model, despite the strong nonlinearities associated
with the contact problems. Nevertheless, it has been proved that
squeal instability occurs in linear conditions. A detailed description
of linearization of the broke model can be found in [8].
3. Optimization design of a disc brake system with hybrid
uncertainties for squeal reduction
3.1. Surrogate model based on RMS
In engineering design, the coupling of optimization algorithm
with finite element analysis (FEA) maybe inefficient, since the iter-
ative analyses during optimization usually require enormous itera-
tions and high computational cost, especially for the RBDO prob-
lem. As a result, the surrogate models have been widely adopted
in engineering applications to alleviate the computational burden.
This approach establishes an explicit mathematical relationship be-
tween design variables and functional responses with a moderate
number of FEA runs. As an effective approach, the response sur-
face (RS) model is one of the simplest and most popular surrogate
models, which can be treated as an effective alternative to FEA and
has been widely adopted in optimization design [36].
Mathematically, by using the RS model, the real part of a com-
plex eigenvalue of FEA can be defined in terms of basis functions
as

N
α (x ) = ak ϕk ( x ) (6)
k=1
where α (x) is the real part of a complex eigenvalue; ϕk (x ) (k =
(1) 1, 2, · · · , N ) are the basis functions; N is the number of basis func-
tions; ak (k = 1, 2, · · · , N ) represent the regression coefficients.
When using a quadratic model [37], the full set of the second
order polynomials of ϕ k (x) are given as
1, x1 , x2 , · · · , xn , x21 , x1 x2 , · · · , x1 xn , · · · , x2n (7)
where n is the number of the variables. The surrogate model of
α (x) could be thus defined as

n  
n
(2) α ( x ) = a0 + ai xi + ai j xi x j + aii x2i (8)
i=1 i j (i< j ) i=1
where x = [x1 , x2 , · · · , xn ]T denotes the variable vector that deter-
(3)
gression coefficients which can be obtained from the design of ex-
periment (DOE) and the least square method (LSM) [38]. The cross
product terms xi xj represent the two-variable interactions and the
square terms x2i represent the second order nonlinearity. After a
RS model is established, its accuracy should be assessed. An anal-
(4)
ysis of variance (ANOVA) is necessary to be conducted to test the
model, so as to ensure the fitting accuracy and significance of the
model [39]. Generally speaking, each technique has its own ad-
vantages and disadvantages. One of the major drawbacks of the
quadratic polynomial RSM is that the good fitting accuracy can
hardly be got when the problem is highly nonlinear. Nevertheless,
(5)
from Refs. [25] and [37], it can be seen that the unstable eigen-
value of a disc brake can be well fitted by the second order RS
model.
3.2. Disc brake systems with hybrid uncertainties
In the traditional deterministic researches based on FEA, system
parameters of the disc brake are assumed to be known precisely
and defined exactly. However, most of the data used in the solu-
tion process of many practical engineering systems is either col-
lected from experiments or acquired as empirical data from the
past, which is usually ill defined, imprecise and uncertain in na-
ture. Due to the complex working environment and operating con-
ditions, the occurrence of brake squeal is intermittent or perhaps
even casual in general, and it is difficult to be captured and re-
produced artificially. This is possibly related to the uncertainties
existing in the brake system. For instance, the friction coefficient
changes, but is not constant, in the course of braking. This friction
characteristic makes it possible to provide the energy source for
brake squeal [40,41]. So it will be more reasonable and of great
significance if uncertainties are introduced to the stability analysis
of disc brake systems.
For an actual disc brake system, uncertainties in material, load-
ing and contact properties and manufacturing errors are unavoid-
able. Usually, the most common model to deal with the uncertain-
ties is the probabilistic model. By the probabilistic model, the α (x)
 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122
in Eq. (8) for the uncertain optimization problem, can be generally
expressed as
α (x ) = α (xD , xU )
where xD is the design variable vector; xU is the uncertain vari-
able vector which is taken as non-design variables and treated as
probabilistic variables.
In actual situations, the disc brake system is a rather complex
system and it is difficult or even impossible to obtain the precise
values of the distribution parameters of some variables (e.g. the
friction coefficient). In order to model the uncertainties of these
variables more close to the actual conditions, these uncertain vari-
ables can be treated as interval variables, whose variation ranges
are well-defined. The α (x) in Eq. (9) then becomes an interval
probabilistic objective function for a hybrid uncertain optimization
problem, and can be expressed as
α (x ) = α (xD , xU )=α (xD , xP , xI )
where xU ={xP , xI }; xD = [xD,1 , xD,2 , · · · , xD,k ]T is the design
variable vector and k is the number of design variables;
xP = [xP,1 , xP,2 , · · · , xP,l ]T is the probabilistic variable vector which is
taken as non-design variables and l is the number of probabilistic
variables; xI = [xI,1 , xI,2 , · · · , xI,m ]T is the interval variables vector
which is taken as non-design variables as well and m is the
number of interval variables.
Due to the manufacturing or measuring errors, the actual values
of the design variable xD may deviate from the theoretical values
in engineering practice. In this case, the design variables xD can be
treated as the uncertain variables as well.
3.3. The optimization design based on reliability and confidence
interval for a disc brake system with hybrid uncertainties
Mathematically, the RBDO model of a disc brake system with
hybrid uncertainties can be generally expressed as
min
xD
α ( x ) = α ( xD , xP , xI )
s.t. P (gr (xD , xP , xI ) ≥ 0 ) ≥ ηr , r = 1, 2, · · · , t
xD,lower ≤ xD ≤ xD,upper
where α (xD , xP , xI ) is the interval probabilistic objective func-
tion. gr (xD , xP , xI ) is the r − th limit–state function, and t is the
number of limit–state functions. gr (xD , xP , xI ) = 0 is the limit–
state equation by which the design space is divided into the safe
region gr (xD , xP , xI ) > 0 and the failure region gr (xD , xP , xI ) < 0.
P (gr (xD , xP , xI ) ≥ 0 ) is the probability for gr (xD , xP , xI ) ≥ 0. ηr is
the r − th system reliability index which is close to 1 with the in-
crease of the required reliability level.
The objective function α (xD , xP , xI ) is an interval probabilistic
function. Without considering the interval variables xI temporar-
ily, the objective function α (xD , xP , xI ) can be treated as a pure
probabilistic function. The confidence interval can be used to de-
scribe the probabilistic characteristic of the values of α (xD , xP , xI ).
The confidence interval for the probabilistic response is a range
within which the true response likely falls. The probability that
the confidence interval contains the true response is called the
confidence level (marked as cl). The confidence level is a chosen
value close to 1, usually 0.90, 0.95 or 0.99. For example, if the
confidence level is cl = 0.99, it will indicate that we are 99% con-
fident that the true response is in the confidence interval. The
upper bound of the confidence interval of α (xD , xP , xI ) at confi-
dence level cl (marked as Ccl (α (xD , xP , xI ))upper ) is the maximal
value of α (xD , xP , xI ) which may occur at the considered confi-
dence level cl. As the interval variables exist actually, the obtained
Ccl (α (xD , xP , xI ))upper is a function of interval variables. The un-
certainty of Ccl (α (xD , xP , xI ))upper suffering from interval variables
115
can be expressed as the variation range of Ccl (α (xD , xP , xI ))upper . If
the upper bound of Ccl (α (xD , xP , xI ))upper satisfies the design re-
quirement, the other values of Ccl (α (xD , xP , xI ))upper in the varia-
(9)
tion range will also satisfy the design requirement. Thus, the upper
bound of Ccl (α (xD , xP , xI ))upper can be considered as the objective
function. In other words, the purpose of the optimization based on
reliability and confidence interval design can be expressed as
min C cl (α (xD , xP , xI ))upper (12)
xD
where C cl (α (xD , xP , xI ))upper is the upper bound of
Ccl (α (xD , xP , xI ))upper .
Without considering the interval variables xI temporarily, the
r − th limit–state function gr (xD , xP , xI ) can be treated as a pure
probabilistic function. The probability of gr (xD , xP , xI ) ≥ 0, namely,
P (gr (xD , xP , xI ) ≥ 0 ) is a deterministic value. As the interval vari-
ables xI exist actually, P (gr (xD , xP , xI ) ≥ 0 ) is a function of interval
(10) variables. P (gr (xD , xP , xI ) ≥ 0 ) is uncertain suffering from the ef-
fects of interval variables, and the uncertainty of P (gr (xD , xP , xI ) ≥
0 ) can be the variation range of P (gr (xD , xP , xI ) ≥ 0 ). It is unneces-
sary to consider all values of P (gr (xD , xP , xI ) ≥ 0 ), this is because if
the lower bound of P (gr (xD , xP , xI ) ≥ 0 ) is larger than ηr , the other
values of P (gr (xD , xP , xI ) ≥ 0 ) within the variation range will be
also larger than ηr . Thus, the constraint conditions can be rewrit-
ten as
P ( g r ( x D , x P , x I ) ≥ 0 ) ≥ ηr , r = 1 , 2 , · · · , t (13)
where P (gr (xD , xP , xI ) ≥ 0 ) is the lower bound of P (gr (xD , xP , xI ) ≥
0 ).
Based on the above discussions, the optimization design based
on reliability and confidence interval for a disc brake system with
hybrid uncertainties can be expressed as
min C cl (α (xD , xP , xI ))upper
xD
s.t. P (gr (xD , xP , xI ) ≥ 0 ) ≥ ηr , r = 1, 2, · · · , t
xD,lower ≤ xD ≤ xD,upper (14)
The optimization problem defined in Eq. (14) involves
two uncertain analysis procedures. One is the calculation of
(11) C cl (α (xD , xP , xI ))upper and the other one is the estimation of
P ( gr ( xD , xP , xI ) ≥ 0 ).
3.4. Calculations of the objective function and constraint function
In practical engineering, the Monte-Carlo method can be
considered as the potential method to calculate the objec-
tive function C cl (α (xD , xP , xI ))upper and the constraint function
P (gr (xD , xP , xI ) ≥ 0 ). The main procedures of the calculations based
on the Monte-Carlo method can be described in the following
Step 1: Generate i − th (i = 1, 2, · · · , Ni ) sample of the interval
variables xI . With the sample values of the interval variables, xI
degenerates to the vector of deterministic variables and can be
marked as xI(i ) ;
Step 2: Generate j − th ( j = 1, 2, · · · , N j ) sample of the proba-
bilistic variable vectors xP . With the sample values of the prob-
abilistic variables, xP degenerates to the vector of deterministic
( j)
value variables, and can be marked as xP ;
( j)
Step 3: Calculate the values ofα (xD , xP , xI )(i, j ) = α (xD , xP , xI(i ) )
( j)
and gr (xD , xP , xI )(i, j ) = gr (xD , xP , xI(i ) );
Step 4: Is the sample of the probabilistic variable vectors xP
ended up? Namely, does j equate to Nj (the number of samples
of probabilistic variable vectors)? If no, j = j + 1 and go on to
Step 2. If yes, calculate the upper bound of the confidence in-
(i )
terval Ccl (α (xD , xP , xI ))upper at the considered confidence level cl
and the probabilityP (gr (xD , xP , xI ) ≥ 0 )(i ) based on all values of
α (xD , xP1 , xP2 )(i, j ) and gr (xD , xP1 , xP2 )(i, j ) , j = 1, 2, · · · , N j ;
116 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122

Start Start

Mathematic model of Mathematic model of RBDO

( x D , x P , x I ) and g r ( x D , x P , x I )

Initialize the design variables

Generate sample of the interval variables x I xD = xD 0

Estimate C cl ( ( x D , x P , x I )) upper and

Generate sample of probabilistic variable s x P
P (g r ( x D , x P , x I ) 0) Update the
design variables
Calculate (x D , x P , x I ) ( i, j) and g r ( x D , x P , x I ) ( i, j)
The end N
condition is met ?
N
Is the sample ended up ? Y

Y Output the optimal design

(i )
Calculate Ccl ( ( x D , x P , x I )) upper and P (g r (x D , x P , x I ) 0) ( i)
End

N Fig. 3. Flowchart of the proposed optimization for a disc brake system with hybrid
Is the sample ended up ? uncertainties.

Y
Select C cl ( ( x D , x P , x I ))upper and P (g r (x D , x P , x I ) 0) , and export them

End

Fig. 2. Flowchart of the calculations of the objective function and constraint func-
tion by using the Monte-Carlo method.

Fig. 4. The FE model of a simplified disc brake.

Step 5: Is the sample of interval variables xI ended up? Namely,
does i equate to Ni (the number of samples of interval vari-
able vector)? If no, i = i + 1 and go on to Step 1. If yes, select Chamfer Slot
Chamfer
C cl (α (xD , xP , xI ))upper and P (gr (xD , xP , xI ) ≥ 0 ) from all values of
(i )
Ccl (α (xD , xP , xI ))upper and P (gr (xD , xP , xI ) ≥ 0 )(i ) , i = 1, 2, · · · , Ni ,
and export them.
The procedures can be summarized in the flowchart shown in
Back plate
Fig. 2.
Fig. 5. The chamfers and slot of the friction material.
4. The procedure of the proposed optimization design

Based on Sections 2 and 3, the main procedure of the optimiza-
tion design based on reliability and confidence interval for a disc
brake system with hybrid uncertainties can be summarized as fol-
lows.
Step 1. Construct the mathematic model of optimization (shown
in Eq. (14)) for the disc brake system with hybrid uncertainties,
based on CEA (shown in Eq. (4)) and RSM (shown in Eq. (8));
Step 2. Initialize the design variables;
Step 3. Estimate the objective function C cl (α (xD , xP , xI ))upper
and the constrain function P (gr (xD , xP , xI ) ≥ 0 ) based on the
Monte–Carlo method;
Step 4. Whether the end condition of the optimization algo-
rithm is met? If no, update the design variables and go back to
Step 3. If yes, output the optimal design.
The flowchart of the optimization is given in Fig. 3.
5. Numerical example
5.1. FE model of a disc brake system
The simplified disc brake system shown in Fig. 1 is adopted to
investigate in this section. The FE model of the brake system is
Brake pressure
Bolt holes
Uy free
Uy free
Fig. 6. Constraints and loadings of the brake system.
illustrated in Fig. 4. The FE meshes are generated by using three-
dimensional continuum elements. The element type is C3D8I and
fine meshes are used in the contact areas of disc and pads.
For this commercial brake system shown in Fig. 1, some com-
mon and effective measures have been taken to reduce the likeli-
hood of squeal in the early design stage, for example, two cham-
fers and a slot have been respectively provided on both sides and
in the middle position of the friction material of each original pad,
as shown in Fig. 5. Therefore, the degree of system stability of this
brake is already relatively high.
Fig. 6 presents the constraints and loadings of this disc brake
assembly. The disc is completely fixed at the five counter-bolt
holes. The back plates are constrained in the ears, and can just
 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122 117

Table 1
The investigated parameters and their nominal values. 100
Investigated parameter Unit Nominal value
50
Friction coefficient f / 0.35

Real part
Brake pressure p MPa 0.50
Density of back plate ρ 1 kg/dm3 7.82 0
Elastic modulus of back plate E1 GPa 207
Thickness of back plate d1 mm 6.50 -50
Density of friction material ρ 2 kg/dm3 2.51
Elastic modulus of friction material E2 GPa 5.94
Density of disc ρ 3 kg/dm3 7.20 -100
Elastic modulus of disc E3 GPa 125
0 2000 4000 6000 8000 10000 12000 14000 16000
Imaginary part /Hz

move along the axial direction. The brake pressure is directly act- Fig. 7. The distribution of the complex eigenvalues of the brake system for some
ing on the back plate in the contact area between the pad and the samples.
piston, and an equal magnitude of brake pressure is similarly ap-
plied to the other back plate.
This study aims to optimize the real part of the dominant
unstable eigenvalue for reducing potential squeal. The eigenvalue
analysis is depended on a number of factors and mainly based on
the stiffness and mass characteristics and operating conditions of
the system. Based upon this, the brake system parameters that we
select to investigate and their nominal values are listed in Table 1.
It should be specially pointed out that an assumption of the
elastic property for the friction material has been made by us.
Fig. 8. The unstable mode of the brake at about 2 kHz.
For the investigated disc brake system, the friction material has
the anisotropic properties. Referring to some earlier works, e.g. the
studies of [35] and [42], we treat the friction material as isotropic
other unstable eigenvalues can also be found at other frequencies.
materials but with varied Elastic modulus, which can be seen in
However, CEA usually over-predicts the number of unstable eigen-
Table 5. This assumption is reasonable and acceptable in an uncer-
values. It means that not all the unstable eigenvalues obtained by
tainty analysis.
CEA should be taken for study, as in [27]. Thus, just the dominant
unstable eigenvalue is considered in this study.
5.2. Determination of the dominant unstable eigenvalue
Under the nominal values of system parameters, the mode
shape of the unstable mode corresponding to the dominant un-
Latin Hypercube Sampling (LHS) method [43] is adopted to gen-
stable eigenvalue is displayed in Fig. 8. It can be seen that the un-
erate the samples for building up the RS model. Regarding the size
stable mode appears mainly in the form of out-of-plane mode of
of samples, it has been found that 1.5 times of the number of the
the brake disc. It is also worth mentioning that the unstable modes
model coefficients are needed to generate a reasonably accurate RS
corresponding to the dominant unstable eigenvalues under differ-
model in practical engineering [36,44,45]. It means that minimum
ent samples all appear in the form of this mode shown in Fig. 8.
83 samples are needed to generate in this study. Therefore, a total
Since CEA usually over-predicts the number of unstable eigen-
of 90 samples are generated in the sample space, which is con-
values [8,27,33], a dynamic transient analysis model of brake
structed by the variations ( ± 15% from the nominal values) of
squeal is constructed for the same disc brake system to validate
the investigated parameters shown in Table 1. Appropriate more
the dominant unstable eigenvalue. A detailed discussion on the
samples will be useful for improving the accuracy of the RS model
loadings and constraints of this analysis model is available in Ref.
certainly. It has great mathematical and theoretical significance to
[33]. The dynamic transient analysis is carried out under the nom-
study the accuracy of the surrogate model under more samples, as
inal values of system parameters, and the analysis results are illus-
in [27]. However, this study is mainly focused on the application
trated in Fig. 9. In Fig. 9(a), the acceleration response of the cen-
of the RS model but not its construction and mathematical details.
troid of a pad in the time period from 0.02 s to 0.03 s is displayed
Thus, more samples are no longer considered for the construction
(loadings and constraints application is in the time period from 0
of the RS model while its accuracy meets the engineering require-
to 0.02 s). In Fig. 9(b), the acceleration response of time history is
ment. For each of the 90 samples, CEA is performed in a personal
converted into frequency domain using the fast Fourier transfor-
computer based upon Eq. (4) and the FE model. The frequency
mation through a Matlab code.
range of eigenvalue analysis is from 0 to 16 kHz, where the brake
Comparing the results shown in Fig. 7 and Fig. 9, it can be seen
squeal commonly occurs. The analysis results show that, the un-
that there is a good agreement between the CEA results and the
stable eigenvalues mainly appear around the frequency of 2 kHz in
transient analysis results on the prediction of dominant unstable
this study. Based on some earlier works, such as [27] and [37], we
eigenvalue. A more reliable validation could be achieved if it is
take the unstable eigenvalues at the certain frequency for study-
possible to compare these numerical simulation results with an ex-
ing. Therefore, the unstable eigenvalue at 2 kHz for each sample is
perimental result. Unfortunately, carrying out an experimental test
regarded as the dominant unstable eigenvalue whose real part is
is not affordable in this current study.
positive and needed to be reduced to improve system stability.
For a clearer understanding, some samples are selected arbitrar-
ily from the 90 samples for investigation, and the complex eigen- 5.3. Construction of RS model
values of the brake system corresponding to these samples are
depicted in Fig. 7. In the figure, the same marker represents the Based on the LSM approach and the CEA results of the 90
complex eigenvalues of the same sample. It can be seen that the samples, the second order RS model of the real part of the
major unstable frequency is approximately 2 kHz. In addition, some dominant unstable eigenvalue is constructed according to Eq. (8).
118
a b
8 5
x 10
Acceleration a /(mm/s 2 )
5
0 Y: 2.089e+005
-5
0.02
Table 2
ANOVA results of the RS model α (x).
Source
Regression
Residuals
Total
The RS model is marked as α (x), where x = [ f, p, ρ1 , ρ2 , ρ3 ,
E1 , E2 , E3 , d1 ]T is the vector of investigated parameters. The detailed
quadratic polynomials of α (x) is shown in Appendix A. Indeed, a
single RS model cannot completely present all unstable eigenval-
ues of the brake. However, only the dominant unstable frequency
is found in the transient analysis. It means that considering the
dominant complex eigenvalue is appropriate and sufficient in this
study. There is no need to consider other unstable eigenvalues of
this brake.
It is critical to validate the significance and adequacy of α (x).
For this purpose, ANOVA is carried out and the analysis results are
summarized in Table 2.
The coefficient of determination R2 is defined as the ratio of
the explained variable to the total variation. It is a measure of the
degree of fitness. When R2 is close to 1, the model fits the actual
experimental data better. The suggested R2 value should be at least
0.80 for a good fit of a model [46]. It can be seen from Table 2 that
the R2 value of the RS model α (x) is higher than 0.80, which in-
dicates that α (x) can explain the observed response well. While
the p value is lower than 0.01, it indicates that the model is con-
sidered to be statistical significant [46]. It can be seen from Table
2 that the RS model α (x) is highly significant, as the p value of the
model is less than 0.0 0 01. Hence, the RS model α (x) is successful
in describing the correlation between the system parameters and
the system response.
Furthermore, just as [27,36], additional 15 samples are gener-
ated by LHS to assess the prediction accuracy of the RS model α (x).
The real parts of the dominant unstable eigenvalues produced by
the original FE model and by the RS model of the 15 samples are
shown in Table 3. It can be seen that the prediction accuracy of
the RS model is very well from the very low errors (all lower than
± 5%).
5.4. Hybrid uncertainties existing in the disc brake
In practice, uncertainties widely exist in a disc brake system.
In order to model the uncertainties more accurately, the parame-
ters of the brake system with enough information are treated as
probabilistic variables, while the parameters with limited data are
treated as interval variables. Namely, the uncertain variables of the
disc brake system are divided into probabilistic variables xP and
interval variables xI in this study.
H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122
x 10
3
Amplitude
2
X: 1942
1
0
0.022 0.024 0.026 0.028 0.03 0 5000 10000 15000
Time t /s Frequency f /Hz
Fig. 9. Transient analysis results of the brake system with the nominal values.
Sum of squares Degree of freedom Mean square F value p value
12551.19 54 232.43 695.83 <0.0 0 01
8.35 25 0.33
12559.54 79
R=0.9996 R2 =0.9993
The disc and the friction materials are the most key compo-
nents of the brake system, which directly influence the brake per-
formance. Therefore, they should not be changed if possible and
they are considered as non-design components. The parameters of
the disc and the friction materials are all selected as non-design
and uncertain variables.
According to the engineering experience, the brake pressure
and the Elastic modulus of the disc are considered as probabilistic
variables, namely xP = [ p, E3 ]T . Their probability distributions and
the values of distribution parameters are listed in Table 4.
According to the engineering experience, the densities of the
friction material and the disc are assumed as interval variables. For
the assumption of the elastic property of the friction material as
previously mentioned, the Elastic modulus of the friction material
is assumed as interval variables as well. Since the friction coeffi-
cient changes with respect to the sliding velocity between the pad
and disc in the process of braking, it is difficult or even impossible
to obtain its precise value. As a consequence, the friction coeffi-
cient is also taken as interval variable. Namely xI = [ f, ρ2 , ρ3 , E2 ]T .
The values of the characteristic parameters of the interval variables
are listed in Table 5.
For the uncertain variables listed in Tables 4 and 5, it is worth
noting that, their variations should be obtained by some reliable
experimental data. Unfortunately, there is few published data for
the uncertain variables used in this study, as pointing out by [27].
Carrying out some experiments to obtain the variations of these
uncertain variables is not affordable in the current investigation.
In practical engineering, the variation of an uncertain parameters
without reliable experimental data is usually taken as within ± 5%
from its nominal value [27]. Therefore, the variations of the vari-
ables listed in Tables 4 and 5, besides f and E2 , are all taken as
within ± 5% from their nominal values. The friction coefficient f
takes the common range [0.3,0.4] while the variation of E2 is taken
as within ± 10% from its nominal value due to the assumption on
its elastic property as previously mentioned. In addition, the vari-
ations of the probabilistic variables are normally distributed about
their mean values, and then it is easy to know that the standard
deviations of p and E3 are about 0.008 MPa and 2 GPa, respectively,
according to three-sigma principle. The variations of the design
variables used in this studied is defined similarly.
The back plate of the disc brake system is considered as the
design component. The density, the Elastic modulus and the thick-
 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122 119

Table 3
Comparison of the FE results and predicted results of the new samples.

No. f p(MPa) ρ 1 (kg/dm3 ) E1 (GPa) d(mm) ρ 2 (kg/dm3 ) E2 (GPa) ρ 3 (kg/dm3 ) E3 (GPa) FE results Predicted results Error(%)

1 0.2975 0.5429 7.652 2.833 7.2 207 6.195 116.96 6.779 29.65 30.49 2.83
2 0.305 0.5214 6.647 2.779 6.891 198.13 6.067 108.93 5.664 26.70 26.80 0.37
3 0.3125 0.425 7.82 2.8865 7.354 238.05 6.322 125 5.525 60.72 62.61 3.11
4 0.32 0.4464 8.993 2.187 7.509 193.69 6.831 141.07 6.639 31.63 30.59 −3.29
5 0.3275 0.5321 8.155 2.241 7.046 189.26 5.431 122.32 7.475 59.72 58.92 −1.34
6 0.335 0.4786 7.485 2.349 7.817 224.74 5.558 143.75 6.361 38.45 37.70 −1.95
7 0.3425 0.4679 6.982 2.51 6.274 229.18 6.704 133.04 7.057 68.61 70.06 2.11
8 0.35 0.5 7.317 2.618 6.429 233.61 6.576 119.64 7.336 32.01 31.09 −2.87
9 0.3575 0.575 7.15 2.402 6.12 180.39 5.685 111.61 6.082 67.72 66.93 −1.17
10 0.365 0.4357 8.658 2.564 8.28 211.44 5.304 130.36 5.804 62.84 62.14 −1.11
11 0.3725 0.5107 6.815 2.456 6.737 175.95 5.176 138.39 5.943 45.37 45.69 0.71
12 0.38 0.4893 7.988 2.725 7.971 184.82 5.94 114.29 7.196 38.88 38.57 −0.80
13 0.3875 0.4571 8.825 2.1335 6.583 202.56 6.449 135.71 6.5 63.32 63.98 1.04
14 0.395 0.5536 8.323 2.671 7.663 215.87 5.813 106.25 6.918 30.92 31.33 1.33
15 0.4025 0.5643 8.49 2.295 8.126 220.31 5.049 127.68 6.221 74.63 71.56 -4.11

Table 4
The probability distributions and distribution parameters of probabilistic variables.

Probabilistic variable Unit Distribution Type Mean Standard deviation

Brake pressure p MPa Normal 0.50 0.008

Elastic modulus of disc E3 GPa Normal 125 2.0

Table 5
The parameter values of interval variables.

Interval probabilistic parameter Unit Lower bound Upper bound

Friction coefficient f / 0.30 0.40

Density of friction material ρ 2 kg/dm3 2.3845 2.6355
Density of disc ρ 3 kg/dm3 6.84 7.56
Elastic modulus of friction material E2 GPa 5.346 6.534

Table 6 expressed as 0.202 − ρ1 s1 d1 ≥ 0. s1 is the cross-sectional area of

The initial values and design ranges of the design parameters.
the back plate and its value is 3.78 × 10−3 m2 . The real part of the
Design parameter Unit Initial value Design range dominant unstable eigenvalue is taken as design objective and the
mass of back plate is taken as design constraint. The basic math-
ρ1 kg/dm3 7.82 7.04 ≤ ρ1 ≤ 8.60
μE1 GPa 207 186.3 ≤ μE1 ≤ 227.7
ematical model of the RBDO of the disc brake system with hybrid
μd1 mm 6.50 5.85 ≤ μd1 ≤ 7.15 uncertainties for squeal reduction can be expressed as

ness of the back plate are taken as the design variables, namely min α (x )
ρ  1 , μE 1 , μd 1
xD = [ρ1 , E1 , d1 ]T . Due to the manufacturing/measuring errors, the
s.t. P ( g( x ) ≥ 0 ) ≥ η
actual values of these design variables may deviate from the the-
g(x ) = 0.202 − ρ1 s1 d1
oretical values in engineering practice. In this case, E1 and d1
x = {xD , xP , xI }
are treated as probabilistic variables which are assumed to fol-
x D = [ ρ1 , E 1 , d 1 ] T (15)
low the normal distributions, and their means (marked as μE1
xP = [ p, E3 ]T
and μd1 ) are taken as design parameters and their standard de-
xI = [ f, ρ2 , ρ3 , E2 ]T
viations (marked as σE1 and σd1 ) are taken as the deterministic
7.04kg/dm3 ≤ ρ  1 ≤ 8.60kg/dm3
values. ρ 1 is treated as interval variable which can be expressed
186.3GPa ≤ μE1 ≤ 227.7GPa
as ρ1 = ρ1 [1 − γ , 1 + γ ], where ρ1 is the theoretical density of the
5.85mm ≤ μd1 ≤ 7.15mm
back plate which is taken as design parameter, and γ is the uncer-
tain level of interval variable ρ 1 .
The initial values and the design ranges of the design parame-
ters ρ1 , μE1 and μd1 are listed in Table 6, where the initial values where ρ1 = ρ1 [1 − γ , 1 + γ ], μE1 and μd1 are the means of E1 and
take the nominal values of ρ 1 , E1 and d1 , respectively. The stan- d1 , respectively; ρ1 , μE1 and μd1 are the design parameters; η is
dard deviations σE1 , σd1 and the uncertain level γ are considered the reliability index according to the design requirement; g(x ) =
as σE1 = 3.0GPa, σd1 = 0.1mm and γ = 0.05, respectively. ρ1 s1 d1 − 0.202 is the limit-state function; s1 is the cross-sectional
area of the back plate.
According to Section. 3, the design constraint P(g(x) ≥ 0) ≥ η
5.5. The optimization design based on reliability and confidence can be transformed to P (g(x ) ≥ 0 ) ≥ η, and the design objective
interval for a disc brake system with hybrid uncertainties min α (x ) can be converted to min C cl (α (x ))upper .
ρ  1 , μE 1 , μd ρ  1 , μE 1 , μd
1 1
For the optimization, the constraint condition is that the mass Therefore, the mathematical model of the optimization design
of the design component does not exceed 0.202 kg, which can be based on reliability and confidence interval for the disc brake
120 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122
Table 7
The values of C cl (α (x ))upper and P (g(x ) ≥ 0 ) at initial and
optimal values of the design parameters.
Function At initial values At optimal values
C cl (α (x ))upper 95.21 76.96
P ( g( x ) ≥ 0 ) 0.492 0.999
system with hybrid uncertainties can be approximated as
min C cl (α (x ))upper
ρ  1 , μE 1 , μd 1
s.t. P ( g( x ) ≥ 0 ) ≥ η
g(x ) = 0.202 − ρ1 s1 d1
x = {xD , xP , xI }
x D = [ ρ1 , E 1 , d 1 ] T
xP = [ p, E3 ]T
xI = [ f, ρ2 , ρ3 , E2 ]T
7.04kg/dm3 ≤ ρ  1 ≤ 8.60kg/dm3
186.3GPa ≤ μE1 ≤ 227.7GPa
5.85mm ≤ μd1 ≤ 7.15mm
Both the design confidence level cl and the design reliabil-
ity index η are assumed as 0.997. C cl (α (x ))upper and P (g(x ) ≥
0 ) are calculated by the Monte-Carlo method with 107 samples
(Ni = 103 , N j = 104 ) at the initial values of the design parameters
(ρ1 =7.82kg/dm3 , μE1 =207GPa and μd1 =6.50mm), and the results
are listed in Table 7.
By using of the nominal values of the investigated parameters
shown in Table 1, we can approximately calculate the initial value
of the real part of the dominant unstable eigenvalue. The calcu-
lated result is α0 = 63.72. From Table 7, it can be found that the
values of C cl (α (x ))upper is 95.21, which is much larger than the ini-
tial deterministic value and indicates that the stability of the disc
brake system may deteriorate in the course of using under uncer-
tainties with the initial values of the design parameters. Further-
more, it can be found that P (g(x ) ≥ 0 ) is just 0.492, which is much
less than the design requirement (η=0.997) and it indicates that
the constraint condition may be seriously destroyed while the un-
certainties of the brake system are taken into account. Therefore,
the performance of the brake system should be improved by the
optimization approaches.
The Genetic Algorithm [47] is an important evolutionary opti-
mization algorithm which has been widely applied in engineering
practices, and it is selected as the optimization approach in the im-
plementation of the proposed optimization. The parameter values
of the back plate are optimized by the Genetic Algorithm based
on the Monte-Carlo method, and the design parameters after opti-
mization are: ρ1 =7.04kg/dm3 , μE1 =227.70GPa and μd1 =5.88mm.
For comparison, the values of C cl (α (x ))upper and P (g(x ) ≥ 0 )
at the optimal values of the design parameters are also listed in
Table 7.
From Table 7, we can see that C cl (α (x ))upper is reduced from
95.21 to 76.96. Namely, it reduces about 19.17% after optimization.
So the stability of the brake system will be improved. Furthermore,
it can be found from Table 7 that the reliability of the constraint
at the initial design parameters violates the design requirement,
while the reliability at the optimal values satisfies the design re-
quirement. Therefore, the proposed approach can be considered as
an effective method for the optimization of the disc brake system
with hybrid uncertainties for squeal reduction.
The complex eigenvalues of the brake system with the optimal
values of design parameters are plotted in Fig. 10. The complex
eigenvalues of the brake system with the initial values of design
parameters are also plotted in Fig. 10. The other parameters are
all fixed at the nominal values when the design parameters take
the initial or the optimal values, and the results of the complex
100
Initial design
Optimal design
Real part
0
-100
0 2000 4000 6000 8000 10000 12000 14000 16000
Imaginary part /Hz
Fig. 10. The distribution of system complex eigenvalues of initial and optimal de-
(16) signs.
8
x 10
Initial design
Acceleration a /(mm/s )
5
2
Optimal design
0
-5
0.02 0.022 0.024 0.026 0.028 0.03
Time t /s
Fig. 11. The dynamic transient analysis results of initial and optimal designs.
eigenvalues are obtained by performing CAE based on the FE
model. It can be seen from Fig. 10 that the real part of the domi-
nant unstable eigenvalue of the brake system is reduced apparently
after the optimization. The real parts of the other complex eigen-
values are also reduced. Thus, the system stability is improved.
The dynamic transient analysis is also carried out to validate
the optimal design. The transient analysis results under the opti-
mal values of system parameters are shown in Fig. 11. For compari-
son, the transient analysis results under the initial values of system
parameters are plotted in Fig. 11 as well. The measurement point
of the acceleration is also laid on the centroid of a pad. From Fig.
11, it can be observed that the acceleration amplitude of the pad is
reduced significantly after optimization. Both the CEA results and
the transient analysis results reveal that the brake system becomes
more stable after optimization.
For further comparison, the system performances at the op-
timal values of design parameters obtained by the deterministic
model, the probabilistic model and the hybrid uncertain model are
shown in Table 8. When implementing the deterministic model,
the probabilistic variables in this study are treated as determin-
istic parameters whose values equate to the means of the prob-
abilistic variables, while the interval variables are also treated as
deterministic parameters whose values equate to the mid-values
of the intervals. When implementing the probabilistic model, the
interval variables are treated as deterministic parameters whose
values equate to the mid-values of the intervals. Namely, all the
uncertainties of the brake system are neglected in the determinis-
tic model, whereas the interval uncertainties are neglected in the
probabilistic model. In Table 8, the optimal values of the design pa-
rameters are obtained by the optimization of deterministic model,
the optimization of probabilistic model and the proposed optimiza-
tion, respectively. And the values of C cl (α (x ))upper and P (g(x ) ≥ 0 )
for the deterministic model and the probabilistic model are
 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122
Table 8
The comparison of the results among the deterministic model, the probabilistic model and the hybrid
uncertain model.
Type Optimal values of the design parameters
ρ  (kg/dm3 )
1 μE1 (GPa )
Initial value 7.82 207.00
Deterministic model 7.04 225.10
Probabilistic model 7.04 227.21
Hybrid uncertain model 7.04 227.70
calculated by the procedure of the proposed approach with the
design parameters (ρ1 =7.04kg/dm3 ,μE1 =225.10GPa,μd1 =7.15mm)
and (ρ1 =7.04kg/dm3 , μE1 =227.21GPa, μd1 =7.11mm), respectively.
From Table 8, we can find that the optimal values of
C cl (α (x ))upper obtained by the optimization of the deterministic
model, the probabilistic model and the proposed optimization are
all less than the initial value, which means that all these ap-
proaches can achieve the goal of optimization. On the other hand,
the reliability of the constraint P (g(x ) ≥ 0 ) at the optimal values
of design parameters obtained by the deterministic optimization is
much less than the design requirement (η=0.997). It indicates that
the constraint condition after optimization may be seriously vio-
lated if the uncertainties of the brake system are neglected. Fur-
thermore, P (g(x ) ≥ 0 ) at the optimal values of design parameters
obtained by the optimization of the probabilistic model is still fail-
ing to achieve the design target. Thus, if we treat the hybrid uncer-
tain model as the probabilistic model in which the uncertainties of
interval variables are neglected, the constraint condition after opti-
mization may be still violated. However, at the optimal values ob-
tained by the proposed optimization, P (g(x ) ≥ 0 ) is not less than
the design reliability index. Namely, the constraint condition af-
ter optimization is satisfied strictly. Based on the discussion men-
tioned above, the proposed optimization design based on reliability
and confidence interval can be considered as a potential method
for squeal reduction of the disc brake system with hybrid uncer-
tainties.
In Guan et al. [11], also provided an optimization method for
disc brake squeal reduction. However, there is a significant dif-
ference between Ref. [11] and this study. The biggest difference
is that Ref. [11] is based on the deterministic techniques which
do not consider the parameter uncertainties; whereas the uncer-
tain parameters with and without sufficient information are all
taken into account for investigation in this study. In addition,
the design constraint of the numerical example is limited to the
lightweight requirement of suppliers. Nevertheless, the main aim
of this study is to present an optimization approach consider-
ing hybrid uncertainties for brake squeal reduction. The proposed
approach has the capacity to take into account more other de-
sign constraints/objectives, more other uncertain variables or other
variations of uncertain variables according to the engineering re-
quirements. Moreover, the proposed approach considering hybrid
uncertainties is not only applicable to the optimization problem
of brake squeal reduction, but also applicable to the other engi-
neering problems with probabilistic and interval uncertainties. The
proposed optimization has important theoretical significances and
broad scope of application in practice.
6. Conclusions
In engineering applications, uncertainties with and without suf-
ficient information may exist simultaneously. For this case, a hy-
brid uncertain model with probabilistic and interval variables is
introduced to deal with the optimization problem of squeal reduc-
tion for a disc brake system in this paper.
To explore the optimal design of a disc brake system for squeal
reduction, a model of the optimization design based on reliabil-
121
C cl (α (x ))upper P ( g( x ) ≥ 0 )
μd1 (mm )
6.50 95.21 0.492
7.15 69.37 0.732
7.11 71.84 0.854
5.88 76.96 0.999
ity and confidence interval is constructed by using the methods of
RSM, CEA and hybrid uncertain analysis. In the proposed optimiza-
tion, the real part of the dominant unstable eigenvalue is selected
as the design objective while the mass of the design component
is adopted as the reliability constraints, both the design objective
and design constraint are interval probabilistic functions due to the
effects of hybrid uncertainties. For this case, the maximum of the
upper bound of confidence interval of design objective is selected
as the objective function, while the minimal value of the proba-
bilistic constraint is selected as the constraint function. The combi-
national algorithm of Genetic Algorithm and Monte-Carlo method
is employed to perform the optimization.
By introducing the proposed optimization for a commercial disc
brake system with hybrid uncertainties, the real part of the do-
main unstable eigenvalue is reduced effectively and the constraint
reliability satisfies the design requirement well at the optimal val-
ues of design parameters. The numerical results also show that if
the interval uncertainties or the hybrid uncertainties existing in
the disc brake system are neglected, the optimal results obtained
by the common optimizations may violate the constraint condi-
tions seriously. The proposed optimization design based on reli-
ability and confidence interval can be considered as a potential
method for squeal reduction of the disc brake system with hybrid
uncertainties.
Acknowledgments
The paper is supported by National Natural Science Foundation
of China (No. 11572121). The authors would also like to thank the
reviewers for their valuable suggestions.
Appendix A
α (x ) = α ( f, p, ρ1 , ρ2 , ρ3 , E1 , E2 , E3 , d1 )
= −1098.33678 − 822.07413 f +2.37734 p − 108.20223ρ1
−157.92349ρ2 − 9.78798ρ3 +5.54512E1 +16.72707E2
−1.84784E3 +428.07868d1 +444.63456 f p−72.15916 f ρ1
+21.12025 f ρ2 +117.77880 f ρ3 +2.35209 f E1 +8.99633 f E2
−2.53044 f E3 +110.65793 f d1 +131.47472 pρ1 − 64.92017 pρ2
−16.63370 pρ3 − 0.055478 pE1 +28.53294 pE2 +0.85694 pE3
−146.38686 pd1 +9.92251ρ1 ρ2 +14.87162ρ1 ρ3 +0.40092ρ1 E1
+0.85994ρ1 E2 − 0.52102ρ1 E3 +12.57852ρ1 d1 +21.43085ρ2 ρ3
−0.19325ρ2 E1 − 3.53690ρ2 E2 +0.34647ρ2 E3 − 30.81064ρ2 d1
−0.39980ρ3 E1 − 2.16935ρ3 E2 +0.55328ρ3 E3 − 25.48857ρ3 d1
+0.0018E1 E2 +0.026698E1 E3 − 0.74327E1 d1 − 0.16965E2 E3
+0.65233E2 d1 +0.72428E3 d1 − 570.09382 f 2 − 175.88387 p2
−9.49232ρ12 +40.10924ρ22 − 2.31084ρ32 − 0.011803E12
−0.15765E22 − 0.026902E32 − 13.91395d12
References
[1] Adnan A. Acoustics of friction. J Acoust Soc Am 2002;111(4):1525–48.
[2] Wagner UV, Hochlenert D, Hagedorn P. Minimal models for disk brake squeal.
J Sound Vib 2007;302(3):527–39.
[3] Yang S. Brake vibration and noise: reviews, comments, and proposals. Int J
Mater Prod Technol 1997;12(4-5):496–513.
122 H. Lü, D. Yu / Advances in Engineering Software 98 (2016) 112–122
[4] Nishiwaki M. Generalized theory of brake noise. In: Proceedings of the insti-
tution of mechanical engineers part D: journal of automobile engineering, 40;
1988. p. 499–513.
[5] Papinniemi A, Lai JCS, Zhao J, Loader L. Brake squeal: a literature review. Appl
Acoust 2002;63(4):391–400.
[6] Kinkaid NM, O’Reilly OM, Papadopoulos P. Automotive disc brake squeal. J
Sound Vib 2003;267(1):105–66.
[7] Ouyang H, Nack W, Yuan Y, Chen F. Numerical analysis of automotive disc
brake squeal: a review. Int J Vehicle Noise Vib 2005;1(3-4):207–30.
[8] Massi F, Baillet L, Giannini O, Sestieri A. Brake squeal: linear and nonlinear
numerical approaches. Mech Syst Signal Process 2007;21(6):2374–93.
[9] Akay A, Giannini O, Massi F, Sestieri A. Disc brake squeal characterization
through simplified test rigs. Mech Syst Signal Process 2009;23(23):2590–607.
[10] Ghazaly NM, Ahmed I, El-Sharkawy M. A review of automotive brake squeal
mechanisms. J Mech Des Vib 2014;1(1):5–9.
[11] Guan D, Su X, Zhang F. Sensitivity analysis of brake squeal tendency to sub-
structures’ modal parameters. J Sound Vib 2006;291(1):72–80.
[12] Spelsberg-Korspeter G. Structural optimization for the avoidance of self-excited
vibrations based on analytical models. J Sound Vib 2010;329(23):4829–40.
[13] Spelsberg-Korspeter G. Eigenvalue optimization against brake squeal:
symmetry, mathematical background and experiments. J Sound Vib
2012;331(19):4259–68.
[14] Lakkam S, Koetniyom S. Optimization of constrained layer damping for
strain energy minimization of vibrating pads. Songklanakarin J Sci Technol
2012;34(2):179–87.
[15] Shintani K, Azegami H. Shape optimization for suppressing brake squeal. Struct
Multidiscip Optim 2014;50(6):1127–35.
[16] Sun G, Li G, Gong Z, Cui X, Yang X, Li Q. Multiobjective robust opti-
mization method for drawbead design in sheet metal forming. Mater Des
2010;31(4):1917–29.
[17] Sun G, Li G, Zhou S, Li H, Hou S, Li Q. Crashworthiness design of ve-
hicle by using multiobjective robust optimization. Struct Multidiscip Optim
2011;44(1):99–110.
[18] Lagaros ND, Plevris V, Papadrakakis M. Neurocomputing strategies for
solving reliability-robust design optimization problems. Eng Comput
2010;27(27):819–40.
[19] Lee I, Choi KK, Gorsich D. Sensitivity analyses of FORM-based and DRM-based
performance measure approach (PMA) for reliability-based design optimiza-
tion (RBDO). Int J Numer Methods Eng 2010;82(82):26–46.
[20] Yu H, Gillot F, Ichchou M. Reliability based robust design optimization for
tuned mass damper in passive vibration control of deterministic/uncertain
structures. J Sound Vib 2013;332(9):2222–38.
[21] Chen Z, Qiu H, Gao L, Su L, Li P. An adaptive decoupling approach for reliabil-
ity-based design optimization. Comput Struct 2013;117(3):58–66.
[22] Grange P, Clair D, Baillet L, Fogli M. Brake squeal analysis by coupling spec-
tral linearization and modal identification methods. Mech Syst Signal Process
2009;23(8):2575–89.
[23] Sarrouy E, Dessombz O, Sinou JJ. Piecewise polynomial chaos expansion
with an application to brake squeal of a linear brake system. J Sound Vib
2013;332(3):577–94.
[24] Tison T, Heussaff A, Massa F, Turpin I, Nunes RF. Improvement in the
predictivity of squeal simulations: uncertainty and robustness. J Sound Vib
2014;333(333):3394–412.
[25] Lu H, Yu D. Brake squeal reduction of vehicle disc brake system with interval
parameters by uncertain optimization. J Sound Vib 2014;333(26):7313–25.
[26] Lu H, Yu D. Stability analysis and improvement of uncertain disc brake
system with random and interval parameters. J Vib Acoust Trans ASME
2015;137:051003.
[27] Nobari A, Ouyang H, Bannister P. Uncertainty quantification of squeal instabil-
ity via surrogate modelling. Mech Syst Signal Process 2015;60-61:887–908.
[28] Stefanou G. The stochastic finite element method: past, present and future.
Comput Methods Appl Mech Eng 2009;198(198):1031–51.
[29] Elishakoff I, Colombi P. Combination of probabilistic and convex models of un-
certainty when scarce knowledge is present on acoustic excitation parameters.
Comput Methods Appl Mech Eng 1993;104(2):187–209.
[30] Elishakoff I, Cai GQ, Starnes JH. Non-linear buckling of a column with initial
imperfection via stochastic and non-stochastic convex models. Int J Non Linear
Mech 1994;29(1):71–82.
[31] Zhu LP, Elishakoff I. Robust optimization for structures using non-probabilistic
convex method of set theory. Math Probl Eng 1996;2(2):143–63.
[32] Zhou MJ. A design optimization method using evidence theory. J Mech Des
2006;128(4):1153–61.
[33] Abubakar AR, Ouyang H. Complex eigenvalue analysis and dynamic tran-
sient analysis in predicting disc brake squeal. Int J Vehicle Noise Vib
2006;2(2):143–55.
[34] Liu P, Zheng H, Cai C, Wang YY, Lu C, Ang KH, Liu GR. Analysis of disc
brake squeal using the complex eigenvalue method. Appl Acoust
2007;68(6):603–15.
[35] Júnior MT, Gerges SNY, Jordan R. Analysis of brake squeal noise using the finite
element method: a parametric study. Appl Acoust 2008;69(5):147–62.
[36] Fang J, Gao Y, Sun G, Li Q. Multiobjective reliability-based optimization for
design of a vehicledoor. Finite Elem Anal Des 2013;67(5):13–21.
[37] Nouby M, Mathivanan D, Srinivasan K. A combined approach of complex eigen-
value analysis and design of experiments (DOE) to study disc brake squeal. Int
J Eng Sci Technol 2009;1(1):254–71.
[38] Baş D, İsmail HB. Modeling and optimization I: usability of response surface
methodology. J Food Eng 2007;78(3):836–45.
[39] Mayers RH, Montgomery DC. Response surface methodology, process and
product optimization using design experiment. New Jersey: John Wiley and
Sons; 2002.
[40] Lyes N, Sébastien B, Evelyne A. Prediction of random self friction-induced vi-
brations in uncertain dry friction systems using a multi-element generalized
polynomial chaos approach. J Vib Acoust Trans ASME 2012;134(4):1002–10.
[41] Kruse S, Hoffmann NP. On the robustness of instabilities in friction-induced
vibration. J Vib Acoust Trans ASME 2013;135:2627–37.
[42] Cao Q, Ouyang H, Friswell MI, Mottershead JE. Linear eigenvalue analysis of
the disc-brake squeal problem. Int J Numer Methods Eng 2004;61(9):1546–63.
[43] Papila M Accuracy of response surface approximations for weight equations
based on structural optimization. PhD thesis, University of Florida; 2001.
[44] Kaufman M, Balabanov V, Giunta AA, Grossman B, Mason WH, Burgee SL,
Haftka RT, Watson LT. Variable-complexity response surface approximations
for wing structural weight in HSCT design. Comput Mech 1996;18(2):112–26.
[45] Redhe M, Forsberg J, Jansson T, Marklund PO, Nilsson L. Using the response
surface methodology and the D-optimality criterion in crashworthiness related
problems. Struct Multidiscip Optim 2002;24(3):185–94.
[46] Fu J, Zhao Y, Wu Q. Optimising photoelectrocatalytic oxidation of fulvic acid
using response surface methodology. J Hazard Mater 2007;144(1-2):499–505.
[47] Goldberg DE. Genetic algorithms in search, optimization & machine learning.
New York: Addison-Wesley; 1989.