Journal of Sound and Vibration 338 (2015) 169–183

Contents lists available at ScienceDirect

Journal of Sound and Vibration

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

Statistics of complex eigenvalues in friction-induced vibration

Amir Nobari a, Huajiang Ouyang a,n, Paul Bannister b
a
School of Engineering, University of Liverpool, the Quadrangle, Brownlow Street, Liverpool, L69 3GH, United Kingdom
b
Jaguar Land Rover, Abbey Road, Whitley, Coventry, CV3 4LF, United Kingdom

a r t i c l e i n f o abstract

Article history: Self-excited vibrations appear in many mechanical systems with sliding contacts. There
Received 19 July 2014 are several mechanisms whereby friction can cause the self-excited vibration to become
Received in revised form unstable. Of these mechanisms, mode coupling is thought to be responsible for generating
5 October 2014
annoying high-frequency noise and vibration in brakes. Conventionally, in order to
Accepted 13 October 2014
identify whether a system is stable or not, complex eigenvalue analysis is performed.
Handling Editor: L.N. Virgin
Available online 28 November 2014 However, what has recently received much attention of researchers is the variability and
uncertainty of input variables in the stability analysis of self-excited vibrations. For this
purpose, a second-order perturbation method is extended and employed in the current
study. The moments of the output distribution along with its joint moment generating
function are used for quantifying the statistics of the complex eigenvalues. Moreover, the
eigen-derivatives required for the perturbation method are presented in a way that they
can deal with the asymmetry of the stiffness matrix and non-proportional damping. Since
the eigen-derivatives of such systems are complex-valued numbers, it is mathematically
more informative and convenient to derive the statistics of the eigenvalues in a complex
form, without decomposing them into two real-valued real and imaginary parts. Then, the
variance and pseudo-variance of the complex eigenvalues are used for determining the
statistics of the real and imaginary parts. The reliability and robustness of the system in
terms of stability can also be quantified by the approximated output distribution.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction

Friction in mechanical systems can cause irritating noise and vibration. The sliding velocity, normal load, friction
coefficient and stiffness of sliding components are key factors making contributions to the excitation mechanisms.
For example, in a simple “mass-on-belt” slider model, when the stiffness and relative velocity of the slider are fairly low,
the difference between the static and kinetic friction can result in a repetitive motion with sticking and slipping phases. This
motion is known as stick–slip [1]. However, instead of this discrete function, a number of smooth and non-smooth functions
may be assigned to the relation between the friction coefficient and sliding velocity [2]. On this premise, Mills [3] developed
a model for brake squeak instability. In this model, a decreasing function of the sliding velocity is considered for the
transition from static to kinetic friction.
Brake systems experience several types of noises and vibrations and a number of frictional mechanisms have been
proposed for the reasons behind these instabilities. By means of a lumped-mass brake model, North [4] showed even a

n
Corresponding author.
E-mail address: h.ouyang@liverpool.ac.uk (H. Ouyang).

http://dx.doi.org/10.1016/j.jsv.2014.10.017
0022-460X/& 2014 Elsevier Ltd. All rights reserved.
170 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183

constant friction coefficient can change stability of the system as a result of taking friction as a follower force. The idea of
follower forces was followed by Ouyang et al. in [5]. Spurr [6] developed a theory in which the normal load was a function of
the angle of inclination of a frictional slider in contact with a surface. It was illustrated that a reciprocating motion could be
formed due to the locking (spragging) of the slider to the surface even if the friction coefficient remained constant. This
mechanism is cited as sprag–slip in the literature. Kinkaid et al. [7] comprehensively reviewed several frictional mechanisms
causing the instabilities of brake systems.
More recently, it is believed that mode-coupling is the main root cause of unstable oscillations in many mechanical systems with
friction. Akay [8] cited a number of examples in which mode lock-in could generate self-exited vibrations. Perhaps the most
challenging example is automotive brake squeal due to its significant warranty cost to the manufacturers. Although there is no
general consensus on its root cause, most researchers believe that the reason behind brake squeal is mode-coupling. Hoffmann
et al. [9] demonstrated this mechanism in an undamped system with two degrees-of-freedom. It is shown that friction breaks the
symmetry of the stiffness matrix since it relates the normal force to the tangential direction. Through increasing the friction
coefficient, the imaginary parts of a pair of system eigenvalues get closer and at a unique point they merge. Simultaneously, the real
parts of the system eigenvalues bifurcate. This unique point is called a bifurcation point. After the bifurcation point, one of the real
parts of the eigenvalues becomes positive and causes the amplitude of vibration to grow exponentially.
The behaviour of friction-induced vibration about bifurcation points has been studied by a number of researchers.
Oestreich et al. [2] did a bifurcation analysis for stick–slip instability. Hetzler et al. [10] studied the bifurcation behaviour of a
simple slider whose friction coefficient was a function of the sliding velocity. Sinou et al. [11] investigated the nonlinear
behaviour of a system with sprag–slip instability using the centre manifold approach. The bifurcation behaviour of mode-
coupling instability under the influence of damping was discussed in [12,13]. Damping can cause the destabilization paradox
[14,15] in friction-induced vibration — having either “lowering effect” or “smoothing effect” on mode-coupling instability.
If two equally damped modes are coupled, damping makes the system more stable, i.e. “lowering effect”, as the real parts of
the system eigenvalues become smaller under its influence. On the other hand, damping can shift the eigenvalues towards
unstable region, i.e. “smoothing effect”, when the modes in coalescence are unequally damped. It is worth mentioning that
“lowering effect” can also occur for the unequally damped modes depending on the values of damping coefficients. This
phenomenon was studied in [12,13] for low-order models. Fritz et al. [16] investigated the impact of the destabilization
paradox on mode-coupling instability of a finite element brake model.
Based on the mode-coupling mechanism, the amplitude of vibration grows without bound in a linear system. However,
what is evident in practice is that the existence of nonlinearities in a structure may impede this growth after a number of
cycles. Subsequently, the reciprocating motion of the system becomes periodic and its trajectories tend to a closed curve
known as Limit Cycle. Depending on the central aim of a study, size of the problem and accessible computational facilities, a
linear or nonlinear approach may be taken toward friction-induced vibration. In [17], Ouyang et al. discussed the pros and
cons of the linear and nonlinear approaches. Although the computational efficiency of the linear stability analysis, i.e.
complex eigenvalue analysis, is appreciated, it is emphasised that the nonlinear approaches can be more informative.
Unfortunately, the nonlinear approaches are mostly performed by numerical integrations which are computationally very
expensive. Therefore, some researchers used different simplification and linearisation techniques for including the
contributions of nonlinearities in the linearised system. In order to predict the limit cycle of friction-induced instability,
the methods of Describing Functions [18], Centre Manifold [19] and Constrained Harmonic Balance [20] have been employed
in the literature. Interestingly, Coudeyras et al. [21,22] determined the limit cycle of nonlinear oscillations via Constrained
Harmonic Balance Method for a simplified brake system.
In addition to the effects of nonlinearities, two inseparable parts of any friction-induced problem are variability and
uncertainty [23]. In the literature, structural uncertainties are divided into two categories: reducible and irreducible
uncertainties [24]. Reducible uncertainties, also simply known as uncertainty, are mostly due to the lack of analysts'
knowledge. Friction, contact, thermal effects, humidity, wear and loading distributions are typical examples, and each one
introduces a degree of uncertainty to the problem. If more realistic models of these factors are incorporated in the problem,
more reliable results can subsequently be obtained. On the other hand, irreducible uncertainties, also called variability,
originate in the deviation of material properties from the nominal values, imperfection of component geometries and
dissimilarity of assemblies. As this group of uncertainties are inherent in the manufacturing processes, it is referred to as
irreducible uncertainty.
Uncertainty analysis of friction-induced vibration can be performed based on either a linear stability analysis or a
nonlinear transient analysis. While the main objective of the first approach is to compute stochastic eigenvalues of a system,
the second approach mostly aims at determining limit cycles of a nonlinear system in the stochastic case. For a system with
two-degrees-of-freedom, Nechak et al. [25] investigated the influence of uncertain friction coefficient on the system
eigenvalues. The same system was studied in [26] in order to determine both uncertain eigenvalues and stochastic limit
cycles of nonlinear oscillations. In both studies, uncertainty analyses were done based on the Polynomial Chaos approach.
Then, Sarrouy et al. [27] extended the Polynomial Chaos approach to the brake squeal problem. Computing stochastic
complex eigenvalues of a simplified brake system formed the central aim of [27].
Regarding the brake squeal problem, Tison et al. [28] demonstrated that the prediction of unstable modes would be
improved if the variability of the pad-to-disc contact interface was modelled by a random field. Oberst and Lai [29] took an
experimental approach to the statistical analysis of squeal instability. Culla and Massi [30] employed Monte Carlo simulation
for investigating contact instability under the influence of uncertainties, which leads to squeal noise.
 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183 171

In order to incorporate the influence of variability and uncertainty in friction-induced instabilities, probabilistic, non-
probabilistic, or mixed uncertainty techniques may be employed. The advantage of probabilistic approaches is that they are
usually more informative in terms of the statistics of outputs. However, their major drawback is that they can be
computationally expensive. For example, Monte Carlo simulation is known as the most practical, yet simplest, technique of
uncertainty analysis [31]. In practice, nevertheless, it is not always feasible to collect a large number of deterministic results
which are required by this approach. Therefore, researchers have been encouraged to use more efficient uncertainty
techniques such as the perturbation methods and polynomial chaos expansions.
The technique to be employed in this paper is the 2nd order perturbation method. Adhikari and Friswell [32] applied this
method for random matrix eigenvalue problems. As there was no sliding contact with friction in their examples, the stiffness matrix
remained symmetric. Moreover, the effect of non-proportional damping was not discussed in [32]. Consequently, the distributions
of the imaginary parts of eigenvalues, i.e. frequencies, were the point of interest in their work. However, this paper focuses on
friction-induced vibration with non-proportional damping. Friction makes the stiffness matrix asymmetric and introducing non-
proportional damping means that the problem should be solved in the state-space. What is important here is to obtain the
probabilistic distributions of the real parts of eigenvalues for evaluating the system stability. For this purpose, a few expressions are
derived for the statistics of the complex eigenvalues. The first moment of the output distribution is used for the mean value and the
joint moment generating function is employed for the calculation of the variance of the scatter of the complex eigenvalues.
The statistics of the real and imaginary parts is then calculated by the use of the derived expressions.

2. Stability analysis of friction-induced vibration with non-proportional damping

In general, there are two numerical approaches toward friction-induced problems: complex eigenvalue analysis and
transient analysis. As complex eigenvalue analysis deals with linear systems, it can only predict the onset of instability.
In fact, this approach finds eigenvalues of a system in the complex form, i.e. λj ¼ αj þ i ωj where αj and ωj are the real and the
imaginary parts of the jth eigenvalue, and then assesses the system stability by the sign of the real part. Based on the mode-
coupling mechanism, one or more real parts of the eigenvalues being positive indicate instability. A positive real part
enlarges the amplitude of oscillations without bound.
However, in reality, the amplitude of vibration often grows to some level and then tends to a steady state motion which is
known as Limit Cycle Oscillation (LCO). In general, the only way to safely deal with any form of nonlinearity is numerical
integration. Although there are some methods such as Harmonic Balance and Describing Functions, which are able to
linearise nonlinear equations of motion about a static equilibrium point, the main shortcoming of these approaches is that
there are a variety of nonlinearities that cannot be studied in this way. For example, the limit cycle of a system with a cubic
nonlinearity can adequately be determined through these techniques, while it is arduous to take the same approaches for a
system with a non-smooth nonlinearity.
As a result, in order to investigate LCO of friction-induced vibration, transient analysis is the most reliable way. For squeal
instability, Sinou et al. emphasised the significance of both complex eigenvalue analysis and transient analysis in [33,34].
For simplified finite element brake models, transient analysis was carried out along with complex eigenvalue analysis in
[33–35]. However, depending on the number of degrees-of-freedom of a finite element model, transient analysis can
sometimes become very expensive. In such cases, industry only relies on the results of complex eigenvalue analysis and
attempts to design a system so that it remains stable under various circumstances.
The equations of motion for an N degree-of-freedom system with a sliding contact and non-proportional damping can be
written as
€ þCuðtÞ
MuðtÞ _ þ KuðtÞ ¼ 0 (1)

where M, C and K are the mass, damping and asymmetric stiffness matrices, respectively. In order to perform the complex
eigenvalue analysis, two aspects of the above equation should be considered. First, the stiffness matrix is not symmetric;
the left and right eigenvectors are not the same. Secondly, damping of the system can be either proportional or non-
proportional. As a result, the state-space equations of motion should be used in place of the second-order equations. It is
worth mentioning that one may employ the second-order equations of motion directly [36], but it is mathematically more
convenient to use the state-space form:

Az_ ðtÞ þBzðtÞ ¼ 0 (2)

   

 C M K 0
_ T, A ¼
where z ¼ u; u and B ¼ .
M 0 0 M
To derive the eigenvalues and eigenvectors of Eq. (2), the complex eigenvalue analysis is carried out. In other words,
z ¼ Z eλt where λ represents the eigenvalues of the system. The solution consists of N pairs of eigenvalues which are
complex conjugates. Since the stiffness matrix is not symmetric, the associated eigenvectors are derived in the following
way:
   
λj A þB uj ¼ 0 ; vTj λj A þB ¼ 0 (3)
172 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183

where uj and vj are the right and left eigenvectors, respectively. Later on, the normalised eigenvectors will be used in the
equations. The normalisation means [37]:

vTj Auj ¼ 1: (4)

Up to this point, no uncertainties are considered in input variables. In fact, the above procedure is the general
deterministic approach toward friction-induced vibration. Now what is important to know is how variability and
uncertainty of the input variables can be propagated to the output space by means of an efficient statistical approach.
For this purpose, the second-order perturbation method is extended in this paper.

3. Uncertainty analysis via perturbation method

In the literature, structural uncertainties are classified into two groups: aleatory and epistemic [24]. Yet it is not always
possible to make a distinction between them. Aleatory uncertainty, also called irreducible uncertainty or variability, refers to
the random nature in the system properties, which originates from manufacturing processes. Variations of material
properties, component geometries and assemblies are the most significant examples of this type of uncertainty. Epistemic
uncertainty, known as reducible uncertainty, is mainly due to the lack of information. This type of uncertainty is called
reducible since future advances and/or investigations can provide new insights into the problems which are not fully
understood yet.
In spite of the extensive investigations that have been conducted so far, predicting the behaviour of friction, contact and
wear remains very complicated. In addition to the tribological interactions, the degree of uncertainty in a friction-induced
vibration problem is increased by the thermal effects, humidity and diverse loading cases. Variability also imposes a
considerable level of uncertainty to the problem. For example, in brake systems, the properties of friction materials may vary
from one brake to another significantly. Ignoring the effect of variability in the complex eigenvalue analysis then causes
underestimation or overestimation of the number of unstable modes. Therefore, in order to make more reliable predictions
of unstable modes, it is necessary to perform uncertainty analysis.
Uncertainty propagation may be carried out via probabilistic, non-probabilistic or mixed uncertainty techniques. Briefly
speaking, non-probabilistic approaches are not as informative as probabilistic ones. If one is interested in finding the ranges
within which the output variables fluctuate, non-probabilistic approaches will be the right choice. Of course, these
predictions can also be made by a fuzzy method. This method evaluates fuzzy membership functions of the outputs, which
are more informative than crisp intervals [38]. However, in most applications, it is well worth quantifying the probability
distributions of the results in order to evaluate the reliability and robustness of a system.
The most fundamental, yet practical, probabilistic method of uncertainty analysis is Monte Carlo simulation. The
conventional Monte Carlo simulation spreads a large number of samples over the design space and collects the results of the
sampled points by means of a mapping function. In fact, the mapping function is the relation between the outputs and input
variables. Then, the statistics of the outputs are quantified by the use of the output scatter. In the case of mode-coupling
instability, the mapping function is the complex eigenvalue analysis. Note there is no explicit expression for the relation
between the real or imaginary parts of eigenvalues and the input variables unless a system consists of a very limited number
of degrees-of-freedom. The complex eigenvalues and eigenvectors of every sampled point should be calculated in order to
obtain the scatter of output space. In this study, the conventional Monte Carlo simulation is used for the validation of the
proposed method. What is obvious though is that the computational workload of the Monte Carlo simulation is massive.
Specifically, in large-scale finite element models, even running a single complex eigenvalue analysis of a detailed brake
model may take one or two days. It is thus impractical to run a great number of such simulations.
Fortunately, the perturbation method can overcome the issue of computational costs. In contrast with Monte Carlo
simulation, this method approximates the statistics of the outputs in one run. In this method, the outputs, i.e. the complex
eigenvalues, are assumed to be a function of the random input variables. Then, this function is expanded about a point of the
input distributions, which has the most contribution to the joint probability distribution. Depending on the orders retained
in the perturbation series, the method is called the first- or second-order perturbation. Higher order terms are very small in
comparison and thus are usually neglected in practice.
If the joint probability distribution of the input variables is Gaussian, the expansion of the complex eigenvalues about the
mean value ξ can be written in a quadratic form as [32]:
       1    
λj θ  λj ξ þ dTλj ξ θ  ξ þ θ  ξ T Dλj ξ θ  ξ (5)
2
n  o

where θi represents ith random input variable, dλj is the gradient vector dλj ξ ¼ ∂λj ðθÞ=∂θk θ ¼ ξ and Dλj is the Hessian
n  o

k

matrix Dλj ξ ¼ ∂2 λj ðθÞ=∂θk ∂θl θ ¼ ξ .

kl
The eigen-derivatives reported in [32] are not directly applicable to the problem under the current study due to the
asymmetry of the stiffness matrix and the presence of non-proportional damping. Plaut and Huseyin [39] derived the
expressions of eigenvalue and eigenvector derivatives in non-self-adjoint systems. These derivatives can be re-written for
 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183 173

the eigenvalues and eigenvectors of Eq. (2) as:

 h      i
∂λj ðθÞ=∂θk ¼  vTj θ Gjk θ uj θ (6)
   
where Gjk θ ¼ ½ð∂BðθÞ=∂θk Þ þ λj θ ð∂AðθÞ=∂θk Þ. Moreover,

2     h∂2 BðθÞ  ∂2 AðθÞi  
∂ λj θ ∂θk ∂θl ¼ vTj θ ∂θ ∂θ þ λj θ ∂θ ∂θ uj θ þ . . .
k l k l
 
     
 
     

∂AðθÞ   ∂AðθÞ  
vTj θ ∂θ uj θ vTj θ Gjl θ uj θ þ vTj θ ∂θ uj θ vTj θ Gjk θ uj θ þ . . .
k l
     
            
      
N vTj θ Gjk θ ur θ vTr θ Gjl θ uj θ þ vTj θ Gjl θ ur θ vTr θ Gjk θ uj θ
∑   : (7)
r¼1 ðλj θ  λr ðθÞÞ
r aj

The quadratic form of the complex eigenvalues which are given in Eq. (5) cannot directly be used since θi is a random
variable. In other words, this equation is not deterministic. However, the statistics of the complex eigenvalues can be
obtained by the use of this quadratic form. For the convenience of readers, the theoretical background of this approach is
briefly discussed in the following section.

4. Statistics of complex eigenvalues

The main objective of this section is to predict the statistics of the complex eigenvalues without decomposing them into
the real and imaginary parts. As the eigen-derivatives expressed in (6) and (7) produce complex-valued numbers, the
moments and cumulants of the output distributions are also derived in a complex form.

4.1. Derivation of the mean value

In the theory of statistics, the mean value of a random variable is its first moment about the origin. Considering the fact
that the input variables are real-valued while λj ðθÞ is a complex function of the inputs, the mean value or the first moment of
the jth eigenvalue can be calculated by:
Z þ1 Z þ1
 
λ^ j ¼ E λj ðθÞ ¼ ⋯ λj ðθÞ pθ ðθÞdθ1 ⋯dθm (8)
1 1
 
where E λj ðθÞ and pθ ðθÞ represent the expected value of λj and the joint probability density function of θ, respectively. Vector
θ consists of m real-valued random input variables. Any joint probability density function can be substituted in place of pθ .
However, in many cases but the multivariate Gaussian distribution, it is not feasible to derive a closed-form expression for λ^ . j
In [40], the joint probability density function of an m-dimensional Gaussian distribution is given in a quadratic form:
 
1 T  
pθ ðθÞ ¼ ð2πÞ  m=2 jVj  1=2 exp  θ  ξ V  1 θ  ξ (9)
2

where V is the covariance matrix of the input variables θ.

It is more convenient to transform the variable of integration (8) to y ¼ θ  ξ:
Z þ1 Z þ1    
    1   1
λ^ j ¼ ð2πÞ  m=2 jVj  1=2 ... λj ξ þ dTλj ξ y þ yT Dλj ξ y exp  yT V  1 y dy1 . . . dym : (10)
1 1 2 2
Due to the fact that the covariance matrix is positive definite, it can be decomposed as V ¼ LLT , where L is a non-singular
lower triangle matrix [41]. By the use of the Jacobian matrix, the variable of integration (10) is transformed to z ¼ L  1 y:
Z þ1 Z þ1    
    1   1
λ^ j ¼ ð2πÞ  m=2 ... λj ξ þ dTλj ξ Lz þ zT LT Dλj ξ Lz exp  zT z dz1 . . . dzm : (11)
1 1 2 2
Since zT z ¼ ∑m 2
j ¼ 1 zj , the multiple integral given in (11) turns into the product of mRsingle variable
 integrals.
 pIndeed,
ffiffiffiffiffiffi that is
þ1 R þ1
one of the advantages of using the theory of quadratic forms. Considering  1 exp  ðx2 =2Þ dx ¼ 2π and  1
 
xexp ðx =2Þ dx ¼ 0, Eq. (11) becomes:
2

Z þ1 Z þ1    
  1 T T   1
λ^ j ¼ λj ξ þ ð2πÞ  m=2 ... z L Dλj ξ Lz exp  zT z dz1 . . . dzm (12)
1 1 2 2
 
The symmetry of the matrix LT Dλj ξ L brings about the same left and right eigenvectors. If Δ and Ψ respectively
 
represent the diagonal matrix of the eigenvalues and the matrix of the normalised eigenvectors of LT Dλj ξ L, pre- and post-
multiplying it by Ψ and Ψ results in:
T

Z þ1 Z þ1    
  1 T 1
λ^ j ¼ λj ξ þð2πÞ  m=2 ... w Δw exp  wT w dw1 . . . dwm (13)
1 1 2 2
174 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183

where w ¼ Ψ z. As matrix Δ is diagonal, wT Δw ¼ ∑m

i ¼ 1 Δi wi . Therefore, the mean value of the jth complex eigenvalues can
T 2

explicitly be determined as:

 
1 m
λ^ j ¼ λj ξ þ
∑ Δ: (14)
2i¼1 i
   
The eigenvalues of LT Dλj ξ L are equivalent to those of VDλj ξ [41]. Furthermore, the summation of eigenvalues of any
matrix is evidently equal to the trace of that matrix. Therefore, Eq. (14) can be re-written as
  1  

λ^ j ¼ λj ξ þ Trace VDλj ξ : (15)

2
It is worth mentioning that the expression derived for the mean value of the complex eigenvalues has the same form as
the one reported in [32] for a symmetric system without damping. However, Eq. (15) produces a complex number whose
real part is the mean value of the real parts of the jth complex eigenvalues and whose imaginary part is the mean value of
the imaginary parts.

4.2. Derivation of the variance

What is known as the variance of a complex random variable is a real-valued number which is calculated by
   h     i
Var λj θ ¼ E ðλj θ  λ^ j Þ ðλj θ  λ^ j Þn (16)

where “n” denotes conjugate of a complex number. Before starting to derive the expression of the variance, it is important to
re-consider the central aim of this study, which is the evaluation of the system stability under the influence of structural
uncertainties. In other words, the distribution of the real parts of the complex eigenvalues is the main point of interest in
this study.
If Σ αj and Σ ωj represent variances of the real and imaginary parts, and also Σ αj ωj and Σ ωj αj stand for the covariance
between the real and imaginary parts, Eq. (16) can be expressed with these terms as:
  

Var λj θ ¼ Σ αj þ Σ ωj (17)

due to the fact that Σ αj ωj ¼ Σ ωj αj .

Obviously, Eq. (17) is not sufficient to calculate the variance of the real and imaginary parts. However, what can resolve
this issue is the use of pseudo-variance:

2 
    
g λj θ ¼ E λj θ  λ^ j
Var : (18)

Similar to the variance, the pseudo-variance can also be related to the variances of the real and imaginary parts:
  

g λj θ ¼ Σ α  Σ ω þ i 2 Σ ω α :
Var (19)
j j j j

  
g λj θ is a complex number and it provides a way to determine the variance of
As seen, in contrast with the variance, Var
the real and imaginary parts along with Eq. (17).
Now, in order to derive an expression for the variance of the complex eigenvalues, the joint moment generating function
is used. In fact, the moment generating function is an alternative way of calculating the moments of a random variable [42].
  n 
The joint moment generating function of λj θ and λj θ can be written as
h   i
n 
M λj ; λn ðs1 ; s2 Þ ¼ E expðs1 λj θ þ s2 λj θ (20)
j

Using the definition of the expected value results in:

Z þ1  h i 
  n  T  T   1 h    i
M λj ;λn ðs1 ; s2 Þ ¼ exp s1 λj ξ þs2 λj ξ þ s1 dλj ξ þs2 dλnj ξ y  yT V  1  s1 Dλj ξ  s2 Dλn ξ y dy (21)
j
1 2 j

This integral is, in fact, the generalised version of Gaussian integral to an m-dimensional domain. Similarly to the
procedure followed for the mean value, the solution of this integral can also be found:
Z þ1    
1 1
exp  xT Ax þ Bx dx ¼ ð2π Þm=2 jAj  1=2 BA  1 B (22)
1 2 2
where A is a symmetric matrix.
According to (22), the joint moment generating function can explicitly be determined:
expðs1 λj ðξÞ þ s2 λj ðξÞÞ
n

M λj ;λn ðs1 ; s2 Þ ¼  ⋯


j
I  s1 Dλj ðξÞ  s2 Dλn ðξÞ
nh i h (23)
 io
j

T  T        
exp s1 dλj ξ þ s2 dλn ξ V ½I s1 Dλj ξ  s2 Dλn ξ   1 s1 dλj ξ þs2 dλn ξ :
j j j
 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183 175

In the theory of statistics, it is quite common to use the cumulant generating function rather than the moment
generating function. The cumulant generating function is the natural logarithm of the moment generating function and the
second cumulant of a distribution is variance. Hence, the variance of the complex eigenvalues is obtained via
   n o
Var λj θ ¼ ð∂2 =∂s1 ∂s2 Þln M λj ; λn ðs1 ; s2 Þ :
j s1 ¼ s2 ¼ 0
After differentiating ln M λj ; λn ðs1 ; s2 Þ and setting s1 and s2 to zero, the variance of the complex eigenvalues can be
j

expressed as
    

T    1  
Var λj θ ¼ dλj ξ Vdλn ξ þ Trace VDλj ξ VDλn ξ : (24)
j 2 j

Similarly, the pseudo-variance is obtained by

        

g λj θ ¼ dT ξ Vd ξ þ 1Trace ½VD ξ 2 :
Var (25)
λj λj λj
2
The pseudo-variance is the same with the variance of eigenvalues given in [32], yet the difference here is that it produces
complex-valued numbers. As discussed earlier, expressions (24) and (25) will be employed together for the calculation of
Σ αj and Σ ωj .

5. A lumped mass model

In the literature, friction-induced vibration is often investigated in systems with limited degrees-of-freedom. However,
Butlin and Woodhouse [43] discovered that very low-order models may not represent the problem well. It was shown that
more reliable results could be achieved when enough degrees-of-freedom were included in a reduced model and the
minimum number of degrees-of-freedom of a reduced brake model was four. Therefore, the model under this study consists
of four degrees-of-freedom. Ouyang [44] employed this model in order to demonstrate the technique of pole-assignment for
stabilizing friction-induced vibrations. This model is shown in Fig. 1. Here, the only difference is that there is contact
damping at the slider-belt interface in addition to the grounded damping in [44]. This lumped mass model may be imagined
as a simplified brake system in which the slider serves as the disc and m2 acts as the brake pad with the stiffness of k2, k3
and k4 in different directions. The angle of inclination for k3 is 451. The contact stiffness and damping at the slider-belt
interface are kc and cc , respectively. Masses m1 and m3 may be considered as representing the abutment and the pad
backplate. The stiffness of the abutment is k1 and the backplate is k5. Mass m1 has one degree-of-freedom along x- direction,
m2 is free to move in both x- and y- directions and m3 can only oscillate in y-direction. The normal force is n1 and the friction
force is f1. It is assumed that the belt velocity is constant and no stick–slip motion is experienced. In other words, the static
and kinetic friction coefficients are the same.

T
The matrices of mass, damping and stiffness corresponding to the vector u ¼ x1 ; y3 ; x2 ; y2 are:
2 3 2 3 2 3
m1 0 0 0 c1 0  c1 0 k1 þ k2 0 k2 0
6 7 6 7 6 7
6 0 m3 0 0 7 6 0 0 0 0 7 6 0 k4 þ k5 0 k4 7
M¼6 6 0
7; C ¼ 6
7 6
7; K ¼ 6
7 6
7
7 (26)
4 0 m 2 0 5 4  c1 0 c 1 μc c 5 4  k2 0 k 2 þ0:5k 3 μ k c  0:5k3 5
0 0 0 m2 0 0 0 c0 þ cc 0  k4 0:5k3 k4 þ0:5k3 þkc

where mi ¼ 1 kg (i ¼ 1; 2; 3), ki ¼ 100 N=m (i ¼ 1; 2; ⋯5), kc ¼ 2k1 , c1 ¼ c0 ¼ 0:5 Ns=m and cc ¼ 0:1 Ns=m. These values are
the same as those of in [44]. However, adding the contact damping to the model causes the value of the critical friction

Fig. 1. A lumped mass model.

176 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183

Table 1
The complex eigenvalues of the system with four different friction coefficients (μ).

Eigenvalue First pair Second pair Third pair Fourth pair

μ¼0.4000  0.00127 8.7580i  0.0626 712.1832i  0.5125716.7569i  0.2236 7 19.8421i

μ¼0.4205 0.00007 8.7963i  0.0632 712.1736i  0.5154716.7700i  0.22157 19.8201i
μ¼1.2165 0.10277 10.3224i  0.1395 711.5937i  0.7632 717.4289i 0.00007 18.8467i
μ¼1.3000 0.40977 10.8802i  0.4388 711.2679  1.0005717.6370i 0.2295 7 18.5560i

Table 2
Normalised sensitivities of the real parts of the eigenvalues.

Variable First pair Second pair Third pair Fourth pair

kc  9.7175  1.2743 0.0263 0.3497

k2  8.5165 0.0854 0.1006  0.2079
k3  1.6676  0.4114 0.0018 0.1203
k4 0.2253 0.9047  0.0812  0.0686
cc 1.6344 0.1566 0.0026 0.1648
μ  19.3073 0.1628 0.1077  0.1872

coefficient to become different from the one in [44]. The matrices given in (26) reveal how the friction force breaks the
symmetry of stiffness and damping matrices by relating the normal force to the tangential one.
Before talking about the uncertainty analysis of the system shown in Fig. 1, it is well worth looking at the results of the
deterministic approaches. As discussed earlier, the complex eigenvalue analysis is the typical deterministic approach for
determining stability of a system. For the example set of input parameters, the complex eigenvalue analysis is performed with
four different friction coefficients. The results are shown in Table 1. Although each input parameter makes a contribution to the
system stability, the friction coefficient usually plays the most significant role in low order models. This fact will be proved for
this system via the sensitivity analysis presented in Section 6. As seen in Table 1, the real part of the first eigenvalue becomes
positive when μ is larger than 0.4205. The same behaviour is observed for the fourth eigenvalue if μ 41.2165.
The effect of input parameters on the bifurcation behaviour of mode-coupling instability has extensively been discussed in
[12,13]. It is worth mentioning that damping can cause the destabilisation paradox [14,15] in friction-induced vibration.
In brief, the configuration and values of damping in a system can results in instability known as “smoothing effect”. This fact is
in contrast with the general expectation of damping which usually shifts the eigenvalues towards the stable region (“lowering
effect”). This phenomenon has fully been investigated in [12,13] for mode-coupling instability. However, it is out of the scope
of this study to investigate these relationships for the system shown in Fig. 1. What is important here is to find out in what
manner the variability and uncertainty of input parameters can be propagated to the output space cost-effectively.

6. The effect of variability and uncertainty

Imagine the system shown in Fig. 1 represents a very simplified brake model. Many experimental studies have shown
that the level of variability and uncertainty of friction material is very high. Not only the variability of material properties
(k2, k3 and k4), but also the friction coefficient (μ), contact stiffness (kc ) and contact damping (cc ) are major sources of
uncertainties in this problem. When the friction coefficient is 0.4, according to Table 1, all eigenvalues of the system are
stable. It would be very useful to find out whether the system would remain stable subjected to uncertainty.
Although these parameters are selected based on the fact whether they are major sources of uncertainties in brake
systems, one effective way of selecting the right uncertain input variables is to carry out a sensitivity analysis. In
particular, when a large-scale model is under study, including several variables can lead to a significant amount of
computational workloads. Thus, only those parameters to which the unstable modes are highly sensitive should be
included in the uncertainty analysis.
To find out how sensitive the real parts of the eigenvalues are with respect to the selected input variables, a sensitivity
analysis is done by means of Eq. (6). Meanwhile, one consideration bringing about more reliable results is to normalize the
sensitivities in order to
 avoid scaling issues. The mean values of the input variables and corresponding real parts are used for
this purpose, i.e. ð∂λj θ =λj ðξÞÞ=ð∂θk =ξk Þ. Table 2 lists the normalised sensitivities of the real parts of the eigenvalues.
As seen, the real part of the first eigenvalue which is in the vicinity of instability shows high sensitivity to the friction
coefficient. The contact stiffness and k2 take the second and third places. Now, it is aimed to evaluate the efficacy of the
second-order perturbation method for quantifying the statistics of the eigenvalues under the uncertainty and
variability of all selected variables. Thereafter, a further study is made by randomising only the two most active
variables, i.e. μ and kc .
 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183 177

Table 3
Mean values of the real parts of the complex eigenvalues.

Real part of the 1st pair Real part of the 2nd pair Real part of the 3rd pair Real part of the 4th pair

Perturbation  1:4447  10  3  63:0119  10  3  512:8562  10  3  222:6897  10  3

Monte Carlo  1:4564  10  3  62:9989  10  3  512:9682  10  3  222:5765  10  3

Table 4
Mean values of the imaginary parts of the complex eigenvalues.

Imaginary part of the 1st pair Imaginary part of the 2nd pair Imaginary part of the 3rd pair Imaginary part of the 4th pair

Perturbation 8.7516 12.1715 16.7552 19.8448

Monte Carlo 8.7515 12.1716 16.7553 19.8450

Table 5
Variances of the real parts of the complex eigenvalues.

Real part of the 1st pair Real part of the 2nd pair Real part of the 3rd pair Real part of the 4th pair

Perturbation 14:4840  10  6 77:3989  10  6 85:3810  10  6 126:1762  10  6

Monte Carlo 14:9092  10  6 76:3613  10  6 101:5279  10  6 139:0542  10  6

Table 6
Variances of the imaginary parts of the complex eigenvalues.

Imaginary part of the 1st pair Imaginary part of the 2nd pair Imaginary part of the 3rd pair Imaginary part of the 4th pair

Perturbation 20:6140  10  3 29:1216  10  3 97:0271  10  3 202:2144  10  3

Monte Carlo 20:6556  10  3 29:4715  10  3 97:5162  10  3 201:9559  10  3

6.1. Uncertainty quantification of all selected variables

To perform an uncertainty analysis, it is assumed that the deviation of k2, k3 and k4 from their nominal values is 5 per
cent and the deviation of kc and cc is 10 per cent. As μ can change significantly in a brake application, 15 per cent variation is
allocated to the friction coefficient.
In order to validate the proposed approach, the conventional Monte Carlo simulation is run for the verification of the
results. One-million samples are made via a multivariate Gaussian random generator. The mean value and variance of the
real and imaginary parts are calculated separately and compared with the outcomes of the expressions given in Section 4.
Tables 3–6 list the results of the perturbation method and Monte Carlo simulation. According to the results, the error in the
prediction of the mean values is practically zero. The maximum error in the approximations of the variances is 15 per cent.
However, due to the small values of the variances, the error seems fairly large.
The distinct advantage of the second-order perturbation method is that it produces the results in just one run while a
significant number of simulations should be used for determining the same results via a Monte Carlo simulation. In actual
applications, running Monte Carlo simulation for large models is almost impossible due to the time required for the
computation of the results. Instead, the second-order perturbation method suggests an efficient way of uncertainty analysis
with remarkable accuracy in the prediction of results.
Generally speaking, when the mapping function is nonlinear, the projection of a multivariate Gaussian variable to the
output space is no longer Gaussian. Depending on how nonlinear a function is, the output distribution can be skewed or its
kurtosis may be different from the normal distribution. In the theory of statistics, the skewness and kurtosis are the shape
descriptors for real-valued random variables. Although it is possible to quantify the higher moments of a complex random
variable, the interpretation and relation of these moments to its real and imaginary parts are not straightforward. Therefore,
the approximation of output distributions with Gaussian probability density function does not perfectly represent the
results. However, it is useful to portray the output profile via Gaussian probability density function in order to have a better
understanding of the distributions of the results. In the literature, it is also common to use the normal/Gaussian probability
density function for this purpose [32].
178 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183

Fig. 2. Contours of the real and imaginary parts of the eigenvalues.

Fig. 3. Contours of the real and imaginary parts of the third eigenvalues.
 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183 179

Fig. 4. Distributions of real parts of the eigenvalues.

A bivariate Gaussian distribution is considered for the real and imaginary parts of each eigenvalue. This assumption helps
to plot the elliptical contours of the scatter of the results. The ellipse equation is [40]:
!2 ! ! !2
xαj  ξαj xαj  ξαj xωj  ξωj xωj  ξωj
 2ρ þ ¼ Constant (27)
σ αj σ αj σ ωj σ ωj

where ξαj and ξωj are the mean values of the real and imaginary parts of the jth eigenvalue. σ αj and σ ωj are the associated
standard deviations. Parameter ρ is the correlation between the real and imaginary parts, i.e. ρ ¼ Σ αj ωj =σ αj σ ωj , which is
calculated by (17) and (19). Moreover, the least-squares fit of the real and imaginary parts of the complex eigenvalues, which
are produced by Monte Carlo simulation is determined. The results are shown in Fig. 2.
In Fig. 2, the predicted elliptical contours of the first and second eigenvalues are well-matched with the least-squares fits.
However, a noticeable discrepancy is found for the third and fourth eigenvalues. The reason behind is believed to be the
level of assumed variances in the input parameters. To support this statement, the variation of all input parameters is now
assumed to be one per cent. For this case, Fig. 3 shows the least-squares fit and the predicted elliptical contour of the real
part of the third mode. As the perturbation method estimates the outputs about the mean value of the inputs, a smaller
variation of the input variables would result in more accurate predictions. The variance of the real part of this mode is now
7.6670  10  7 via Monte Carlo simulation and 7.6349  10  7 using the perturbation method.
When fairly large variations are considered for the input variables, like the first set of the input parameters, the
nonlinearity of the mapping function causes the skewness and kurtosis of the outputs to deviate from those of the normal/
Gaussian distribution. The distributions of the real parts of the eigenvalues are plotted in Fig. 4.
The histograms show the results of the Monte Carlo simulation, while the red circle-dashed lines indicate the
approximated distributions via Gaussian probability density function. Likewise, the distributions of the imaginary parts
are shown in Fig. 5. As seen, the distributions of the real parts of the first and second eigenvalue are nearly Gaussian.
However, the real parts of the third and fourth eigenvalues are negatively and positively skewed, respectively. Tables 7 and 8
180 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183

list the skewness and kurtosis of the real and imaginary parts of all eigenvalues. It can be seen that all imaginary parts of the
eigenvalues are almost Gaussian. The reason is mainly due to the fact that the imaginary parts are not much influenced by
the friction coefficient. It is worth mentioning that the kurtosis of the normal distribution is equal to three, but in the theory
of statistics, it is also common to use the excess kurtosis, which is defined as kurtosis minus three. However, in this study,
the excess kurtosis is not used.

Fig. 5. Distributions of the imaginary parts of the eigenvalues.

Table 7
The skewness of the real and imaginary parts.

Eigenvalue The 1st pair The 2nd pair The 3rd pair The 4th pair

Real parts  0.030  0.278  0.965 0.706

Imaginary parts 0.049  0.290  0.014  0.017

Table 8
The kurtosis of the real and imaginary parts.

Eigenvalue The 1st pair The 2nd pair The 3rd pair The 4th pair

Real parts 3.208 3.055 8.391 6.119

Imaginary parts 3.005 3.200 3.024 2.989
 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183 181

The second important conclusion is that although all of the eigenvalues of the system were stable before the uncertainty
analysis, the variation of the input parameters causes the first eigenvalue of the system to exceed the stable region and
become positive for some samples. Once again, it is imagined that the system represents a simplified brake model. There is
no doubt that at the end of the production, well designed and manufactured brakes are still likely to experience friction-
induced instability during services because of inherent uncertainty and variability in them. The probability distributions of
outputs then help to evaluate how robust and reliable a design is in terms of noise and vibration. Due to this fact,
uncertainty analysis of brake systems has received intensive attention of car manufacturers recently.

6.2. Randomising the most active parameters (μ and kc )

As discussed earlier, the major problem in the uncertainty analysis of industrial-scale finite element models is the issue
of computational workload. Although the perturbation method can sufficiently overcome this issue, reducing the number of
uncertain variables can help further. If a sensitivity analysis reveals that some of the selected input variables are of low
activity in a particular model, excluding them from the uncertainty analysis will be a right decision. Even if there is a high
level of uncertainty about one variable, it can be excluded when its influence on a particular unstable mode is negligible. In
other words, the levels of both sensitivity and uncertainty should be considered when right uncertain input parameters are
selected.
To illustrate this point, it is worthwhile to re-consider the results of the sensitivity analysis. Table 2 shows that the real
part of the first eigenvalue which is in the vicinity of instability is highly sensitive to the friction coefficient (μ). Its sensitivity
to the contact stiffness (kc ) is also considerable. However, for example, the contact damping has a minor effect on this
particular mode. Although the damping coefficient associated with contact is always considered one of the major sources of
uncertainty, here it is not necessary to include this variable in the uncertainty analysis.
In order to get a better understanding of what is being discussed, the uncertainty analysis is re-performed in this section.
However, only the variations of μ and kc are considered this time. Since the first eigenvalue is at high risk to become
unstable, only the probability distribution of its real part is shown here. Fig. 6 displays the distribution of the real part of the
first eigenvalue.
Two interesting points may be concluded based on these results. First, when the parameters of low activity are set aside
from the analysis, the results are almost the same with less computational workload. The mean value and variance of the
output distribution for the real part of the first eigenvalue is  1.3743  10  3 and 13.8163  10  6 in Fig. 6, while they were
1.4564  10  3 and 14.9092  10  6 in fig. 4. Secondly, the number of samples in the conventional Monte Carlo simulation
is 10,000 this time. Consequently, the output histogram is not as smooth as the one presented in the first example. In fact,
this is another drawback of the conventional Monte Carlo simulation. Although the convergence in the values of the mean
and variance may occur when an enough number of samples are collected, a smoother distribution will be achieved when
the number of samples is significantly increased. Incidentally, even collecting 10,000 samples is not feasible in many
practical applications.

7. Reliability analysis

In general, a reliability analysis is performed for evaluating the robustness of a design. In fact, the reliability analysis
quantifies the failure probability of the design. The term “failure” is widely used in engineering for various undesired states
of a system. In this study, failure means that one of the system eigenvalues becomes unstable, i.e. the sign of the associated
real part turns positive. For brake analysts, it is very important to know what percentage of a brake design will ‘fail’ either

Fig. 6. Distribution of the real part of the 1st eigenvalue.

182 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183

Fig. 7. Reliability analysis.

for production variability or due to usage and aging effects. Then, the design modifications which are usually done to reduce
brake noises can be carried out for reducing the likelihood of unstable vibration. The failure probability is obtained as:


P ℱ ¼ P αj : αj 4 0 (28)

Among the four modes of the system shown in Fig. 1, only the first one exceeds the stable region under the influence of
the uncertainties. Fig. 7 shows the distribution of the real part of the first eigenvalue under the influence of uncertainty of all
selected variables.
In Fig. 7, the green triangles highlight the border of the unstable region. By the use of the results of Monte Carlo
simulation, the percentage of failure for the first mode is 35.7. If the output distribution is approximated by Gaussian
probability density function, the predicted probability of failure is calculated by
" #
αj  ξαj
Pℱ ¼ 1  Ψ (29)
σ αj
pffiffiffiffiffiffi R z  
where ΨðzÞ ¼ ð1= 2π Þ  1 exp  u2 =2 du and αj ¼ 0. Accordingly, the approximated failure probability of the first mode is
35.2 per cent. This percentage is calculated by the estimated mean value ξαj and standard deviation σ αj via the perturbation
method. Moreover, one may say that the failure probability is quite large for the system under this study. In fact, the design
point is deliberately set slightly below the critical friction coefficient in order to demonstrate how bad a design can be if the
variability and uncertainty of the inputs are not considered carefully.

8. Conclusions

Mode-coupling instability may occur in many mechanical systems with sliding contacts. According to this mechanism,
the real parts of the complex eigenvalues of a system become positive when certain system parameters vary and this
indicates unstable oscillations. The common deterministic approach is to perform complex eigenvalue analysis in order to
evaluate stability of the system. However, complex eigenvalue analysis cannot deal with the variability and uncertainty of
input variables, which are inherent in any systems with frictional contacts. Uncertainty quantification of the outputs is then
vital for the reliable stability analysis of such problems.
The statistics of the complex eigenvalues are studied in this paper with the application to friction-induced vibration
problems. In order to deal with non-proportional damping and the asymmetry of the stiffness matrix, the state-space
equations are used. The second-order perturbation method is then extended for incorporating the variability and
uncertainty of input variables. A few expressions for calculating the mean value, variance and pseudo-variance of the
complex eigenvalues are derived. The mean value of the eigenvalues is, in fact, a complex number which provides the mean
value of the real parts and the mean value of the imaginary parts. The variance and pseudo-variance of the complex
eigenvalues include information of the variance of the real parts, the variance of imaginary parts and the covariance of the
real and imaginary parts. The correlation between the real and imaginary parts of the eigenvalues can also be obtained by
these statistical measures.
The most important outcome of this study is that the distributions of the real and imaginary parts of the complex
eigenvalues can efficiently be approximated without decomposing them into two real-valued numbers (real and imaginary
parts). In this way, the results are produced in just one-run, while Monte Carlo simulation requires a large number of
samples to find the statistical measures of the outputs.
This study also allows the likelihood of unstable vibration to be predicted and then provides a very useful tool for design.
 A. Nobari et al. / Journal of Sound and Vibration 338 (2015) 169–183 183

Acknowledgements

This work is sponsored by Jaguar Land Rover.

References

[1] K. Popp, P. Stelter, Stick–slip vibrations and chaos, Philosophical Transactions of the Royal Society of London A 332 (1990) 89–105.
[2] M. Oestreich, N. Hinrichs, K. Popp, Bifurcation and stability analysis for a non-smooth friction oscillator, Archive of Applied Mechanics 66 (1996)
301–314.
[3] H.R. Mills, Brake Squeak, Technical Report 9000 B, Institution of Automobile Engineers, 1938.
[4] M.R. North, Disc brake squeal – A Theoretical Model, Technical report 1972/5, Motor Industry Research Association, Warwickshire, England (1972).
[5] H. Ouyang, J.E. Mottershead, D.J. Brookfield, S. James, M.P. Cartmell, A methodology for the determination of dynamic instabilities in a car disc brake,
International Journal of Vehicle Design 23 (2000) 241–262.
[6] R.T. Spurr, A theory of brake squeal, Proceeding of the Institution of Mechanical Engineers: Automobile Division 15 (1961) 33–52.
[7] N.M. Kinkaid, O.M. O'Reilly, P. Papadopoulos, Automotive disc brake squeal, Journal of Sound and Vibration 267 (2003) 105–166.
[8] A. Akay, Acoustics of friction, Journal of the Acoustical Society of America 111 (2002) 1525–1548.
[9] N. Hoffmann, M. Fischer, R. Allgaier, L. Gaul, A minimal model for studying properties of the mode-coupling type instability in friction induced
oscillations, Mechanics Research Communications 29 (2002) 197–205.
[10] H. Hetzler, D. Schwarzer, W. Seemann, Analytical investigation of steady-state stability and Hopf-bifurcations occurring in sliding friction oscillators
with application to low-frequency disc brake noise, Communication in Nonlinear Science and Numerical Simulation 12 (2007) 83–99.
[11] J.-J. Sinou, F. Thouverez, L. Jézéquel, Analysis of friction and instability by the centre manifold theory for a non-linear sprag–slip model, Journal of Sound
and Vibration 265 (2003) 527–559.
[12] N. Hoffmann, L. Gaul, Effects of damping on mode-coupling instability in friction induced oscillations, Journal of Applied Mathematics and Mechanics 83
(2003) 524–534.
[13] J.-J. Sinou, L. Jézéquel, Mode coupling instability in friction-induced vibrations and its dependency on system parameters including damping, European
Journal of Mechanics A/Solids 26 (2007) 106–122.
[14] O.N. Kirillov, Destabilization paradox, Doklady Physics 49 (2004) 239–245.
[15] O.N. Kirillov, A.O. Seyranian, The effect of small internal and external damping on the stability of distributed non-conservative systems, Journal of
Applied Mathematics and Mechanics 69 (2005) 529–552.
[16] G. Fritz, J.-J. Sinou, J.-M. Duffal, L. Jézéquel, Investigation of the relationship between damping and mode-coupling patterns in case of brake squeal,
Journal of Sound and Vibration 307 (2007) 591–609.
[17] H. Ouyang, W. Nack, Y. Yuan, F. Chen, Numerical analysis of automotive disc brake squeal: a review, International Journal of Vehicle Noise and Vibration 1
(2005) 207–231.
[18] A.F. D'Souza, A.H. Dweib, Self-excited vibrations induced by dry friction, part 2: Stability and limit-cycle analysis, Journal of Sound and Vibration 137
(1990) 177–190.
[19] J.-J. Sinou, F. Thouverez, L. Jézéquel, Methods to reduce non-linear mechanical systems for instability computation, Archives of Computational Methods
in Engineering 11 (2004) 257–344.
[20] B. Hervé, J.-J. Sinou, H. Mahé, L. Jézéquel, Extension of the destabilization paradox to limit cycle amplitudes for a nonlinear self-excited system subject
to gyroscopic and circulatory actions, Journal of Sound and Vibration 323 (2009) 944–973.
[21] N. Coudeyras, J.-J. Sinou, S. Nacivet, A new treatment for predicting the self-excited vibrations of nonlinear systems with frictional interfaces: The
Constrained Harmonic Balance Method, with application to disc brake squeal, Journal of Sound and Vibration 319 (2009) 1175–1199.
[22] N. Coudeyras, S. Nacivet, J.-J. Sinou, Periodic and quasi-periodic solutions for multi-instabilities involved in brake squeal, Journal of Sound and Vibration
328 (2009) 520–540.
[23] T. Butlin, J. Woodhouse, Friction-induced vibration: Quantifying sensitivity and uncertainty, Journal of Sound and Vibration 329 (2010) 509–526.
[24] B. Möller, M. Beer, Engineering computation under uncertainty – Capabilities of non-traditional models, Computers and Structures 86 (2008)
1024–1041.
[25] L. Nechak, S. Berger, E. Aubry, Non-intrusive generalised polynomial chaos for the robust stability analysis of uncertain nonlinear dynamic friction
systems, Journal of Sound and Vibration 332 (2013) 1204–1215.
[26] E. Sarrouy, O. Dessombz, J.-J. Sinou, Stochastic study of a non-linear self-excited system with friction, European Journal of Mechanics A/Solids 40 (2013)
1–10.
[27] E. Sarrouy, O. Dessombz, J.-J. Sinou, Piecewise polynomial chaos expansion with an application to brake squeal of a linear brake system, Journal of
Sound and Vibration 332 (2013) 577–594.
[28] T. Tison, A. Heussaff, F. Massa, I. Turpin, R.F. Nunes, Improvement in the predictivity of squeal simulations: Uncertainty and robustness, Journal of
Sound and Vibration 333 (2014) 3394–3412.
[29] S. Oberst, J.C.S. Lai, Statistical analysis of brake squeal noise, Journal of Sound and Vibration 330 (2011) 2978–2994.
[30] A. Culla, F. Massi, Uncertainty model for contact instability prediction, Journal of the Acoustical Society of America 126 (2009) 1111–1119.
[31] C.A. Schenk, G.I. Schuëller, Uncertainty Assessment of Large Finite Element Systems, Springer, Berlin Heidelberg, 2005.
[32] S. Adhikari, M.I. Friswell, Random matrix eigenvalue problems in structural dynamics, International Journal for Numerical Methods in Engineering 69
(2007) 562–591.
[33] J.-J. Sinou, Transient non-linear dynamic analysis of automotive disc brake squeal - On the need to consider both stability and non-linear analysis,
Mechanics Research Communications 37 (2010) 96–105.
[34] J.-J. Sinou, A. Loyer, O. Chiello, G. Mogenier, X. Lorang, F. Cocheteux, S. Bellaj, A global strategy based on experiments and simulations for squeal
prediction on industrial railway brakes, Journal of Sound and Vibration 332 (2013) 5068–5085.
[35] K. Soobbarayen, S. Besset, J.-J. Sinou, Noise and vibration for a self-excited mechanical system with friction, Applied Acoustics 74 (2013) 1191–1204.
[36] S. Adhikari, M. Friswell, Eigenderivative analysis of asymmetric non-conservative systems, International Journal for Numerical Methods in Engineering
51 (2001) 709–733.
[37] D.B. Murthy, R.T. Haftka, Derivatives of eigenvalues and eigenvectors of a general complex matrix, International Journal for Numerical Methods in
Engineering 26 (1988) 293–311.
[38] S. Adhikari, H. Haddad Khodaparast, A spectral approach for fuzzy uncertainty propagation in finite element analysis, Fuzzy Sets and Systems 243
(2014) 1–24.
[39] R.H. Plaut, K. Huseyin, Derivatives of eigenvalues and eigenvectors in non-self-adjoint systems, AIAA Journal 11 (1973) 250–251.
[40] N.L. Johnson, S. Kotz, Distributions in Statistics: Continuous multivariate distributions, John Wiley & Sons Inc., New York, 1972.
[41] N.L. Johnson, S. Kotz, Distributions in Statistics: Continuous univariate distributions-2, John Wiley & Sons Inc., New York, 1970.
[42] N.L. Johnson, S. Kotz, Distributions in Statistics: Discrete distributions, John Wiley & Sons Inc., New York, 1969.
[43] T. Butlin, J. Woodhouse, Friction-induced vibration: Should low-order model be believed, Journal of Sound of Vibration 328 (2009) 92–108.
[44] H. Ouyang, Pole assignment of friction-induced vibration for stabilisation through state-feedback control, Journal of Sound and Vibration 329 (2010)
1985–1991.