Mechanical Systems and Signal Processing 139 (2020) 106615

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

Traction-based multi-scale nonlinear dynamic modeling

of bolted joints: Formulation, application, and trends
in micro-scale interface evolution
Nidish Narayanaa Balaji a, Wei Chen b, Matthew R.W. Brake a,⇑
a
Department of Mechanical Engineering, Rice University, Houston, TX 77005, United States
b
School of Aerospace Engineering and Applied Mechanics, Tongji University, Shanghai 200092, PR China

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

Article history: A new framework for modeling the dynamics of bolted structures is proposed that consid-
Received 15 July 2019 ers the nonlinear interfacial modeling of bolted structures using multi-scale traction-based
Received in revised form 29 October 2019 contact constitutive laws implemented through Zero-Thickness Elements (ZTE). Using
Accepted 3 January 2020
rough contact theory, it is possible to establish fundamental constitutive relationships with
Available online 23 January 2020
parameters estimated from micro-scale surface scans. Such a model is employed in the
framework for a three bolt lap-joint benchmark (the so-called ‘‘Brake-Reuß-Beam”).
Keywords:
Since the characterization of the interface is conducted in a full-field manner on top of a
Frictional systems
Rough contact
finite element mesh, the framework is also demonstrated to be applicable for conducting
Zero thickness elements full-field micro-scale interface evolution studies. Preliminary studies are conducted to
Surface characterization establish correlations of local changes in relevant roughness parameters with predicted
Bolted joints local tractions and dissipation fluxes.
Multi-scale modeling Ó 2020 Elsevier Ltd. All rights reserved.

1. Introduction

Developing a definitive understanding of the dynamical behavior of jointed systems can be significantly enabling for
design engineering [1]. The interfacial phenomena in a jointed connection can be highly nontrivial, and physical insight
in a fundamental level is of utmost importance for predictive modeling. Dealing primarily with bolted joints, the current
work focuses on modeling such systems in a physically consistent manner, placing emphasis on assessing the predictive
capabilities of the developed model.
Just the static loading condition of a bolted joint has been demonstrated to be highly variable due to uncertainties asso-
ciated with the torque-tension relationship [2–5]. Furthermore, even for a given level of bolt tension, the lack of consensus in
a reasonably accurate contact model complicates the prediction of the pre-stress configuration (see discussions between
compliant and ‘‘hard” contact, such as [6]).
Static considerations aside, it has been experimentally demonstrated that bolted interfaces experience very rich kine-
matic phenomena in dynamic excitation [7–9], which makes it necessary for any predictive approach to have sufficient fide-
lity in interfacial representation. Moreover, the contact interface is governed by highly nonlinear mechanics, ranging from
plastic flow and damage [10,11] to asperity fractures and adhesion [12,13]. However, accounting for all of these would result

⇑ Corresponding author.
E-mail addresses: [email protected] (N.N. Balaji), [email protected] (W. Chen), [email protected] (M.R.W. Brake).

https://doi.org/10.1016/j.ymssp.2020.106615
0888-3270/Ó 2020 Elsevier Ltd. All rights reserved.
2 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

in a model that would be nearly infeasible for practical deployment. It is thus of relevance to determine the most significant
underlying physics that can provide sufficiently accurate predictions while still retaining low computational complexity.
From a modeling standpoint, since the study of pre-stress and interfacial friction are geared toward different objectives, it
is sometimes thought that it is admissible to decouple the static simulation and the dynamic simulation [14,15]. In such
studies, the conventional approach is to first conduct a static simulation for the prestress state and then the dynamic anal-
yses are applied in the sense of perturbations about this state. However, as will be shown in the current work, this approach
has a significant disadvantage in that it defeats the purpose of applying physically derived nonlinear contact models, espe-
cially those pertaining to normal contact.
In order to model the contact behavior, it is common practice among phenomenological models to consider the normal
and the tangential behavior separately or to only allow for one way coupling (i.e., the tangential behavior does not affect the
normal behavior, but at the same time is dependent on it) [16,17]. From a mathematical standpoint, the contact problem,
more formally referred to as the Signorini problem [18], is formulated as a variational inequality (see [19] for details). How-
ever, the inequality problem can be expressed easily only for infinitely smooth surfaces, while most real surfaces have rough-
ness, modeled usually as randomly distributed micro-scale asperities on the surface [20]. Models of the response of such
surfaces may be achieved either by creating an exact computational representation of the interface by modeling each asper-
ity and solving the Signorini problem here or by developing compliant models accounting for the statistical distribution of
asperities [21]. Owing to the magnitudes of length-scales involved (asperities in the order of lm and components in the
order of milli- and/or centi-meters), the former approach tends often to be computationally prohibitive and thus has very
limited applicability in a system-level framework. The latter approach, is computationally cheaper, while at the same time
capturing the effects of the surface asperities in a statistically averaged sense. There exist many such formulations for the
normal contact, for instance the linear penalty stiffness approach as in [22,23] and the nonlinear penalty models in [24].
Some of these studies have also been used to account for plastic flow in the asperity level [25], leading to considerably
improved dissipative compliant models.
For the tangential behavior, physical observations of ‘‘stick” and ‘‘slip” have traditionally been used to derive independent
phenomenological models (see [26] for a review of some of the earliest efforts). These are usually modeled as discrete sliders
with a finite elastic limit, formally known as Prandtl-Ishlinskiı̆ models [27] (consisting of plays [28] and/or stops or Jenkins
elements [29]), which form the basis of some of the most popular elastic dry friction models in use today [14,30,31]. There
have been several reconciliations between the parameters involved here and surface roughness (see, for instance, [22,23,32–
35]). Infinite distributions of these discrete elements lead to the formulation of the Iwan models [36] (first proposed for plas-
ticity), which have gained significantly in popularity since the four-parameter Iwan model was proposed in [37]. However,
the drawbacks with applying the latter approach in the current work is twofold: first, the non-uniform and unsteady pres-
sure distribution in the interface (see [7–9], for some experimental studies) during dynamical operation makes them inap-
plicable for the test problems currently considered; and second, reconciling the involved parameters with the surface
roughness is not as direct as for the former.

1.1. Overview of modeling approach

The current work proposes a physically congruent structural dynamics modeling framework that attempts to tackle each
of the above issues. Experimentally, following the procedure in [5], bolt tensions are applied in a controlled manner by mon-
itoring the axial strains developed in the bolts so that the bolt loading conditions are specified without having to resort to
approximated torque-tension relationships.
The interface is represented using Zero-Thickness Elements [38] (ZTEs) so that the model is capable of representing local
interfacial kinematics sufficiently well. Adjustments for meso-scale irregularities in the topology of the surface are employed
so that machining imperfections (such as waviness) may be accounted for. A compliant rough contact model is formulated
using parameters estimated from surface scans for modeling the interfacial mechanics. The model involves the macro-scale
model, meso-scale imperfections, and the micro-scale asperities in one-way relationships, i.e., the lower length-scales influ-
ence the higher ones (micro- and meso- influencing the macro-), while the vise-versa is not explicitly captured in the model
(but explored in a post hoc fashion). Interfacial scans are conducted before and after dynamical testing in order to attempt at
establishing statistical correlations between predicated interfacial tractions and dissipation fluxes to changes in surface
roughness parameters. The purpose is to establish surface scanning as a possible independent verification of the developed
contact model(s).
A coupled modal quasi-static approach is formulated in the spirit of the QSMA approach in [15,39]. The main motivation
for this is in order to for it to be possible to employ fully non-linear contact models without making assumptions beyond the
ones necessary for QSMA. It was observed in [40] that the prestress analysis preceding the dynamical analysis influences the
nature of the response in very undesirable ways. The coupled QSMA is formulated so as to not encounter such issues in
practice.
It must be noted here that the above approach does not model modal interactions and still assumes that the modal forcing
may be taken to remain approximately constant throughout the operational regime (as in QSMA). If a more detailed under-
standing of the steady state dynamical behavior of the structure is sought, one must opt for time-domain (shooting, direct
integration, etc.) or frequency domain (harmonic balance) methods, which are both several orders of magnitude more
expensive than the current approach.
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 3

The experimental dynamic characterization in this work involves the identification of effective nonlinear modal charac-
teristics from ring down data extracted from impact hammer tests using a fairly recent approach [41,42]. Since the nonlinear
modal modeling approach predicts these quantities directly, a physically meaningful comparison may be made between
model prediction and experimental observations.
In what follows, section 2 describes the interface representation (section 2.1) and modeling (section 2.2) followed in the
current work. Section 3 describes the benchmark structure (section 3.1) and interfacial parameter identification (section 3.2).
The results of the benchmark investigation for the dynamic tests the surface roughness evolution studies are presented in
section 4. And Section 5 provides key discussions, conclusions, and the outlook the authors take at the end of the study.

2. Interface modeling

In order to model the interface, two things must be established: how the interface is represented, and how its mechanical
behavior (response to kinematic phenomena) is defined. Zero-Thickness Elements (ZTE), occurring first in [38], are used to
represent the contact interface. The implementation using traction-based compliant models [14,43] and super-convergent-
nodal recasts is described in section 2.1. The 3D nonlinear contact model used is adapted from [21,32] for the normal and
from [44–46] for the tangential contact relationships. The surface friction concept has been adapted and slightly modified
from [47], where it was applied for a phenomenological model. Classical results, as some of these may be, have, as far as
the authors’ knowledge, not been applied to model bolted joints in the current spirit. All of the constitutive laws are formu-
lated using empirical asperity distributions in each discretized element on the interface, through which element-wise
traction-displacement laws are established (see section 2.2). Here, the word traction is used to refer to the surface traction
vector (see Cauchy’s stress theorem [48]). No assumptions are made with regards to the actual distribution of the asperities.
It must be noted that ZTE’s may be used only for small displacement applications, which is the expected case for tightly
bolted joints such as the Brake-Reuß Beam benchmark (BRB) [49]. This gives the ZTE approach a significant computational
advantage since contact-searches are completely avoided in the formulation. For applications with larger displacements,
there exist more general approaches (see, for instance [50]) involving contact and closest point searches (usually resulting
in computationally burdensome implementations).
As a natural requirement of element-based modeling, it is necessary to describe the continuum behavior of each element
to relative distortions. Thus, traction-based nonlinear constitutive relationships need to be employed in such a way as to
consistently calculate nodal forces by assembling traction-integrals across the elements in the interface. One of the main
advantages of element-traction approaches is that the model may be consistently employed for arbitrarily shaped meshes
in interfaces. Note that the same may not be said about node-to-node contact models [51], which are limited in their appli-
cability for irregular or biased mesh scenarios. This requirement is particularly germane for bolted interfaces, where the
presence of the holes makes the mesh highly irregular.
Cheap computational implementation of these models may be achieved by evaluating the nonlinear traction vectors at a set
of quadrature locations (extracted using interpolation matrices computed offline) in each element and using appropriate
weights to integrate them with the corresponding shape functions (achieved with an integration matrix, also computed offline)
to obtain the integrated effects (forces and Jacobians/stiffnesses) as nodal forces and stiffnesses to be used by the solver. This
offers an implementation that is cheap, yet consistent with the weak form that finite element theory is formulated to solve.

2.1. Zero-thickness elements

Fig. 1 depicts a schematic view of an interfacial element with numbered nodes. In the figure, nodes 1-2-3-4 form an ele-
ment that belongs to one body; 5–6-7–8 belong to another; and the ZTE establishes an ‘‘interfacial element” as a combination
of these nodes. Being defined as interfacial elements, the nodes in each element are taken to be fixed through the operation. In
other words, the mesh (nodal connectivity of the elements) is taken to be fixed throughout the simulation. As noted already,
this is a simplification that can be made only for small tangential relative displacement applications (in the ðn
gÞ plane). A
similar but slightly different implementation involves thin-layer elements [52–54], which enforce a small but finite thickness
in each element in order to regularize rigid contacts. The zero thickness elements on the other hand, are better suited for com-
pliant contact implementations. The tangential plane natural coordinate axes are n; g and the normal direction is n. For flat
interfaces, relative kinematics in the n
g plane are completely governed by frictional laws while those in the n direction
are governed by normal contact laws. For non-flat interfaces an additional step involving coordinate transformations will
be necessary in order to derive corresponding contact relationships since there might not be a nodal Degree-of-Freedom
(DoF) that is aligned along the normal direction. In the current study however, since only small relative displacements and
initial gaps are investigated, the z-displacement of the nodes is always taken to be along the n-direction.
Although for the quadrilateral-on-quadrilateral element case (as shown above), the resulting interfacial element is an 8-
noded element, only the planar shape functions, in the n
g plane, are used for the kinematic description in each element.
Denoting u as some displacement and using the subscripts TOP and BOT to denote the top and bottom counterparts respec-
tively, the finite element interpolation may be represented in terms of the shape functions N i (using Einstein summation
notation) as
4 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 1. Schematic of a Zero-Thickness Element (ZTE).

uTOP ðn; gÞ ¼ Ni ðn; gÞui i ¼ 1; 2; 3; 4

ð1Þ
uBOT ðn; gÞ ¼ NI ðn; gÞu I
I ¼ 5; 6; 7; 8:

Eq. (1) represents the interpolation for linear elements, i.e., with each node having the displacements as the Degrees-Of-
Freedom (DOFs). The same approach can easily be extended to higher order elements using Hermite shape functions (see
[55], for instance, where quadratic elements were used). Using relative coordinates Du ¼ uTOP
uBOT , and defining
Dui ¼ ui
uIðiÞ by denoting by i and IðiÞ the top and corresponding bottom nodes, it is possible to represent the kinematics
consistently using just the relative DOFs. A convenient framework for expressing the traction non-linearities can be estab-
lished via
Duðn; gÞ ¼ Ni ðn; gÞDui ; ð2Þ
which is based on just the local relative displacements (see section 2.2 for one particular form). For a structure governed by
linear elasticity away from the interface, the corresponding weak form term for a particular displacement-traction pair
becomes,
R R
W int ¼ CTOP wTOP ðn; gÞtTOP ðn; gÞdC þ CBOT wBOT ðn; gÞt BOT ðn; gÞdC
R
¼ C ðwTOP ðn; gÞ
wBOT ðn; gÞÞtðn; gÞdC since tTOP ð:Þ ¼
tBOT ð:Þ ¼ t ð:Þ ð3Þ
R
C ½N i ðn; gÞ t ðn; gÞdC:
T
¼ fDwi gT

Here, wðn; gÞ represents the weight functions; fwi g are nodal weight function values, interpolated using the same shape
functions N i ðn; gÞ (Galerkin projection); and tTOP ; tBOT represent the traction fields on the top and bottom faces respectively.
The area domains CTOP and CBOT , taken from the areas of the top and bottom surfaces of the interface, are assumed to have
opposite normal directions. This allows for the recast of the sum as an integral of the traction (say, tTOP ) over just a single
reference domain C (taken to be CTOP here). Using the variational principle on the weak form, the nodal forces and Jacobians
are given by
n o R
@W int
¼ fF i g ¼ C fNi ðn; gÞgT t ðn; gÞdC
@ Dwi
h i R   ð4Þ
@F i
@ Duj
¼ C fNi ðn; gÞgT @@t
Du
Nj ðn; gÞ dC:

Since there are 3 DOFs per node, the 3 tractions, in general, depend on all 3 DOFs. Numerically, the above integration may be
implemented using numerical quadrature, provided that the tractions and their derivatives are known in the quadrature
locations. This may be achieved by first evaluating the relative displacements at the quadrature locations, computing the
constitutive laws here, and then conducting the weighted summation to integrate and obtain the nodal quantities. The
two operations may be computed offline and stored in the form of matrices Q and T. Suppose there are N n nodes and N e
 
elements with N q quadrature points per element, Q will be of dimension N e N q  ðN n Þ and T will be of dimension
ðN n Þ  ðN e N n Þ. The former is built using just the shape functions evaluated at the quadrature locations while the latter
involves the shape functions multiplied by the mapping Jacobi determinant and the quadrature weights for each point.
The relevant quantities are thus computed via,
fDugqp ¼ Q fDugn
fF gn ¼ Tftgqp ð5Þ
h i h i
@ fF gn @ ft g
@ fug
¼ T @ fDugqp Q T ;
n qp
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 5

where the subscripts qp and n denote quadrature-point and nodal quantities respectively, with the former being of size N e N q
and the latter being of size N n . Since the matrices may be computed offline, the exact computation of the consistent quan-
tities may thus be achieved in an efficient manner by storing these matrices. Different types of elements may be accommo-
dated by considering these during the construction of these matrices, making the approach extensible even to higher order
elements in the same fashion.

2.2. Contact model

Since the ZTE formulation requires traction constitutive laws, standard phenomenological models are interpreted in a
force-per-unit-area manner localized to each element in order to be applied. This is in contrast to classical phenomenological
models, which relate forces to displacements, and to constitutive models, which relate stresses to strains.

2.2.1. Normal contact

It is first assumed that the rough surface may be modeled as a set of ellipsoids of uniform effective radii (planar axes) b
(estimated by the mean from measurements) located z distance away from the nominal axis of each surface. The assumption
of uniform effective radii is commonly encountered in rough contact literature, starting from [21]. In [22] the average effec-
tive radius is used to derive analytical relationships thus establishing force-displacement relationships in each sphere-pair as
an integral contact formulation. Fig. 2a shows a schematic of the contact between two bodies with circular profiles. Denoting
the distance between the centers in the undeformed and deformed configurations by r0 and r respectively, the contact area
(a) and load (f)-displacement (w ¼ r
r0 ) relationship in the elastic regime is expressed using the Hertzian solution [56] for
this case by
a ¼ pbw with w ¼ r
r 0
pffiffiffi
f ¼ 23 1
Em2 bw3=2 ;
ð Þ ð6Þ
 
1
where; b ¼ R1x þ R1y þ R1x þ R1y :
1 1 2 2

Here, Rqj is the radius of the ellipsoid in the q direction (x; y) on surface j (1; 2). Although formulations accounting for possible
generic orthotropy can be found in the literature [57], the axes are assumed to coincide with the coordinate/measurement x
and y directions in the current study.
Further, even though multiple contact scenarios may be considered (e.g., [56]), the current work only uses Eq. (6), para-
metrized by average effective radii. To calculate the average traction over an entire element/segment, an asperity peak dis-
tribution function /ðzÞ is defined for each segment as in [21]. For a segment with N asperities, the number of asperities with
heights between z and z þ Dz is given by
nðzÞ ¼ N/ðzÞDz: ð7Þ
Over the range of the asperity heights z 2 ½0; 1Þ, the individual contributions of each asperity must be summed up over all of
the asperities (in contact) to obtain the behavior of the whole segment. Applying summations to the quantities in eq. (6)
yields the sums
X X
Au nðzk Þpbwk ¼ N/ðzk Þpbwk Dzk ; and
k2K k2K
X pffiffiffi 2=3 X pffiffiffi 2=3 ð8Þ
Fu nðzk Þ 23 E
bwk ¼ N/ðzk Þ 23 E
bwk Dzk :
ð1
m2 Þ ð1
m2 Þ
k2K k2K

Here the set K is used to denote the set of asperities with heights zk that are in contact. Denoting the imposed relative normal
displacement with dn , the reference is defined in such a way that dn ¼ 0 for the case with the tallest peaks just touching. The
surface interference d and the deflection wk for the asperity with height zk are given by
d ¼ zp
dn
ð9Þ
wk ¼ zk
d:

In the limit Dzk ! 0, the sums in eq. (8) may be transformed to the corresponding Riemann integrals. Furthermore, since only
the asperities with peak heights greater than or equal to d are involved in the contact, the set K may be described fully by
limiting the integrals to ½d; 1Þ. Therefore, the integral relationships are
6 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 2. Normal contact development for rough surfaces: (a) Spherical asperity-on-asperity contact; (b) Contact of rough surfaces.

R1
A ¼ Npb d ðz
dÞ/ðzÞdz ¼ NpbE½z
djz > d
   
ð10Þ
¼ NpbE z
zp
dn jz > zp
dn
pffiffiffi R 1 pffiffiffi h i
F ¼ N 23 1
Em2 b d ðz
dÞ3=2 /ðzÞdz ¼ N 23 1
Em2 bE ðz
dÞ3=2 jz > d
ð Þ ð Þ
pffiffiffi h  3=2  i ð11Þ
¼ N 23 1
Em2 bE z
zp
dn jz > zp
dn :
ð Þ

In the final form of the expression, the term E½ g  denotes the expectation statistic of the term within the square brackets
given that z, the asperity heights, are distributed according to the density function /ðzÞ, and g ð:Þ is some function of z. Math-
ematically, this is either evaluated through the integral
Z 1
E½ g  ¼ g ðzÞ/ðzÞdz; ð12Þ
0

or estimated empirically. A commonly encountered estimator for the expectation statistic is the algebraic average of the
function g ðzÞ evaluated over a finite sample population generated from the distribution function given as

b 1X Ns
E ½g ¼ g ðzi Þ; ð13Þ
N s i¼1

with N s being the finite sample size. Unless explicitly stated otherwise, the average estimator will be referred to as the
empirical estimator henceforth in the current paper.
The effective normal traction t hn over an element (or segment) is defined as the force divided by the total area Ae (as
opposed to the real contact area in eq. (10)),
pffiffiffi h
F 2 E b   3=2  i
thn ¼ e ¼ N e E z
zp
dn jz > zp
dn : ð14Þ
A 3 ð1
m Þ A
2

The superscript h is used to denote that this is derived from the Hertzian solution for elastic contact.
In the current work, the tractions are integrated over interfacial elements, which means that Ae is the area of the finite
element under consideration. This thus requires that the characterization of the roughness of an interface is done on an
element-by-element or segment-by-segment fashion so that the local variations in the surface can be modeled
appropriately.
For a practical implementation, there are three possible ways of evaluating the traction in eq. (14): (1) Constructing
empirical distributions of /ðzÞ using surface scans and evaluating the integrals numerically for each displacement value;
(2) identifying features in /ðzÞ in order to evaluate the statistics using standard results (classical rough contact approaches
include formulations for the normal and exponential distribution families); and (3) constructing a table of thn for different d
values and using a curve fit to come up with an approximate model. Of the three, the first is the most accurate, but compu-
tationally very expensive for dynamic simulations, and the second is usually not generic enough. The third method would
involve conducting simulations at different levels of normal interference (d) and obtaining corresponding tractions. Although
these simulations themselves may be expensive, they need to be conducted just once, since the developed table may be used
to obtain a simpler representation of the macro-response after establishing a suitable curve fit. Studies such as [58] and its
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 7

references develop such approximate models for several standard distribution functions. The current work uses the third
approach and is described in detail in section 2.3.
From a computational stand-point, the normal traction is expressed as
(
thn ðun Þ t hn > 0 ðcontact Þ
t n ðun Þ ¼ ð15Þ
0 otherwise ðseparationÞ:

Note that since the contact model employed here is based on the Hertzian solution, it is implicitly elastic and thus there is no
hysteretic energy dissipation captured in the normal contact. Moreover, as can be seen by the contact condition in eq. (15),
the effects of adhesion are not considered in the current model. A common way of modeling adhesion in normal contact is to
allow for some finite negative normal traction before complete separation [59]. See section 3.2 for some discussions on this
for the current case.

2.2.2. Tangential Contact

Although there exist fully nonlinear rough contact relationships for frictional tangential contact (such as those employed
in [25,60]), the current work uses a simplified elastic dry friction element for this. The ‘‘stuck” regime is characterized by
traction that are linearly proportional to the relative displacements while the ‘‘slipped” regime is a constant force along
the stuck prediction direction. The formulation is similar to [47], except for the traction-based adaptation and the fact that
the stiffnesses are dependent on the normal relative displacement. When fully stuck, the traction developed is given by

tstuck
t ¼ kt ut : ð16Þ

Here, the subscript t refers to quantities relevant to the tangential kinematics with tstuck ; kt ,
and ut denoting the tangential
t
stuck traction, traction-stiffness, and tangential relative displacement respectively. The hysteretic modification on this is
done by introducing a ‘‘stick limit condition” using Coulomb’s law and by rewriting the equation in the rate-form. The limit
condition is usually taken as the magnitude of tangential forces not exceeding the normal traction scaled by the coefficient of
friction (reminiscent of Coulomb’s original law). Thus, the rate-dependent hysteretic modification of this is
(
dt t kt dut
þ du
dkt dun
ut stuck
¼ ds n ds
ð17Þ
ds 0 slipped=separated;
where s is some state-evolution variable (such as time) and un is used to denote the relative normal displacement. Using an
explicit Euler scheme for integrating this between successive points, one can replace the rate-dependence to explicit (hys-
teretic) state dependence. This yields a functional form that is similar (but generalized) to previous path dependent friction
implementations (such as the popular elastic dry friction/Jenkins element in [61]), and is represented as
8
>
> 0 separated
>
>
>
>
<
tt ¼ kt ðut
ut0 Þ þ du
dkt
ðun
un0 Þut þ tt0 stuck ð18Þ
>
> n
>
>
>
>
:
t0t slipped:
The subscripts 0 are used to denote the historical quantities, or the quantities at the ‘‘previous” s-step. The ‘‘Coulomb cone”
idea is used to obtain the post-slip behavior. Slippage is defined as the condition where the total magnitude of the planar
tractions is equal to the stick limit. Since slippage cannot introduce rotational moments in the system, it is observed that
the planar orientation of the traction vector is preserved across stick–slip state transitions. Consequently, a ‘‘stick-
prediction” direction is obtained by norming the traction predicted for the stuck state in eq. (18) to give
8 9 8 x;stuck 9 8 x;stuck 9
< bt sp >
> > > >t >
x
= 1 < tt = 1 < t =
¼ r
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
   ¼ : ð19Þ
>
: by > ; x;stuck
2
y;stuck
2: > y;stuck >
; jjt t jj >
stuck
: y;stuck >
;
t sp tt þ tt tt tt

This completes the formulation, allowing the tangential tractions to be expressed as

8
>
> 0 t n ¼ 0 ðseparationÞ
>
>
>
>
<
q  dkq
t qt ¼ kt uqt
uqt0 þ dutn ðun
un0 Þuqt þ t qt0 jjtstuck jj < lt n ðstickÞ ð20Þ
>
>
t
>
>
>
>
:
ltnbt qsp otherwise ðslipÞ:
The symbol q is used to denote either of the planar coordinates x; y and that the tangential laws for the two directions are
identical in form. Classical marching-based stick–slip estimation techniques [61] could be followed for estimating the Jaco-
bians for a Harmonic balance solver, while specifying appropriate ‘‘initial states” t t0 and ut0 will be sufficient for a quasi-static
8 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

solver (see section 3.1.2 for the current implementation). These quantities are only updated at points of direction change in
the hysteretic cycle since the response is taken to be monotonic between the direction reversal points. At the point of direc-
tion change, all frictional elements are instantaneously stuck, and this information is transmitted by saving the displace-
ments and tractions and accounting for them in the following steps using the above path/history variables. However, in
order to stay consistent with the fact that the normal contact is always in the elastic regime, the normal contact law is
not implemented in this manner.
In studies such as [44,46], a relationship is established between the linearized stiffnesses in the normal and tangential
directions in terms of a constant v (taken to be 2:0 here) and the Poisson’s ratio m. There have been different studies formu-
lating the exact form [16,21,44,46,62–66], and the current work, uses the tangential compliance estimate as in [44,46], which
proposes the tangential force law per asperity as

T qt ¼ ð2
2mvÞðE1þmÞ bq  dq
¼ ð2
m4E b  dq ;
Þð1þmÞ q ð21Þ
 
1
where; bq ¼ R
1
q þ q
R
1
:
1 2

In the above, scripts q denote x or y components, bq is the effective asperity radius from the two surfaces, and dq is the
imposed tangential relative displacement along the q component. As already noted, the orthotropy axes for the interface
are assumed to coincide with the coordinate x; y axes. Carrying out the integrals and dividing by the region area Ae (as in
eq. (14)), the elastic tangential traction is,

tqt ¼ uqt ð2
m4E b Pðz > dn Þ

Þð1þmÞ t
4E   
¼ uqt b P z > zp
d n : ð22Þ
ð2
mÞð1 þ mÞ t
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
q
kt

The tangential stiffness is thus only a function of the normal relative displacement, and is not an explicit function of the tan-
gential relative displacement. Here the subscripts n are used for d to emphasize the fact that it denotes the normal interfer-
ence zp
dn . This is the only difference in formulation in comparison to [47], wherein the tangential stiffnesses were all
assumed to be constant.
  
The statistic occurring above (P z > zp
dn ) is the complementary cumulative density function (ccdf) of z, which may
be evaluated either using the exact integral
Z 1
Pðz > fÞ ¼ F Z ðfÞ ¼ /ðzÞdz; ð23Þ
f

or with a finite sample empirical estimator using the count statistic

XNs
b ðz > fÞ ¼ Fb Z ðfÞ ¼ 1
P 1Z >f : ð24Þ
Ns i¼1 i

Similar to the estimator for the expectation in Eq. (13), this may be evaluated for a finite sample and then fitted to an appro-
priate curve that may be used for the dynamical model.

2.3. Parametric interface characterization

For the traction model of section 2.2, the most relevant quantities that have to be determined from interfacial scans are
the asperity peak heights and asperity radii along each direction. Additionally, ‘‘meso-scale” features must also be extracted
so that flatness imperfections in the topology of the interface (such as from machining or warpage) can be accounted for.
First, a bi-linear plane (of the form z ¼ a0 þ a1 x þ a2 y þ a3 xy) is fitted to the raw data to obtain the approximate under-
lying surface. This is then normalized with respect to the tallest point over the whole interface such that the tallest point
becomes zero and the rest of the surface is a positive distance away. This normalization is carried out since it is assumed
that prestress is the only mechanism through which asperities interact, and for zero prestress, no asperity is undergoing
deformation. This reorients the raw data, so that it is in a convenient form to be incorporated into the contact model. A point
is said to be separated from its counterpart on the other surface if the normal relative displacement at that point is lesser
than the value of the above plane evaluated there. A sample of the raw data and its processed form for a single element
are shown in Fig. 3a and b.
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 9

Fig. 3. Processing interfacial scan data: (a) a 3D view of raw data over a finite element; (b) Data upon surface bi-linear fitting; (c) kernel-interpolated fits of
asperity peaks to circles.

Following this, the asperities, assumed to be manifested as local peaks in the raw data, are analyzed. In [67], the authors
consider a few popular methods for numerically evaluating the relevant roughness parameters from a surface scan. Drawing
inspiration from the three-point-peak (3PP) method, a kernel-integrated approach is followed here. Local kernels of three
data points are fitted to a quadratic polynomial and analyzed using the coefficients if a peak exists within the kernel. After
establishing conditionals to ensure that the same peak is not counted more than once, the kernel is moved across the data in
each scan direction to obtain estimates of corresponding peak properties.
The local quadratic polynomials, expressed in the form z ¼ b0 þ b1 x þ b2 x2 (with x being the position variable and bi being
the coefficients), allow the peak location, height and its circular radius to be estimated by,

b1
xpeak ¼
2b 2

b2
zpeak ¼ b0
4b12 ð25Þ
Rpeak ¼
2b12 :

Once this one-dimensional scan is conducted in each direction, peak heights and radii are obtained for the asperities as
viewed along each scan direction. Fig. 3c depicts the results of a portion of a single line from a single element. It can be
observed that the implementation identifies all of the peaks satisfactorily. A more accurate implementation might conduct
this on the full 2D data, fitting ellipsoids from contours. That would allow for the detection of arbitrarily oriented orthotropy
axes if they exist.
10 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

There are two challenges for the implementation of the traction in Eq. (14) and the traction-stiffness in Eq. (22). First, it is
not always trivial to estimate the exact number of asperities in an interface since the ‘‘peaks” are more wavy than sharp (ob-
serve aspect ratios in Fig. 3c). Second, it will be very expensive to save complete non-parametric models of the interfacial
asperities for each element in a mesh. Even if it is possible to accommodate such fidelity computationally, the storage over-
head this will introduce is prohibitive. While addressing the former forms the rest of the current subsection, the latter will be
dealt with in section 2.4.
For the first issue, an asperity-packing argument is made in order to obtain estimates of the number of asperities on a
given element. Each asperity, on a two-dimensional snapshot, is idealized in Fig. 4a. Representing the radius, height and
effective width with R; z, and w respectively, the relationship for the effective width, which is the projection of the asperity
onto the plane of the interface, is
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
w¼ zð2R
zÞ: ð26Þ

Since the radii are estimated for the two scanning directions as bX and bY in the current study, the projection of each asperity
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
is idealized to be an ellipse with axes wX ¼ zð2bX
zÞ and wY ¼ zð2bY
zÞ. In order to get an estimate of the number of
asperities within a region, it is hypothesized that each peak is adjacent to a valley and the geometry of the valleys are iden-
tical to that of the peaks. This results in a uniform square packing as in Fig. 4b, with packing efficiency 0:5.
Thus, the number of elements in a given region is estimated by the fraction of full ellipses that may be packed into the
element area multiplied by the packing efficiency. This yields the estimate

b ¼g Ae
N ; ð27Þ
pwX wY
where g is used to denote the packing efficiency. Although there are ways of accounting for the gaps that will be present in
the packing efficiency, this will have to be done on an element-by-element basis since the gap distribution on an arbitrary
polygon is not trivial. Packing of ellipses in a generic triangle or a quadrilateral is by itself a fairly challenging problem, but
setting g ¼ 0:5 is assumed to be nominally valid for the current work. The interested reader is directed to [68,69] and similar
works for more details.

2.4. Parametric interface response characterization

Coming to the second issue (from the previous subsection) related to the accurate evaluation of the statistics (the integral
expressions), previous works such as [21,22,58] developed results making specific assumptions about the distribution func-
tion /ðzÞ (such as positing it to be an exponential distribution). In the current study however, the local contact response is
simulated by evaluating the estimators in Eqs. (13) and (24) using bootstrapped peak heights from the region of interest,
followed by fitting the response with an appropriate function. There have been several earlier studies following a similar
idea, such as [60], where an exponential function and a smoothened hysteretic model are fitted to the normal and tangential
traction laws respectively, and [25] where nonlinear piece-wise power law relationships are employed for both laws.
The colored lines in Fig. 5 show the empirically determined normal traction and x-direction tangential traction-stiffness
as the normal displacement is varied for a single element in the interface (note that the normal displacement denotes the
global interference of the two surfaces, with negative values corresponding to contact separation). The first large kink in each
curve denotes the height of the second asperity, i.e., before that, only a single asperity was in contact. The 95% confidence
intervals for the characteristics calculated using 30 bootstrapped samples are depicted for each case (see colored-dashed
lines), from which it is apparent that the maximal uncertainty is in the low displacement regime where the number of
engaged asperities is low.
Inspecting the shape of the curves, smooth functions are fitted to the data to aid the implementation. For the normal trac-
tion, a power law with a ‘‘Gaussian dip” term is used, giving the form,
 2 !
1 log un þ d
log tn ¼ a þ b log un þ c exp
; ð28Þ
2 e

with a; b; c; d; e being the fitting parameters. For the tangential traction-stiffness, a hyperbolic tangent function is employed
to fit the smoothed step-like behavior of the ccdf. This is expressed as,

log kt ¼ a þ b tanh ðdðlog un
cÞÞ: ð29Þ

Here, a; b; c; d are fitting parameters (different from the ones used in Eq. (28)). It must be noted that simplifications of the
former with a simple power law and the latter with a piece-wise constant function were avoided since the operational
regime of the models are around the region where the trends seem to show the dips and transitions respectively. The fits
for the reference element may be observed in Fig. 5 as black lines with dash patterns.
Once these fits are conducted for the asperities in each element of an interface, the parameters may be transmitted across
for system-level simulations. Now that the models are established, the only unknown that will have to be determined is the
coefficient of friction l. Although there are studies that estimate l from rough contact parameters (such as the ones used in
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 11

Fig. 4. (a) 2D geometry and (b) Spatial distribution of asperities in the interface.

x
Fig. 5. Statistical response characterization and fits. The left and right vertical axes are used for tn and kt respectively. The 95% bootstrap confidence limits
are highlighted using dashed lines for each case.

[25]), these are not used in the current formulation and the value is thought to be a constant lying somewhere between 0 and
1 (from physical arguments pertaining to dry friction). The influence of l is studied parametrically for the chosen application
case.
In summary, it must be noted that the current constitutive modeling approach has the advantages that it:

 Is multi-scale in the sense of relating macro-level responses (tractions, stiffnesses) using meso- and micro-level features;
 Is capable of approximating the rough contact response without too much overhead for a macro-level simulation;
 Is capable of accounting for localized variations across different parts of a surface;
 Enables the study of the evolution of relevant roughness properties with experimentation and their correlations with the
system-level responses.

3. Benchmark application

The current section applies the developed modeling framework for a bolted joint benchmark. The different processes are
summarized as a flowchart in Fig. 6. As can be seen, the modeling approach is coupled with surface assessments of the inter-
faces in the system. The non-linear analysis is conducted using the contact model developed through surface scan data
12 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 6. Summary of application framework.

imposed on the finite element mesh. Since such scans are non-destructive, practical deployment of the framework will not
incur any significant overhead in resource requirements beyond the need for specimens to be fabricated.
Following the interfacial scan, dynamic experiments are conducted in order to assess the model results. Since a quasi-
static modal approach is used for the simulations, experimental ring down data from impact hammer tests are transformed
into modal backbones for comparison. Following this, in order to study how features on the interface change during the
dynamical tests, interfacial scans are conducted after disassembly. The differences in the parameters are used for testing cor-
relations with simulated interfacial field quantities.

3.1. The Brake-Reuß beam: a bolted assembly

The Brake-Reuß Beam (BRB) is a bolted assembly structure [49] that consists of two ‘‘half-beams” connected together
using a lap-joint realized using three bolts. Depicted in Fig. 7a, the total length of the assembly is approximately 720 mm
(28.375 inches) and has a 1 in1 in square cross-section. Joining the beams, three sets of 5/16 bolts, nuts and washers
are used. The total length of the interface is 120 mm, with the holes separated by 30 mm center-to-center on both half-
beams. Fig. 7a depicts an ABAQUS model of the beam. Each bolt is pre-stressed equally.
The beam is meshed in such a way that it has a 10  10 grid of cubic (quad, C3D8R) elements in the square cross-section,
and the holes are meshed with 32 quad elements to have a reasonably sized mesh for dynamic analysis. The mesh is shown
in Fig. 7b.
The assembly is set up by creating CAD models of the ‘‘half-beams”, bolts, nuts and washers, and assembling them in ABA-
QUS. Mesh tie constraints are specified between the bolt-washer, washer-half beam, and nuts-washer interface. Note that no
constraint is specified between the interfaces and/or the bolts and nuts at this stage.
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 13

Fig. 7. The Brake-Reuß Beam Benchmark: (a) Assembly; (b) Interfacial Mesh.

3.1.1. Assembly modeling

In order to have a physically appropriate means of applying the prestress that can be used in substructured analyses, it
becomes necessary to model the interaction between the bolts and nuts in a reasonably accurate fashion. The idea followed
here is to arrest any relative motion in the plane perpendicular to the axis of the bolt (preventing any interpenetration) and
allow free translation along it. In order to realize this, virtual nodes are introduced and coupled to the inner and outer contact
surfaces of the nuts and the bolts respectively (see Fig. 8), and the planar degrees of freedom (x; y in this case) are constrained
to be equal for the two nodes using Multi-Point Constraints.
The actual prestress is realized by having appropriately directed nodal forces at these virtual nodes, i.e., a compressive
force along the
^z direction is applied on the nut-coupling nodes and a tensile load along the þ^z direction is applied on
the bolt-coupling nodes, all with magnitudes specified by the prestress level.
It must be noted that the above assembly, in the unstressed state, has seven eigenvalues corresponding to rigid body
modes (the degeneracy of the non-straining zero-energy modes is seven), with six components coming from the full assem-
bly lacking any boundaries and the seventh arising out of the fact that translation along the bolt axis is not constrained. How-
ever, once the loads are applied in the presence of an interfacial model, the degeneracy will reduce to 6, corresponding to the
rigid body modes of the assembled system.
In most solution approaches involving zero-energy modes (ZEMs), the ZEMs are first constrained out by estimating their
null-space and projecting the system onto it. For interfacial modeling applications with assemblies similar to the above,
however, not all of the ZEMs may be constrained out since the one with the translation along the bolt axis has a non-
trivial contribution to the response of the system. It thus becomes necessary to identify the correct set of six ZEMs to con-
strain out. This is carried out here by first transforming the system to a relative coordinate representation and then identi-
fying the ‘‘stuck interface” ZEMs that impart no straining of the interface. Denoting the degrees of freedom of the top and
bottom interfaces and the rest of the degrees of freedom as u~T ; u~B and u~R respectively, the relative coordinate transformation
is achieved as follows:
8 9 2 38 9
< u~T >
> = I I 0 > ~>
< Du =
6 7
u~B ¼ 4 0 I 0 5 u~B :
>
: ~ > ; >
: ~ >; ð30Þ
uR 0 0 I uR
|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}
T R

Here, Du~ ¼ u~T
u~B , is the relative displacement across the interface. This is followed by a Galerkin projection using the same
matrix to transform the original system to the relative coordinate representation. From this representation, the fully stuck
ZEMs may be estimated as the eigenvectors of the system formed by excluding the rows and columns corresponding to the
relative DOFs (setting Du ~ to zero). From physical arguments, it can be seen that the degeneracy of ZEM is just six for the stuck
interface system (and this can be verified numerically). These modes may now be transformed back to the original relative
coordinate representation by prepending it with an appropriate number of zeros. Stacked as columns, this set of six modes
forms the ZEM matrix Z whose null-space basis vectors L (subject to a mass-weighted inner product) is then used to trans-
form the system to be expressed in terms of the null-reduced DOFs u~n . This may be expressed as,
8 9
< Du
> ~>=  
u~B ¼ ½L : ð31Þ
>
: ~ > ; u~n
uR

The final system inertia matrix M, stiffness matrix K, and bolt load vector F~b are given, after the above two transformations,
in terms of their original counterparts (characterized withb’s) as

M ¼ LT T TR c
MT R L
b T RL
K ¼ LT T T K ð32Þ
R
c
F~b ¼ LT T TR F~b :
14 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 8. Coupling constraints used for the assembly: (a) Inner surface of nut coupled to a virtual point; (b) Outer surface of bolt coupled to a virtual point.

The linear system comprising of M and K has a single ZEM corresponding to the translation along the bolt axis that has not
been constrained out in eq. (31). Thus, the system is ill-posed in the absence of a contact model and has to be studied only in
the presence of a model for the interface.
In practice, all of the steps from eq. (30) are carried out on substructures constructed using fixed interface component
modes extracted from the Finite Element program (ABAQUS).

3.1.2. Nonlinear modeling

For the nonlinear modeling, the contact models developed in section 2.2 are applied with the ZTE’s (section 2.1) from the
relative coordinate system and then transformed to the null-reduced system through L. For a general excitation force F~ex ðtÞ
expressed in relative coordinates, the nonlinear dynamic system is

€~n þ K u~n þ LT F~nl ðLu~n ; . . .Þ ¼ F~b þ LT F~ex ðtÞ:

Mu ð33Þ
For static analysis, say of just the bolt prestress, the dynamic terms are dropped and the problem is solved quasi-statically.
For dynamic calculations, an improved nonlinear hysteretic implementation of the Quasi-Static Modal Analysis (QSMA)
approach expounded in [15] is formulated. The difference between the current Coupled Quasi-Static Modal Analysis
(CQSMA) approach and the previous approach is twofold: (a) QSMA decouples the static and dynamic simulation steps, while
CQSMA conducts the simulations in a coupled manner and conducts the decoupling only in the analysis/post-processing
part; and (b) CQSMA is generalized for non-Masing models too, while QSMA has only been used with Masing’s hypothesis
so far.
For QSMA [15], the static pressure distribution is incorporated into the dynamic analysis by conducting all of the analyses
~  and the perturbed solution
on a perturbation of the solution from this step. Denoting the solution of the static problem by u
~¼u
by u  ~u ~
~ þ d (with d being the perturbation), it is shown that the equations of motion can be reduced, using a first order
u

Taylor’s expansion about the static solution, to

ð34Þ

Here, the forcing vectors F~nc ; F~frict ; F~bolt , and F~ext denote the normal contact forcing used for the static analysis, the frictional
contact model used for the dynamic analysis, the constant bolt load, and the external excitation respectively. The main draw-
back here is that it is not appropriate to use a nonlinear normal contact model for the dynamic analysis since it has already
been linearized into the Jacobian, and doing so will introduce an over-stiffening effect on the interface.
Following this, the modal characteristics are extracted by applying a quasi-static forcing in the shape of the corresponding
mode shape, /, of this augmented system. The nonlinear problem becomes

K A ~du þ F~frict ðu ~
~ Þ ¼ aM /; ð35Þ
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 15

where a is the modal forcing amplitude. The natural frequency at a particular forcing amplitude a is given by the secant stiff-
ness of the modal hysteresis loop and the damping factor is given using the area inside the hysteresis loop. The exact expres-
sions commonly encountered are
rffiffiffi
~ T M ~du ; a D
q¼/ x¼ ; f¼ : ð36Þ
q 2pðqxÞ2
The dissipation is sometimes more accurately estimated by calculating it individually for each frictional element in the
model and summing them up.
Taking inspiration from the above, CQSMA starts off by first solving the prestress problem

~  þ F~nl ðu
Ku ~  Þ ¼ F~bolt ; ð37Þ

after which the mode shape, /, ~ is calculated about the Jacobian evaluated at the solution u
~  (similar to QSMA). For the
dynamic analysis, the full system, given by
~€ þ K u
Mu ~ þ F~nl ðu
~Þ ¼ F~bolt þ F~ext ðtÞ; ð38Þ
is used instead of the perturbation. Note that there is no need to separate the normal and frictional contact models here.
However, the modal forcing vector is constructed based on the mode-shapes of the augmented system linearized about
the prestress solution as before,
" ! #
@ F~nl ~
~ ¼ 0:
Kþ
x M /
2
ð39Þ
@u~

This sets up the following system for the modal quasi-static analysis:

~ þ F~nl ðu
Ku ~ Þ ¼ F~bolt þ aM /:
~ ð40Þ
~ from the static
The modal amplitude here is determined by taking the inner product of the deviation of the CQSMA solution u
~  with the chosen linearized mode shape as
solution u
~ T M ðu
q¼/ ~
u
~ Þ: ð41Þ
In a like manner, the dissipation is taken as the work done by the perturbation tractions on the perturbation displacement
field in the interface. Thus, the modal characteristics are estimated by,
rffiffiffi  
a D u ~  ; ~t
~t 
~
u
x¼ ; f¼ : ð42Þ
q 2pðqxÞ2
~t has been used here to denote the tractions at the end of the static analysis. These are subtracted from the traction vector
since the work done by the tractions developed in the pre-stress step is not of any significance for dynamic operation. The
history-dependent manner in which the contact models are formulated in section 2.2 naturally lends them to hysteretic
quasi-static calculations, thus generalizing the approach to non-Masing problems (such as those with separation). In the cur-
rent work, the hysteresis loop is idealized as a polygon with a prescribed number of vertices and the area within it is calcu-
lated using the Surveyor’s/shoelace formula [70]. For a planar polygon defined by vertices ðxi ; yi Þ, with i ¼ 1; . . . ; n ordered
along clockwise, the formula for the area is given by the expression
 
1 Xn
1 X
n
1 

A ¼  xi yiþ1 þ xn y1
xiþ1 yi
x1 yn : ð43Þ
2  i¼1 i¼1


If one estimates the hysteretic areas under the traction-displacement characteristics at any point on the interface, the
derived quantity will be the flux of dissipation, i.e., the amount of energy dissipated per unit area at that location. Visualizing
the flux of dissipation over the interface and classifying different portions of the interface based on it can be a very useful
approach for wear and damage assessment and/or prediction. This will be used in section 4.3 for studying correlations to
evolutionary trends.
Fig. 9 provides a detailed overview of the modeling approach. The different blocks summarize the different modeling and
analysis steps followed in sections 3.1 to 3.3. Recall Fig. 6 for the context of this in the complete framework.

3.2. Experimental setup for interfacial scans

A Keyence in-line profilometer is used in the current study to scan the interfacial surfaces of the jointed connection. The
profilometer measures the heights of points on a surface in a certain range (specifically
23 to 23 mm, where the 0 level
corresponds to the point 80 mm vertically below the head) through a laser transmitter–receiver pair. As shown in
Fig. 10a, the laser probe line sweeps through at a constant velocity (approximately 0.2 mms
1 ) to extract the heights across
the surface.
16 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 9. Overview of Modeling approach.

Fig. 10. Interface scanning setup: (a) The optical profilometer; (b) Raw data, and (b) Plane-adjusted data. (Colouring indicates element).

The vertical axis resolution of the instrument is approximately 0.5 lm and the x and y direction step sizes are (approx-
imately) 30 lm and 50 lm respectively (coordinate systems as depicted in Fig. 10a). Although a much higher resolution
device will be necessary for a more exact tribological characterization, this configuration was deemed sufficient for the cur-
rent exploration.
In order to account for variations in the straightness of the beam and/or misalignment between the sweep direction and
the surface tangential plane, it is first assumed that the interface itself is nominally flat and any variation seen in the
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 17

extracted data is due to physical deviations of the beam from the ideal geometry. The raw data is de-trended by subtracting a
trend plane that is built using the data along each free edge. Fig. 10b to c depict the raw and de-trended data for one such
scan.
Following the procedures outlined in section 2.3, the relevant properties are estimated for each element in the interface
after first localizing the data within the underlying finite element mesh. The element-wise total roughness parameter Rt is
used as the peak heights in the current study (so that the asperity heights are all positive). The results from the scan are pro-
vided in Fig. 11. After this, the contact model parameters (five for normal traction and four each for the two tangential stiff-
nesses) are estimated by conducting the curve fits upon bootstrapped simulations for each element as outlined in section 2.4.
For constructing the interfacial meso-scale gaps fitted using the bilinear planes (see Fig. 11a), the planes are first fitted
separately for both of the interfaces and then added together to estimate the ‘‘gap plane” that is useful for the analysis. This
addition is justified by the assumption that the two interfaces do not interact when they are not pre-stressed (meaning that
the gap function can never be negative). Following this, the contact condition at any location ~ x 2 C (using C to denote the
interface) becomes

untop ð~xÞ
unbot ð~xÞ P g f ð~xÞ; ð44Þ

where g f ð~
xÞ denotes the gap function and top and bot denote corresponding values from the top and bottom faces respec-
tively. Since bi-linear planes are used for this, the function may easily be represented as the shape function-weighted
sum of nodal contributions. In practice, the form a0 þ a1 x þ a2 y þ a3 xy, upon being fitted, is evaluated at four quadrature
locations in each element and nodal values are extracted as the best possible values for representing the gap at the quadra-
ture locations across all the elements in a least-squares sense (see Super-Convergent Points (SCP) in classical finite element
texts such as [71]). Practically however, this was observed to be prone to experimental errors in the current experimental
setup. This is due to the fact that no precise mechanism has been employed to ensure the co-planarity of the probe direction
and the surface plane. Since the meso-scale interface is unreliable, simulations are conducted both with and without taking it
into account in order to observe its influence on the response.
b the interface values are taken as the average of the estimates for each of the faces.
For the other parameters (Rt ; b ; b ; N), X Y
The estimates for these parameters, as opposed to the gap function, were found to be fairly repeatable across multiple scans
and thus there is considerable confidence in the developed contact model parameters.

3.2.1. Adhesive effects

As remarked earlier, the influence of adhesion is assumed to be negligible in the current modeling approach. In order to
assess this, the dimensionless intermolecular distance parameter that is used in [72,73] is computed. A key quantity required
for this is the critical interference parameter
 2
pKH
xC ¼  R: ð45Þ
2E
pffiffiffiffiffiffiffiffiffiffi
Here the effective radius R is taken as bX bY ; the hardness H is taken as 9.47 GPa (from experimental data published in Ref.
[74] for 2.2 mm radii spheres); the parameter K is taken as 0:454 þ 0:41m (with m ¼ 0:29 being Poisson’s ratio, giving
  
K ¼ 0:5729); and the effective Young’s modulus E taken as E= 2 1
m2 (E = 192.85 GPa, yielding E = 105.28 GPa). Using
a typical inter molecular separation [72] of =0.4 nm, the dimensionless parameter =xC comes out to be less that 0.1% for all
the elements (see histogram in Fig. 12a).
Furthermore, the relationship in Ref. [72] and the asperity geometries in section 2.3, may be used to estimate the
expected adhesive traction. Fixing the work of adhesion value to 2:5 J=m2 (data for low carbon steel on low carbon steel con-
tact [75]), Fig. 12b plots the expected adhesive traction along with the fit established before for the elastic traction as func-
tions of the normal displacement (global interference/relative displacement of the two interfaces). Note that the adhesive
tractions, in keeping with the spirit of the presented approach, represent the average adhesive traction over the considered
element and not the traction on any single asperity. It can be seen that there exists a threshold below which adhesive effects
may not be ignored. However, since the operating regime for the current investigation (in the central region of the interface)
involves normal displacements on the order of a micron or more, adhesive tractions are more than one order of magnitude
lesser than elastic tractions. Studying the influence of adhesion in the outer regions of the interface on the dynamics of the
system is beyond the scope of the current paper. The currently developed model therefore does not account for adhesive
effects.

3.3. Experimental setup for static pre-stress

As mentioned earlier, there is a high amount of variability in the torque-tension relationship of bolts. The origins of these
may range from material/geometric imperfections to environmental conditions and thus may not always be amenable to
accurate modeling. Therefore, the current study utilizes an experimental technique developed in [5] to ensure better accu-
racy in the interfacial loading conditions. The threads in the bolts are machined off in the internal regions of the beam and
strain gauges are attached to the curved surface of the bolt (see Fig. 13a).
18 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

b (d)
Fig. 11. Results of surface scan: (a) Interface fitted with bi-linear planar elements; (b) Surface roughness Rt ; (c) Asperity count estimate per element N;
Radii when seen along horizontal (scan along vertical); and (e) Radii when seen along vertical (scan along horizontal).

Fig. 12. Assessing influence of adhesion: (a) histogram of dimensionless intermolecular distance across all elements; and (b) comparison of elastic and
adhesive normal traction for the indicative element (previously used in Fig. 5).
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 19

Fig. 13. Bolt pre-stress characterization (a) Experimental bolt with the strain gauge, (b) Finite element model with loading conditions and output element,
and (c) Force-micro-strain response characteristics.

A finite element model of the modified bolt is prepared to characterize the force-strain characteristics. Since the deflec-
tions are expected to be in the linear domain, a linear finite element model is sufficient to capture the load–strain relation-
ship (see Fig. 13b for the mesh and loading conditions). The model is fixed at the threads and loaded uniformly in the base
surface of the bolt-head. The average strain from an element in the outer surface of the bolt (highlighted in red) is extracted
for the study. The force-strain characteristic is fit to a linear function (see Fig. 13c) yielding the relationship

F ¼ 6:2683  106 : ð46Þ

Here, F denotes the bolt load in N and , the strain. For the required load of 11.580 kN (corresponding to a 20 Nm preload (as
per the analytical solution in [76]), the bolt must develop a strain of 1847 l. However, experimentally the bolts were tight-
ened to strains averaging around 2050 l, which corresponds to a prestress level of 12.845 kN (using the same expression,
corresponding to around a 20 Nm torque level). Thence, this prestress level was used in all the simulations

3.4. Experimental setup for dynamic testing

Fig. 14 depicts the test setup used for the impact hammer tests. The beam, after being assembled (with the prepared
bolts) with the correct prestress levels, is suspended using two bungee cables. Two accelerometers are attached at the
extreme ends and hammer impacts are made from the point shown in the figure. Five ‘‘medium level” impacts (~350 N ampli-
tude), one ‘‘low level” (~80 N) and one ‘‘high level” impact (~850 N) are carried out on the current structure. The ring down
acceleration data is recorded from the accelerometers in order to identify the nonlinear dynamic properties. The wires from
the bolts were wrapped loosely around the assembly during the test. This configuration was determined, through multiple
experiments, to have no appreciable effect on the dynamics of the system.
The current study focuses solely on responses dominated by the first bending mode in the system (depicted for a low
amplitude case in Fig. 15). Thence, the data from one of the accelerometers is filtered around this frequency by a 4th -
order Butterworth bandpass filter and the Peak-Finding and Fitting (PFF) method (see [42]) is used to process the filtered
data to obtain amplitude-modal characteristics of the system. This extracts the modal frequency and damping ratio as func-
tions of the filtered amplitudes at the chosen accelerometer location. Since QSMA and CQSMA (Section 3.1.2) characterize the
system when they are excited solely by the mode of interest, the corresponding amplitudes from the simulations may
directly be correlated with the filtered amplitudes. In the current study, instead of retaining the end nodes as part of the
boundary DOFs, the corresponding part of the recovery matrix (transforming subtructure DOFs to the FE model DOFs) is
employed to extract the necessary amplitude since this is thought to provide a more accurate substructure than the former
approach.

4. Results

4.1. Experimental Identification

In order to first identify the density and Young’s modulus of the beam, low amplitude hammer tests are conducted for the
half beams separately after having them weighed. The extracted values for density and Young’s modulus are 7857.8 kgm
3
and 192.85 GPa respectively.
Following this, the beams are tightened to a prestress of 12.845 kN (as mentioned before) by monitoring the strain expe-
rienced by the bolts and the nonlinear impact hammer tests are carried out following the procedure described in [77].
20 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 14. Experimental Setup for Hammer Test.

Response Node 2
(hidden)

Response Node 1

Y X

Fig. 15. The mode shape of interest: First bending mode (linearized about an Abaqus frictionless hard contact prestress simulation in Abaqus). The recovery
nodes are highlighted in red.

Fig. 16. Experimental results: (a) Frequency domain comparison for the bandpass filtering step; and (b) Identified modal characteristics; lines with
asterices () corresponds to high amplitude hit (others are medium level repeats).

Fig. 16a depicts a sample of the frequency content of the signal before and after the initial bandpass filtering step. Only the
accelerometer data from the sensor in the side of the hammer is used for the current study.
Fig. 16b depicts the final modal characteristics after these are processed using the PFF technique. It can be observed that
the frequencies start off from around 180 Hz, where the damping is very low (approximately 0.1 %), and then proceed to
decrease gradually, while the damping increases by many folds. The largest amplitude hit is depicted using red lines with
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 21

asterices (). The experimental data are averaged and smoothened (using a Savitzky-Golay filter) to get a single characteristic
line for each of the quantities.

4.2. Numerical implementation

4.2.1. Static prestress analysis

The static prestress analysis is conducted using 25 quadrature points in each element. This number was chosen in spite of
the fact that lesser numbers were observed to work well too, since a threshold was observed in some cases below which it
was prohibitively difficult to get the nonlinear solver to converge. All of the plots in this section are generated using a uni-
form coefficient of friction l ¼ 0:6 throughout the interface 1.
Fig. 17 plots the resulting traction field on the interface, with the traction evaluated directly at the element centroids
shown in Fig. 17a. This can be used as an indication about whether or not the element is in contact: if there is a non-zero
traction developed, then the element is in contact. Elements with zero traction are left uncolored (white in the figure). It
can be observed here that there are two large regions near the ends that are out of contact during the static loading. This
can directly be correlated to the large values of the gap function (plotted in Fig. 11a) in these regions.
Fig. 17b to 17d plot out the colour maps of the nodal traction values in each direction interpolated in a least-squares
sense. It must be noted that even though the normal tractions at the quadrature locations are strictly positive (as are the
nodal forces), the least-squares solution will require nodal traction values to be negative in some regions. This is non-
physical and a relic of the fact that the pressure variation in the interface cannot be easily captured by the finite element
shape functions natively 2. However, since the model in itself requires only integrals of these quantities, the integrals can
be carried out accurately without this least-squares fit for the simulations (they were conducted only for the depictions).
The absolute values of the interpolated normal pressures are plotted in Fig. 17b so that the features may be compared with
the centroidal values in Fig. 17a.
As expected, it can be observed that there is very little to no slippage that occurs in the tangential directions (hence nom-
inally zero tangential tractions in Fig. 17c to d). However, it can be observed that there are locations where there are con-
siderable tangential kinematics that must not be ignored.
In order to understand the influence of the gap function and the uniformity of parameters (of the contact models used),
two further simulations are conducted: (a) the interfaces are perfectly flat in the meso-scale, but have a non-uniformly dis-
tributed surface roughness (as characterized); and (b) the interfaces are perfectly flat and the roughness is uniform across
the interface. Fig. 18a to b present the results of the two cases in order.
The first major ramification of assuming flat interfaces is that the regions in and out of contact are dramatically changed.
Both Fig. 18a and b look relatively similar in the elements that are in contact. The influence of making the contact model
uniform is that the interfacial traction becomes more uniform through the interface as can be observed by comparing the
two figures.
Due to the way the model of the assembly was created, it was noted previously that the resulting model is singular since
it does not impose any constraint on the rigid body mode corresponding to the two half beams separating along the bolt axis
direction. However, upon application of the prestress, the perturbation model about the prestressed state does not possess
this zero energy mode. In fact, the linearized eigen-modes of this perturbation state correspond to low amplitude modes of
the assembled system.
Table 1 tabulates the linearized mode frequencies for the first mode for each of the above cases along with the experi-
mental measurement. It can be observed that the flat, non0uniform case underpredicts the frequency by more than 12 Hz
while the other cases are slightly better. Note that the apparent similarity between the ‘‘Flat; Uniform” case and the actual
interface is only by chance; the parameters used here were taken from an element chosen in random and thus cannot be
used to make further inferences. The relative closeness of all three modeling approaches gives confidence that this approach
can be used for blind predictions on a structure that has not yet been fabricated.

4.2.2. Nonlinear dynamic analysis through CQSMA

As noted previously, the coefficient of friction is a persisting unknown in the formulation. For convenience, it is assumed
to be uniform across the interface. Fig. 19 depicts the amplitude characteristics for different l values. The abscissas of the
plots are the amplitudes at the response locations (see Section 3.4).
A few observations can be made upon first glance:

 There is an offset of about 8 Hz in all the models for the low amplitude (nominally stuck case) in the frequency charac-
teristics (see Fig. 19a).
 The coefficient of friction seems to determine the amplitude level characterizing the onset of slip-like behavior in the
interface: for smaller values of l, the onset of slip occurs at lower amplitudes (see Fig. 19b). Thus, for a given amplitude
level, a smaller value of l implies increased dissipation.

1
Little to no appreciable change was observed for different values of l for the prestress analysis
2
This is the same reason that traction values evaluated at element centroids are better indicators of the contact state than the nodal quantities.
22 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 17. Results of static prestress analysis: (a) Contact pressures at element centroids (empty elements indicate separation); (b) Interpolated normal
contact pressure; (c) X-tangential contact pressure; and (d) Y-tangential contact pressure.

Fig. 18. Influence of flatness and non uniformity (empty elements indicate separation): (a) Flat interface with non uniform roughness; and (b) Flat interface
with uniform roughness.

Table 1
Comparison of linearized mode-frequencies for the three
different configurations studied here with the experi-
mental prediction. The ‘‘Non-flat; Non-Uniform” case
corresponds to the characterized interface.

Configuration Mode Frequency (Hz)

Non-Flat, Non-uniform 174.21
Flat; Non-Uniform 168.91
Flat; Uniform 172.76
Experimental 180.92

 There seem to be two markedly different dissipative behaviors in the amplitude regime of interest: for l < 0:20, the dis-
sipation characteristic seems to have a concave region in the region of interest and for l P 0:20, the characteristic seems
to be more convex. This may be attributed to the ‘‘region” in the stick–slip transition the regime of interest lies in.
 Close to l ¼ 1:0, the dissipation characteristics seem to converge to a particular curve (as seen by the 0.75 and 0.99 curves
in Fig. 19b).
 The power-law slopes in the dissipation plot Fig. 19c do not seem to vary much over the different l values. The implica-
tion even a small variation has on the response is however exemplified by the amount of variations the other two sub-
plots seem to exhibit.
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 23

Fig. 19. Amplitude-Modal Characteristics for the first mode for different l values: (a) Variation in natural frequency; (b) Variation in effective damping
factor; and (d) Variation in cyclic dissipation.

The low amplitude offset may be attributed to several factors including persisting uncertainties in the bolt prestress, errors
in the interfacial characterization, epistemic uncertainties in the contact models employed, etc (see Appendix A for some
discussions). Coming to the contact model, it must be noted that the Hertzian solution presents an over-prediction of the
normal traction (and hence the stiffness). Thus it is not realistic to expect a larger ‘‘true” stiffness here 3. For the tangential
contact model on the other hand, a standard form has been assumed for simplicity. Using more involved models (such as those
in [25]) could offer possible improvements.
The observations on the influence of the coefficient of friction may be understood by visualizing the hysteretic behavior of
the tangential oscillations. Consider a constant slip force-constant stiffness case as shown in Fig. 20. For a low coefficient of
friction (red), the element starts dissipating at a much lower displacement amplitude as compared to an element with a
higher friction coefficient (blue). By comparing different displacement amplitudes and where they lie with respect to the
thresholds for each element, one it is obvious that even beyond the point where the second element starts slipping, there
will be a certain threshold until which the first element will dissipate more. Since the interface consists of many such ele-
ments, the above observations may be corroborated with what would be expected from such a trend.
In order to understand the significance of the modal characteristics in Fig. 19, it is important to develop some insight into
the hysteretic phenomena in the interface. This is done here by visualizing the dissipation fluxes through the interface.
Fig. 21 plots the dissipation flux distribution over the interface for four different response (or forcing) levels along the modal
backbone, signifying a nominally stuck level, a stick–slip-transition level, a nominally slipping level, and an unrealistically
high level. Fig. 21a to c depict the numerical and experimental backbones in terms of the natural frequency, effective damp-
ing factor, and total dissipation (respectively) for l ¼ 0:2 over a larger response amplitude level than in the previous figure.
The first feature that is readily apparent from the extended backbones (Fig. 21a to c) is that the behavior of the system
post nominal slippage looks to be slightly more involved than that observed for Single Degree-Of-Freedom (SDOF) models
with elastic friction. In SDOF models there exist distinct stuck and slipped frequencies with the transition between them
being characterized as the nonlinear phenomenon. It has also previously been shown, since the slipped frequency is smaller
than the stuck frequency, that the resulting damping factor will show a downward trend from the slipping point, which is
also accompanied by a change in the power-law slope of the dissipation curve [78,79,49]. In the current application, although
the dissipation and damping factor behaviors seem to be consistently reminiscent of the SDOF systems (Fig. 21b to c), the
frequency shifts are not. The frequency seems to progressively decrease with the forcing level. It must however be noted that
the amplitude levels in the right-most end of these plots are in the order of 1 m, which is an unrealistic scenario for the cur-
rent application since it will result in the system incurring significant plastic deformation, but the results may be helpful for
developing intuition about the underlying frictional phenomena.
Looking at the dissipation fluxes Fig. 21d to g, the first observation is that the magnitudes in the normal directions are
several orders of magnitude lesser than that in the tangential directions. This is consistent with the fact that the normal con-
tact model is perfectly elastic and no dissipation is expected to happen here. Due to the nature of the mode shape (see Sec-
tion 3.4), relative translation on the interface is expected in the X direction only, and all Y directional effects will be due to
Poisson effects alone, since the model is perfectly symmetric about the vertical plane (XZ plane). The influence of the Y dis-
sipation, although much smaller than that of the X dissipation, are nonetheless considerable, as may be seen from the figures
in all cases.
For Point 1 (Fig. 21d), the system is expected to be nearly linear since the amplitude level is extremely small. The interface
however does dissipate, and this may be observed from the figure too. Most of this comes from slippage near the ends of the
interface where the contact pressure is expected to be lesser since they are not directly under the bolts. Considering the pre-

3
For non–Hertzian contact such as pin-in-hole or conformal contact, the elastic stiffness could be up to 25% higher than for the Hertzian case [16]
24 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 20. Idealized hysteresis loops for a low l (red) and a high l (blue) element.

stress results (Fig 17), where it was remarked that a considerable region towards the ends is found to be out of contact as per
the meso-interface gap function, the distribution of the X dissipation here appears similar (there is a low dissipation region
near the ends).
Point 2 (Fig. 21e) corresponds to a response level just before the frequency makes a considerable ‘‘down-ward turn”. It can
be seen that the dissipation fluxes are now larger in magnitude and also that more regions are participating in the frictional
interactions. This can be corroborated with the fact that the damping factor is steadily increasing around this level. An addi-
tional observation that can be made here pertains to the detection of nonlinearity. Looking at the frequency response alone,
the system appears to still be in the near-linear regime, but the damping factor plot shows that the frictional nonlinearities
are very much active at this point. From an identification perspective, this implies that the detection of dissipation nonlin-
earities will be a much more reliable way of characterizing bolted joints than frequency-shift nonlinearities.
At point 3 (Fig. 21f), the system may be said to be in a fully slipped condition, had this been an SDOF frictional element,
since it is chosen at the peak of the damping factor curve and the point of slope-change in the power-law amplitude-
dissipation slope. The flux distributions indicate that at this level, almost the complete interface is participating near-
equally in the energy dissipation. This is characteristic of a nominally slipped state. The exact nature of the deflection is how-
ever not akin to macro-slip, but an expansion of the whole interface (left segment expands left-ward; right, right-ward; and
center stays without expansion). There have been some recent studies that indicate that such behaviors can indeed be
observed experimentally [8,9].
At point 4 (Fig. 21d), with an extremely high level of excitation (albeit non-physical for the current system), the dissipa-
tion flux distribution reveals that the interface is no longer being ‘‘utilized” in an equal fashion. Although there is dissipation
throughout the interface, the amount of energy lost near the ends is much greater than that lost in the inner regions of the
interface. This ‘‘loss of optimality” in the participation of different portions of the interface seems to be playing an important
role in the shift of the dissipation power law slope which also characterizes nominal slippage. Further, since the interfacial
separation is expected to occur in a much more drastic fashion, the simulation detects a significant amount of dissipation in
the normal direction too. One must, however, not interpret this as something necessarily physical since one of the basic
assumptions in the CQSMA procedure, i.e., that the mode shape may be assumed to remain constant over the operational
regime, is clearly violated for these high amplitude levels. Further investigation is necessary to understand if these are
merely numerical artifacts or if there is some underlying physics. Another aspect that must be taken into consideration at
these large levels is bolt-pinning. Since no contact model is employed to capture the interaction of the hole and the bolt,
the added stiffness due to the pinning of the bolt is not accounted for in the current model.
Since the Meso-scale topology characterization is prone to experimental error in the current setup, simulations are also
conducted for the fully flat case, where the gap function is taken to be uniformly zero everywhere. The main purpose of this
is to develop insights into the contribution of the gap function to the accuracy of the system level solution. An initial obser-
vation was made from the linearized frequency values in Section 4.2.1, where the ‘‘flattened” system seemed to show a
reduced stiffness. Fig. 22 plots the results of this set of simulations in a format similar to the previous one. The differences
are not appreciable for the modal characteristics plots due to the qualitative similarity of Figs. 22a to 22c their counterparts
in Fig. 21a to c apart from the initial offset in the stuck frequency and some differences in the transition to slip region. How-
ever, the dissipation fluxes show a much more pronounced influence. This seems to suggest that the prediction of stiffness
nonlinearities is a much easier problem than the prediction of dissipative nonlinearities for bolted joint applications, consis-
tent with previous observations [1]. Although internal resonances are primarily driven by frequency shifts, it is the opinion of
the current authors that accurate modeling of the non-dissipative components involved will be a much easier task than the
modeling of dissipative components.
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 25

Fig. 21. Interfacial dissipation fluxes for different loading conditions for l ¼ 0:2. (a)-(c) are the modal response characteristics (as in Fig. 19) over a broader
response level range (dark continuous lines are experimental and the dotted blue ones are simulated curves); (d), (e), (f), and (g) are the dissipation fluxes
plotted over the interface in the tangential (x and y) and normal (z) directions for forcing amplitudes corresponding to point 1, 2, 3, and 4 on (a)-(c)
respectively.

Once again, four points are identified so that the results may be compared with the previous case. It can be seen that the
qualitative understanding that maximal dissipation occurs when the entire interface participates near-equally in energy dis-
sipation still seems to hold in this case. The exact regions of dissipation are, however, slightly different from the previous
case, such as the absence of ‘‘gaps” near the ends of the structure, etc. These are mostly expected since the meso-scale irreg-
ularities seem to be the primary cause of these details.

4.3. Micro-scale trends

Irreversible changes in the tribological and topological properties of the interface may be interpreted as wear. In the cur-
rent context, since the contact laws are all generated from the roughness parameters, it is now possible to isolate only the
employed roughness parameters as the most relevant for modifying the response of the system. Further, since interfacial
traction and dissipation fluxes can be thought of as the prime ‘‘drivers” of wear in the interface, the current section attempts
to understand if it is possible to establish statistical correlations between these parameters and the interfacial parameters.
For this study, the experimental test procedure followed is,

1. Interfacial scan 1;
2. Assembly;
3. Bolt prestress;
4. 7 Hammer impact tests;
5. Disassembly;
6. Interfacial scan 2.

The whole procedure is completed in a single day within a span of about 5 h and thus it is very meaningful to look for pos-
sible correlations in the way interfacial parameters change between steps 1 and 6. The change in roughness DRt ; x- and y-
radii bX ; bY , and the change in the estimated number of asperities per element are taken as the most critical ‘‘responses”.
26 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 22. Interfacial dissipation fluxes for different loading conditions for l ¼ 0:2 for the fully flat interface. (a)-(c) are the modal response characteristics (as
in Fig. 19) over a broader response level range (dark continuous lines are experimental and the dotted blue ones are simulated curves); Additionally the
dark dashed lines represent the simulated responses for the non-flat interface case. (d), (e), (f), and (g) are the dissipation fluxes plotted over the interface in
the tangential (x and y) and normal (z) directions for forcing amplitudes corresponding to point 1, 2, 3, and 4 on (a)-(c) respectively.

Since a considerable amount of experimental variability was observed for the gap function, this is omitted for the current
study. The traction in the interface at the end of the prestress analysis T n , and the total dissipation flux Dx þ Dy þ Dz (see dis-
cussions in Section 4.2.2) at Points 1 and 2 in the response backbone (labeled as D1 and D2 respectively) are treated as the
possible factors for the current study. Since an exhaustive treatment of correlation is not the purpose of the current work, a
simplified approach is followed in order to gain insights for future work. A single value is calculated for each element in the
interfacial mesh for the factors and parameters, and Spearman’s correlation coefficient [80] is used to test for monotonic
relationships between the factors and the responses. The null hypothesis for the test may be stated as,
H0 : There is no ½monotonicrelationship between the factors and the responses½in the population: ð47Þ
The correlation coefficient q may be calculated for each factor-response pair and a p value may be estimated from this for a
significance interpretation. q, bounded between
1 and +1, denotes the amount of relationship, i.e., values close to 1
denote strong monotonic (increasing or decreasing) tendencies, while values closer to 0 indicate a smaller relationship.
The results are interpreted with a significance level of 0:05, so that there is a confidence of 95% to reject the null hypothesis.
Tables 2a to 2b summarize the results of the analysis. As already mentioned, the current approach for characterizing the
gap function is prone to errors and thus factors calculated from a fully flat (but non-uniform roughness) simulation are also
studied for correlation, and are super-scripted with f. From the p values, it can be seen that the null hypothesis may not be
rejected for most of the cases. This can imply either that the current set of factors are insufficient or that the data is not pre-
cise enough (as the resolution is only 30 lm  50 lm  0:5lm in the x; y, and z directions). However, for the changes in
roughness, there seem to be significant confidences in rejecting the null hypothesis pertaining to the normal traction and
dissipation flux. This trend is observed identically for the simulations with and without the gap function. Comparing the
numbers in the two cases, it appears that there are slightly stronger correlations and higher confidences in the relationships
   
for the latter case. Other than this, there are high confidence results for two other pairs, namely, T n ; D N b and Df ; Db .
1 X

From the above discussions and the actual values of q, the following comments may now be made about the statistically
significant cases.
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 27

Table 2
Results of Spearman correlation tests: (a) Spearman q; and (b) the corresponding p values. Cases where the null hypothesis may be rejected with a confidence
of more than 95% are highlighted with boxes. Superscripts f denote the flat interface factors.

DRt DbX DbY b

DN
(a)
Tn 0:0822
0.0578 0.0393
0:1534
D1
0.0670 0.0615
0.0253 0.0533
D2
0:0829 0.0507
0.0361 0.0322

T fn 0:1528
0.0462 0.0345
0.0408

Df1
0.0365
0:0810
0.0411 0.0156
Df2
0:0862 0.0036
0.0261 0.0502

(b)
Tn 0:0463 0.1616 0.3411 0:0002
D1 0.1046 0.1364 0.5395 0.1969
D2 0:0445 0.2197 0.3819 0.4350

T fn 0:0002 0.2632 0.4041 0.3236

Df1 0.3767 0:0497 0.3202 0.7051

Df2 0:0368 0.9315 0.5282 0.2244

 
 Since both qðT n ; DRt Þ as well as q T fn ; DRt are greater than zero, this indicates that the interfacial roughness may have an
increasing relationship with the traction experienced at the end of bolt tightening.
 
 Both qðD2 ; DRt Þ as well as q Df2 ; DRt are less than zero, implying that greater operational dissipation through a region
may be correlated with a decrease in the roughness there.
 
b and qðD2 ; Db Þ, which have also turned out to be statistically significant
Similar interpretations also follow for q T n ; D N X

relationships. All of the correlation coefficients are very small in magnitude, indicating that the onset of wear is typically very
gradual for experiments similar to the current set.

5. Discussions and conclusions

The specific contributions of this work are,

 A multi-scale roughness-based compliant interface model has been formulated for interfaces in contact. Relationships are
drawn from rough contact theory to establish the exact form of the model and it is implemented in a traction-consistent
manner using interfacial zero-thickness elements (Section 2.1, 2.2).
 The approach is demonstrated to lend itself to applications involving minor deviations in the initial surface that can be
expressed in the form of gap functions (Section 2).
 The hysteretic contact laws are implemented in an incremental quasi-static formulation that is derived as a representa-
tion of the rate-dependent form of the nonlinearities (Section 2.2.2). This makes the implementation convenient for
modal quasi-static analysis approaches.
 The use of null-space reductions to remove the zero energy modes of bolted structures in the fully free configuration has
been introduced rigorously, which will be helpful for conducting nonlinear analyses have been highlighted (Section 3.1.1).
 Coupled quasi-static modal analysis, an improved formulation of QSMA [15] has been formulated and successfully
employed for the benchmark implementation (Section 3.1.2). The current improvement, referred to in the text as CQSMA
has the following key differences from QSMA,
– It makes the implementation of nonlinear interfacial laws mathematically consistent. QSMA solves for the linearized
perturbations about the prestress solution, while CQSMA solves for the actual displacements.
– It makes it possible to account for prestressing and dynamic loading in a coupled fashion. By formulation, QSMA had a
fully decoupled static simulation step followed by the modal excitation step, while CQSMA conducts both simultane-
ously, in a way that is more consistent with reality.
 A repeatable procedure has been used for pre-stressing the bolts using strain gauges to monitor the axial extension of the
bolts (Section 3.3).
 Different response regimes of a bolted joint structure are highlighted using a benchmark structure through modal char-
acteristic plots and interfacial dissipation flux distribution maps (Section 4.2.2).
 It is demonstrated that micro-scale trends may be successfully studied through statistical means in order to characterize
correlations between simulated factors and experimentally measured parameters (Section 4.3).
28 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

The main drawback in the current study comes from experimental limitations introduced by the laser profilometer set up.
Since no mechanism has been used for ensuring perpendicularity of the scanning beam and the interfacial surface, the meso-
scale topology (in the form of the gap-function) may be interpreted only as a shifted and/or rotated representation of its
actual counterpart. Also, an increased resolution scanning apparatus such as a confocal microscope [81–83] would also
greatly improve the confidences in the correlation analyses.
In the computational modeling side, the system must be investigated using transient and steady state harmonic methods
in order to gain a deeper understanding of the ‘‘true” nonlinear dynamical behavior of the system in an actual application
scenario. A more consistent rough contact formulation arising either out of a reduced non-parametric representation of
the expectation statistics, or through controlled machining processes yielding more repeatable roughness distributions,
could greatly improve the connect between model and reality. Further, in the asperity level, the employed constitutive
model corresponds to a simple Hertzian contact case in the current work. More advanced models could be employed so
as to capture the influence of adhesion (see Section 3.2.1), plasticity, fracture, etc., that could reduce epistemic uncertainties.
Another key drawback in the current formulation is that no relationships have been employed for fixing the coefficient of
friction l. More physics-based methods of estimating this parameter will greatly reduce the epistemic uncertainties in
the formulation, since its influence on the dissipation properties has been demonstrated to be extremely significant. Appli-
cation of the friction modeling approach for large displacement contact problems has to be investigated in detail since the
simplified ZTE formulation will no longer be valid for such cases and interpretations of the rough contact parameters will
have to be revised.
Lastly, it must be noted that most experimental studies of jointed systems document significant variability in the obser-
vations [77] (seen to a limited extent in Section 4.1), often cited to be due to friction or variations in preload. There have been
several stochastic modeling approaches applied to such systems with varying successes [84–86]. It would be relevant to con-
struct stochastic rough contact models in order to account for such variability from the perspective of uncertainty quantifi-
cation methods.
Going forward, generalizing the approach to minimize the use of exhaustive interfacial scans can expand the applicability
to scenarios where such scans are practically impossible. A possible approach to this would be to establish a consistent way
of working back the asperity distributions on a surface based on machining processes and documented surface finishes. Since
most industrial components are machined to conform to standard specifications, conducting parametric studies on the rela-
tionship of these specifications to the rough contact parameters would add significant value to such studies.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Acknowledgements

The authors thank Mianmian Ruan for helpful assistance for the bolt preparation; and Mengshi Jin for helpful discussions
regarding the ringdown data processing.
Funding: The authors are thankful for the support of the National Science Foundation under Grant No. 1744327 and the
China Scholarship Council (CSC).

Appendix A. Sources of Variability in the low-amplitude natural frequency

The current section seeks to provide an explanation about the offset in the natural frequency for low amplitudes where
the benchmark system seems to behave linearly. In Section 4.2.2 (see Table 1 and Fig. 19), the predicted low amplitude fre-
quency is 174.21 Hz while the experimental value is 180.92 Hz. Two sources of variability are investigated here: (a) addi-
tional stiffening provided by the bungee supports; and (b) variability in the bolt strains measured from the strain gauges.
For studying the amount of frequency offset introduced by the bungee cables, the appropriate regions of the assembly are
first identified as surface sets on the finite element model (see Fig. 23) and then virtual nodes are used to represent these
using distributing coupling elements. These virtual nodes are connected to the ground using linear stiffnesses estimated
assuming that the bungee cords themselves may be approximated as rigid during operation. Since the total mass of the
assembly is 3.68 kg and each bungee cord, suspended at about two feet length (0.6096 m), supports approximately half
of the total weight, the (low-amplitude oscillatory) pendulum stiffness is

kpend ¼ m2a gl u30Nm
1 : ð48Þ

As the focus of this study is to determine the effect of the bungee cord stiffness on the natural frequencies of the flexible body
modes, a simplified analysis is utilized in which the interface is modeled with hard contact and ideal Coulomb friction in
ABAQUS. Table 3 shows the natural frequencies of the first 10 elastic modes of the system (mode 1 has been the focus of
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 29

Fig. 23. Surface set used to enforce the bungee stiffness as boundary condition.

Table 3
Influence of boundary condition model on linearized natural frequencies.

Free-Free B.C. Bungee cord B.C.

Mode Natural freq (Hz) Natural freq (Hz)
1 167.42 167.42
2 227.31 227.31
3 585.22 585.22
4 666.10 666.10
5 1194.7 1194.7
6 1272.5 1272.5
7 1608.0 1608.0
8 1685.1 1685.1
9 2030.8 2030.8
10 2862.6 2862.6

Table 4
Influence of bolt prestress on the linearized natural frequency and Percentage area in contact for the non-flat + measured roughness and the flat + measured
roughness models. The case used for all the studies in the paper (corresponding to 2050 l) is boxed in black.

Bolt strain Bolt Load Non-flat + measured rough surface Flat + measured rough surface
(l) (kN) Nat. Freq. (Hz) Contact area (%) Nat. Freq. (Hz) Contact area (%)
1050 172.62 62.16 165.34 86.7904
1250 170.13 67.41 167.17 86.7904
1550 173.39 72.27 166.46 87.8913
1750 174.36 74.35 167.72 88.5517

2050 174.21 77.59 168.98 89.8727

2250 174.07 79.45 168.57 90.3130
2500 173.33 81.05 168.77 91.1936
2750 173.44 81.72 169.10 91.1936
3000 172.54 82.47 170.12 91.1936

all the studies in this paper). It can be observed that there is no appreciable change to the first two decimal places, showing
that it is not possible to obtain a 6 Hz shift due to the bungees alone.
The second consideration, involving possible errors in the strain measurements, is investigated by conducting the lin-
earization about the prestress state of various bolt strain levels. Table 4 tabulates the mode 1 natural frequencies and per-
centage contact area for different bolt strain levels (in an approximately 50% band). Two cases of contact models developed
from the surface measurements, one with the non-flat surface with measured roughness, and the other with the assumption
that the nominal surface is perfectly flat. It can be seen that the variation in natural frequency with the prestress level in the
first case does not seem to be gradual or monotonic. For the second case however, the natural frequency is observed to be
monotonically increasing with the bolt prestress level until the frequency saturates at 25 kN. These observations may be cor-
roborated with the trends in the area in contact.
Fig. 24 depicts these trends graphically, wherein it may be observed that the non-flat interface nominally has much smal-
ler area in contact for any given prestress level. This shows that the exact topology of the interface has a highly non-trivial
and significant impact on the area in contact, which is extremely important to the response of the system. Erroneous meso-
surface identification is thus a factor that may not be ruled out as a possible cause for the offset seen.
30 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

Fig. 24. Influence of Bolt load (per bolt) on (a) linearized natural frequency; and (b) percentage area of contact. The case used for all the studies in the paper
(corresponding to 2050 l) is encircled in black.

References

[1] D.J. Segalman, D.L. Gregory, M.J. Starr, B.R. Resor, M.D. Jew, J.P. Lauffer, and N.M. Ames. Handbook on Dynamics of Jointed Structures. Technical Report
SAND2009-4164. Sandia National Laboratories, Albuquerque, NM, 2009..
[2] S.A. Nassar, A. Abboud, An improved stiffness model for bolted joints, J. Mech. Des. 131 (12) (2009) 121001, https://doi.org/10.1115/1.4000212, ISSN
1050-0472.
[3] S.A. Nassar, S. Ganeshmurthy, R.M. Ranganathan, G.C. Barber, Effect of tightening speed on the torque-tension and wear pattern in bolted connections,
J. Pressure Vessel Technol. 129 (3) (2006) 426–440, https://doi.org/10.1115/1.274929, ISSN 0094-9930.
[4] S.A. Nassar, P.H. Matin, G.C. Barber, Thread friction torque in bolted joints, J. Pressure Vessel Technol. 127 (4) (2005) 387–393, https://doi.org/10.1115/
1.2042474, ISSN 0094-9930.
[5] M. Ruan. The variability of strains in bolts and the effect on preload in jointed structures. Master’s thesis, Department of Mechanical Engineering, Rice
University, Houston, TX, May 2019..
[6] A. Pazouki, M. Kwarta, K. Williams, W. Likos, R. Serban, P. Jayakumar, D. Negrut, Compliant contact versus rigid contact: a comparison in the context of
granular dynamics, Phys. Rev. E 96 (4) (2017), https://doi.org/10.1103/PhysRevE.96.042905, ISSN 2470-0045, 2470-0053.
[7] B. Seeger, P. Butaud, F. Du, V. Baloglu, M.R.W. Brake, C.W. Schwingshackl. In situ measurements of interfacial contact pressure during impact hammer
tests, in: 36th International Modal Analysis Conference (IMAC XXXVI), Orlando, FL, February 2018..
[8] Wei Chen, Mengshi Jin, Iyabo Lawal, Matthew R.W. Brake, Hanwen Song, Measurement of slip and separation in jointed structures with non-flat
interfaces, Mech. Syst. Signal Process. 134 (2019) 106325.
[9] S.W.B Klaassen, M. Brøns, G. Chauda, T.A. Kasper, C.W. Schwingshackl, M.R.W. Brake. Optical full field monitoring of bolted lap-joint behaviour under
vibration, in: 37th International Modal Analysis Conference (IMAC XXXVII), Orlando, FL, January 2019..
[10] L. Bureau, T. Baumberger, C. Caroli, Contact mechanics for randomly rough surfaces, Eur. Phys. J. E 19 (2) (2006) 163–169, https://doi.org/10.1140/epje/
e2006-00019-2, ISSN 1292-8941, 1292–895X. http://arxiv.org/abs/cond-mat/0510232. arXiv: cond-mat/0510232.
[11] H. Ghaednia, X. Wang, S. Saha, Y. Xu, A. Sharma, R.L. Jackson, A review of elastic-plastic contact mechanics, Appl. Mech. Rev. 69 (6) (2017), https://doi.
org/10.1115/1.4038187, 060804–060804-30, ISSN 0003–6900.
[12] X. Lei, How do asperities fracture? An experimental study of unbroken asperities, Earth Planet. Sci. Lett. 213 (3) (2003) 347–359, https://doi.org/
10.1016/S0012-821X(03)00328-5, ISSN 0012-821X.
[13] A. Fantetti, L.R. Tamatam, M. Volvert, I. Lawal, L. Liu, L. Salles, M.R.W. Brake, C.W. Schwingshackl, D. Nowell, The impact of fretting wear on structural
dynamics: experiment and simulation, Tribol. Int. 138 (2019) 111–124.
[14] M. Krack, L. Salles, F. Thouverez, Vibration prediction of bladed disks coupled by friction joints, Arch. Comput. Methods Eng. 24 (3) (2017) 589–636,
https://doi.org/10.1007/s11831-016-9183-2, ISSN 1134-3060, 1886–1784.
[15] R.M. Lacayo, M.S. Allen, Updating structural models containing nonlinear Iwan joints using quasi-static modal analysis, Mech. Syst. Signal Process. 118
(2019) 133–157, https://doi.org/10.1016/j.ymssp.2018.08.034, ISSN 08883270.
[16] K.L. Johnson, Contact Mechanics, Cambridge University Press, Cambridge, 1985.
[17] M.R.W. Brake, An analytical elastic plastic contact model with strain hardening and frictional effects for normal and oblique impacts, Int. J. Solids
Struct. 62 (2015) 104–123.
[18] A. Signorini, Questioni di elasticit non linearizzata e semilinearizzata (Topics in non linear and semi linear elasticity), Rendiconti di Matematica e delle
sue Applicazioni 5 (18) (1959) 95–139.
[19] G. Fichera, Boundary Value Problems of Elasticity with Unilateral Constraints, in: C. Truesdell (Ed.), Linear Theories of Elasticity and Thermoelasticity,
Springer, Berlin Heidelberg, Berlin, Heidelberg, 1973, pp. 391–424, https://doi.org/10.1007/978-3-662-39776-3_4, ISBN 978-3-662-38853-2 978-3-
662-39776-3.
[20] R.L. Jackson, I. Green, On the modeling of elastic contact between rough surfaces, Tribol. Trans. 54 (2011) 300–314.
[21] J.A. Greenwood, J.B.P. Williamson, Contact of nominally flat surfaces, Proc. R. Soc. London, Series A 295 (1966) 300–319.
[22] S. Medina, D. Nowell, D. Dini, Analytical and numerical models for tangential stiffness of rough elastic contacts, Tribol. Lett. 49 (1) (2013) 103–115,
https://doi.org/10.1007/s11249-012-0049-y, ISSN 1023-8883, 1573–2711.
[23] C. Campana, B.N.J. Persson, M.H. Müser, Transverse and normal interfacial stiffness of solids with randomly rough surfaces, J. Phys.: Condensed Matter
23 (2011) 085001.
[24] M. Eriten, A.A. Polycarpou, L.A. Bergman, Physics-based modeling for fretting behavior of nominally flat rough surfaces, Int. J. Solids Struct. 48 (2011)
1436–1450.
[25] M. Eriten, A.A. Polycarpou, L.A. Bergman, Physics-based modeling for partial slip behavior of spherical contacts, Int. J. Solids Struct. 47 (18) (2010)
2554–2567, https://doi.org/10.1016/j.ijsolstr.2010.05.017, ISSN 0020-7683. DOI: http://www.sciencedirect.com/science/article/pii/
S0020768310001952.
[26] E. Popova, V.L. Popov, The research works of Coulomb and Amontons and generalized laws of friction, Friction 3 (2) (2015) 183–190, https://doi.org/
10.1007/s40544-015-0074-6, ISSN 2223-7690, 2223–7704.
[27] A. Visintin. The Science of Hysteresis Volume 1, chapter Mathematical Models of Hysteresis, pages 1–123. Elsevier Inc., 2005..
 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615 31

[28] M.A. Janaideh, R. Naldi, L. Marconi, P. Krejci, A hybrid model for the play hysteresis operator, Phys. B (2013).
[29] G.M. Jenkins, Analysis of the stress-strain relationships in reactor grade graphite, Br. J. Appl. Phys. 13 (1962) 30–32.
[30] S. Zucca, C.M. Firrone, M.M. Gola, Modeling underplatform dampers for turbine blades: a refined approach in the frequency domain, J. Vib. Control 19
(2013) 1087–1102.
[31] A.T. Mathis, N.N. Balaji, R.J. Kuether, A.R. Brink, M.R.W. Brake, and D.D. Quinn. A review of damping models for structures with mechanical joints.
Applied Mechanics Reviews, Under review..
[32] M.L. Raffa, F. Lebon, G. Vairo, Normal and tangential stiffnesses of rough surfaces in contact via an imperfect interface model, Int. J. Solids Struct. 87
(2016) 245–253, https://doi.org/10.1016/j.ijsolstr.2016.01.025, ISSN 00207683.
[33] C. Putignano, L. Afferrante, G. Carbone, G. Demelio, A new efficient numerical method for contact mechanics of rough surfaces, Int. J. Solids Struct. 49
(2012) 338–343.
[34] S. Akarapu, T. Sharp, M.O. Robbins, Stiffness of contacts between rough surfaces, Phys. Rev. Lett. 106 (2011) 204301.
[35] W.R. Chang, I. Etsion, D.B. Bogy, An elastic-plastic model for the contact of rough surfaces, ASME J. Tribol. 109 (1987) 257–263.
[36] W.D. Iwan, On a class of models for the yielding behavior of continuous and composite systems, ASME J. Appl. Mech. 34 (1967) 612–617.
[37] D.J. Segalman, A four-parameter Iwan model for lap-type joints, J. Appl. Mech. 72 (5) (2005) 752–760.
[38] R.E. Goodman, R.L. Taylor, T.L. Brekke, A model for the mechanics of jointed rock, J. Soil Mech. Found. Division 94 (3) (1968) 637–660.
[39] H. Festjens, G. Chevallier, J.-L. Dion, A numerical tool for the design of assembled structures under dynamic loads, Int. J. Mech. Sci. 75 (2013) 170–177.
[40] N.N. Balaji, M.R.W. Brake, Systems and Signal Processing 126 (2019) 42–64, https://doi.org/10.1016/j.ymssp.2019.02.013, ISSN 0888-3270.
[41] M. Jin, M.R.W. Brake. Nonlinear system identification methods for jointed structures, in: 37th International Modal Analysis Conference (IMAC XXXVII),
Orlando, FL, January 2019..
[42] M. Jin, M.R.W. Brake, H. Song. Comparison of nonlinear system identification methods for free decay measurements with application to jointed
structures. J. Sound Vib., In press..
[43] M. Mayer, L. Gaul, Modeling of contact interfaces using segment-to-segment-elements for FE vibration analysis, in: 23rd International Modal Analysis
Conference (IMAC XXIII), Bethel, CT, 2005.
[44] R.D. Mindlin, Compliance of elastic bodies in contact, ASME J. Appl. Mech. 16 (1949) 259–268.
[45] R.D. Mindlin, H. Deresiewicz, Elastic spheres in contact under varying oblique forces, ASME J. Appl. Mech. 20 (1953) 327–344.
[46] C. Cattaneo. Sul contatto di due corpi elastici: Distribuzione locale degli sforzi. Rendiconti dell’Accademia Nazionale dei Lincei, 27: 342–348, 434–436,
474–478, 1938..
[47] M. Afzal, I. Lopez Arteaga, L. Kari, An analytical calculation of the Jacobian matrix for 3d friction contact model applied to turbine blade shroud contact,
Comput. Struct. 177 (2016) 204–217, https://doi.org/10.1016/j.compstruc.2016.08.014, ISSN 00457949.
[48] K.D. Hjelmstad, Fundamentals of structural mechanics, Springer Science & Business Media, 2007.
[49] M.R.W. Brake (Ed.), The Mechanics of Jointed Structures, Springer, 2017.
[50] B. Yang, T.A. Laursen, X. Meng, Two dimensional mortar contact methods for large deformation frictional sliding, Int. J. Numer. Meth. Eng. 62 (9) (2005)
1183–1225, https://doi.org/10.1002/nme.1222, ISSN 0029-5981, 1097–0207.
[51] L. Pesaresi, J. Armand, C.W. Schwingshackl, L. Salles, W. Wong, An advanced underplatform damper modelling approach based on a microslip contact
model, Journal of Sound and Vibration (2018), https://doi.org/10.1016/j.jsv.2018.08.014, ISSN 0022-460X.
[52] K.G. Sharma, C.S. Desai, Analysis and implementation of thin-layer element for interfaces and joints, J. Eng. Mech. 118 (12) (1992) 2442–2462, https://
doi.org/10.1061/(ASCE)0733-9399(1992)118:12(2442), ISSN 0733-9399, 1943–7889.
[53] X. Wang, L.B. Wang, Continuous interface elements subject to large shear deformations, Int. J. Geomech. 6 (2) (2006) 97–107, https://doi.org/10.1061/
(ASCE)1532-3641(2006)6:2(97), ISSN 1532-3641, 1943–5622.
[54] P. Wriggers, J. Schröder, A. Schwarz, A finite element method for contact using a third medium, Comput. Mech. 52 (4) (2013) 837–847, https://doi.org/
10.1007/s00466-013-0848-5, ISSN 0178-7675, 1432–0924.
[55] G. Beer, An isoparametric joint/interface element for finite element analysis, Int. J. Numer. Meth. Eng. 21 (4) (1985) 585–600, https://doi.org/10.1002/
nme.1620210402, ISSN 0029-5981, 1097–0207.
[56] D.A.H. Hanaor, Y. Gan, I. Einav, Contact mechanics of fractal surfaces by spline assisted discretisation, Int. J. Solids Struct. 59 (2015) 121–131, https://
doi.org/10.1016/j.ijsolstr.2015.01.021, ISSN 00207683.
[57] R. Buczkowski, M. Kleiber, Elasto-plastic interface model for 3d-frictional orthotropic contact problems, Int. J. Numer. Meth. Eng. 40 (4) (1997) 599–
619, https://doi.org/10.1002/(SICI)1097-0207(19970228)40:4<599::AID-NME81>3.0.CO;2-H, ISSN 1097-0207.
[58] R. Jedynak, Exact and approximate solutions of the infinite integrals of the asperity height distribution for the Greenwood-Williamson and the
Greenwood-Tripp asperity contact models, Tribol. Int. 130 (2019) 206–215, https://doi.org/10.1016/j.triboint.2018.09.009, ISSN 0301679X.
[59] M. Eriten, D.T. Petlicki, A.A. Polycarpou, L.A. Bergman, Influence of friction and adhesion on the onset of plasticity during normal loading of spherical
contacts, Mech. Mater. 48 (2012) 26–42, https://doi.org/10.1016/j.mechmat.2012.01.003, ISSN 0167-6636.
[60] L. Gaul, M. Mayer. Efficient modelling of contact interfaces of joints in built-up structures, in: Computer Methods and Experimental Measurements for
Surface Effects and Contact Mechanics VIII, volume I, pages 195–205, The New Forest, UK, May 2007. WIT Press. ISBN 978-1-84564-073-6. DOI:
10.2495/SECM070191. http://library.witpress.com/viewpaper.asp?pcode=SECM07-019-1..
[61] C. Siewert, L. Panning, J. Wallaschek, C Richter, Multiharmonic forced response analysis of a turbine blading coupled by nonlinear contact forces, J. Eng.
Gas Turbines Power 132 (8) (2010) 082501, https://doi.org/10.1115/1.4000266, ISSN 07424795.
[62] J. Králikowski, J. Szczepek, Assessment of tangential and normal stiffness of contact between rough surfaces using ultrasonic method, Wear 160 (2)
(1993) 253–258, https://doi.org/10.1016/0043-1648(93)90428-O, ISSN 00431648.
[63] N. Yoshioka, C.H. Scholz, Elastic properties of contacting surfaces under normal and shear loads: 1 Theory, J. Geophys. Res. 94 (B12) (1989) 17681,
https://doi.org/10.1029/JB094iB12p17681, ISSN 0148-0227.
[64] N. Yoshioka, C.H. Scholz, Elastic properties of contacting surfaces under normal and shear loads: 2 Comparison of theory with experiment, J. Geophys.
Res. 94 (B12) (1989) 17691, https://doi.org/10.1029/JB094iB12p17691, ISSN 0148-0227.
[65] H.A. Sherif, S.S. Kossa, Relationship between normal and tangential contact stiffness of nominally flat surfaces, Wear 151 (1) (1991) 49–62, https://doi.
org/10.1016/0043-1648(91)90345-U, ISSN 0043-1648.
[66] M. Gonzalez-Valadez, A. Baltazar, R.S. Dwyer-Joyce, Study of interfacial stiffness ratio of a rough surface in contact using a spring model, Wear 268 (3)
(2010) 373–379, https://doi.org/10.1016/j.wear.2009.08.022, ISSN 0043-1648.
[67] A. Poganik, M. Kalin, How to determine the number of asperity peaks their radii and their heights for engineering surfaces: A critical appraisal, Wear
300 (1–2) (2013) 143–154, https://doi.org/10.1016/j.wear.2013.01.105, ISSN 00431648.
[68] A. Horwitz. Ellipses Inscribed in Parallelograms. arXiv:0808.0297 [math], August 2008. http://arxiv.org/abs/0808.0297. arXiv: 0808.0297..
[69] A. Horwitz. Dynamics of ellipses inscribed in quadrilaterals. arXiv:1505.01048 [math], May 2015. http://arxiv.org/abs/1505.01048. arXiv: 1505.01048..
[70] B. Braden, The Surveyor’s area formula, College Mathematics J. 17 (4) (1986) 326–337, https://doi.org/10.1080/07468342.1986.11972974, ISSN 0746-
8342.
[71] O.C. Zienkiewicz, R.L. Taylor, The Finite Element Method for Solid and Structural Mechanics, The Finite Element Method, Elsevier Science, 2013, ISBN
9780080951362.
[72] L. Kogut, I. Etsion, Adhesion in elastic-plastic spherical microcontact, J. Colloid Interface Sci. 261 (2) (2003) 372–378, https://doi.org/10.1016/S0021-
9797(03)00071-7, ISSN 0021-9797.
[73] W.R. Chang, I. Etsion, D.B. Bogy, Adhesion model for metallic rough surfaces, J. Tribol. 110 (1) (1988) 50–56, https://doi.org/10.1115/1.3261574, ISSN
0742-4787, 1528–8897.
32 N.N. Balaji et al. / Mechanical Systems and Signal Processing 139 (2020) 106615

[74] A. Ovcharenko, G. Halperin, I. Etsion, M. Varenberg, A novel test rig for in situ and real time optical measurement of the contact area evolution during
pre-sliding of a spherical contact, Tribol. Lett. 23 (1) (2006) 55–63, https://doi.org/10.1007/s11249-006-9113-9, ISSN 1023-8883, 1573–2711.
[75] E. Rabinowicz, Influence of surface energy on friction and wear phenomena, J. Appl. Phys. 32 (8) (1961) 1440–1444, https://doi.org/10.1063/
1.1728375, ISSN 0021-8979.
[76] R.G. Budynas, K. Nisbett, Shigley’s mechanical engineering design, 10th ed., McGraw Hill, 2014.
[77] M.R.W. Brake, C.W. Schwingshackl, P. Reuß, On the observed variability and repeatability in jointed structures, Mech. Syst. Signal Process. 129 (2019)
282–307.
[78] B.J. Deaner, M.S. Allen, M.J. Starr, D.J. Segalman, H. Sumali, Application of viscous and Iwan modal damping models to experimental measurements
from bolted structures, ASME J. Vib. Acoust. 137 (2015) 021012.
[79] D.J. Segalman, M.S. Allen, M. Eriten, K. Hoppman, Experimental assessment of joint-like modal models for structures, in: ASME International Design
Engineering Technical Conferences IDETC/CIE, Boston, MA, 2015.
[80] C. Spearman, The proof and measurement of association between two things, Am. J. Psychol. 15 (1) (1904) 72, https://doi.org/10.2307/1412159, ISSN
00029556.
[81] R. Leach, Optical Measurement of Surface Topography, vol. 14, Springer, 2011.
[82] H.-J. Jordan, M. Wegner, H. Tiziani, Highly accurate non-contact characterization of engineering surfaces using confocal microscopy, Meas. Sci. Technol.
9 (7) (1998) 1142–1151, https://doi.org/10.1088/0957-0233/9/7/023, ISSN 0957–0233.
[83] D.A. Lange, H.M. Jennings, S.P. Shah, Analysis of surface roughness using confocal microscopy, J. Mater. Sci. 28 (14) (1993) 3879–3884, https://doi.org/
10.1007/BF00353195, ISSN 1573-4803.
[84] R. Murthy, B.K. Choi, X.Q. Wang, M.C. Sipperley, M.P. Mignolet, C. Soize, Maximum entropy modeling of discrete uncertain properties with application
to friction, Probab. Eng. Mech. 44 (2016) 128–137.
[85] X.Q. Wang, M.P. Mignolet. Stochastic Iwan-type model of a bolted joint: Formulation and identification. In 32nd International Modal Analysis
Conference (IMAC XXXII), Orlando, FL, February 2014..
[86] M.P. Mignolet, C. Soize, Stochastic reduced order models for uncertain geometrically nonlinear dynamical systems, Comput. Methods Appl. Mech. Eng.
197 (2008) 3951–3963.