Mechanical Systems and Signal Processing 100 (2018) 550–569

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

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

Jacobian projection reduced-order models for dynamic systems

with contact nonlinearities
Chiara Gastaldi a,⇑, Stefano Zucca a, Bogdan I. Epureanu b
a
Department of Mechanical and Aerospace Engineering, Politecnico di Torino, Corso Duca degli Abruzzi 24, Torino, Italy
b
Department of Mechanical Engineering, University of Michigan, 2350 Hayward St., Ann Arbor, MI, USA

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

Article history: In structural dynamics, the prediction of the response of systems with localized nonlinear-
Received 23 March 2017 ities, such as friction dampers, is of particular interest. This task becomes especially cum-
Received in revised form 13 July 2017 bersome when high-resolution finite element models are used. While state-of-the-art
Accepted 27 July 2017
techniques such as Craig-Bampton component mode synthesis are employed to generate
reduced order models, the interface (nonlinear) degrees of freedom must still be solved
in-full. For this reason, a new generation of specialized techniques capable of reducing lin-
Keywords:
ear and nonlinear degrees of freedom alike is emerging. This paper proposes a new tech-
Friction
Nonlinear dynamics
nique that exploits spatial correlations in the dynamics to compute a reduction basis.
Reduced-order models The basis is composed of a set of vectors obtained using the Jacobian of partial derivatives
Jacobian of the contact forces with respect to nodal displacements. These basis vectors correspond
Harmonic balance method to specifically chosen boundary conditions at the contacts over one cycle of vibration. The
Forced response technique is shown to be effective in the reduction of several models studied using multi-
ple harmonics with a coupled static solution.
In addition, this paper addresses another challenge common to all reduction techniques:
it presents and validates a novel a posteriori error estimate capable of evaluating the qual-
ity of the reduced-order solution without involving a comparison with the full-order
solution.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction

The dynamics of structures constrained through frictional contacts can be very complex due to the nonlinear nature of dry
friction [1]. Often, the response is assumed to be periodic so that harmonic balance methods (HBM) can be used in the fre-
quency domain. Methods to compute the steady-state response can use HBM together with alternating frequency-time (AFT)
approaches for systems with friction contacts [2–6].
Spatial reduction techniques like the proper orthogonal decomposition can be used to model the physical DOFs of the
structure (i.e. the nodal DOFs of the FE model) as a linear superposition of shape functions [7]. However these techniques
require the solution of the full system to be computed at given time instants, and this may be unfeasible for
high-resolution FE models. Another approach which also requires the nonlinear system to be solved in advance in order
to generate the ROM is described for turbine bladed disks applications in [8].

⇑ Corresponding author.
E-mail address: chiara.gastaldi@polito.it (C. Gastaldi).

http://dx.doi.org/10.1016/j.ymssp.2017.07.049
0888-3270/Ó 2017 Elsevier Ltd. All rights reserved.
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 551

Nomenclature

Abbreviations
AFT Alternating Frequency Time method
CB-CMS Craig Bampton-Component Mode Synthesis
DoF(s) Degree(s) Of Freedom
FFT Fast Fourier Transform
FE Finite Elements
FO Full Order
FR Full Reduction
HBM Harmonic Balance Method
IFFT Inverse Fast Fourier Transform
JP Jacobian Projection
NLR NonLinear Reducion
RO(M) Reduced Order (Model)

Vectors and matrices

D dynamic stiffness matrix
f vector of forces
U reduction basis
g vector of forces in reduced coordinates - FR
H reduced dynamic stiffness matrix - NLR
J Jacobian matrix (multi-harmonic)
b
K multi-harmonic matrix of partial derivatives of contact forces with respect to displacements
l vector of forces in reduced coordinates - NLR
KJC diagonal matrix of eigenvalues
M; C; K mass, structural damping and stiffness matrices
MJ ; CJ ; KJ multi-harmonic mass, structural damping and stiffness matrices, harmonics: 0-H
MJC ; CJC ; KJC multi-harmonic mass, structural damping and stiffness matrices after static condensation, harmonics: 1-H
(after static condensation)
p vector of displacements in reduced coordinates - FR
W mode used to compute the boundary conditions necessary to the definition of U
q vector of displacements
R rotation matrix
r multi-harmonic vector of residuals
s vector of displacements in reduced coordinates - NLR
S diagonal matrix of singular values
U; V matrices of left and right singular vectors respectively
VJC multi-harmonic eigenvectors matrix

Additional subscripts
cond conditioned
C contact
E external
i iteration step
ST relative to a full stick linear system

relative to a posteriori error estimate
L relative to linear DoFs
N relative to nonlinear DoFs
ROM relative to the reduced order model

Additional superscripts
h
harmonic index h, both real and imaginary part
h;I
harmonic index h, imaginary part
h;R
harmonic index h, real part
ðÞ
referred to a complex-valued quantity where real and imaginary parts have been split into separate entries of
a matrix or vector
T
transpose
552 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

Other variables
a; b generic scalar quantities
a dimensionless parameter that identifies the external excitation level imposed to the system
E Young’s modulus
ED Dissipated energy


ehfc ; ehfc actual and predicted error for a given harmonic order
f ð. . .Þ generic function of.
g ratio between minimum and maximum singular value
h a given harmonic index
H the maximum number of harmonics retained in the nonlinear calculation
ku1 ; ku2 ; kv tangential and normal contact stiffness values
l friction coefficient
NDoF number of physical degrees of freedom
m Poisson’s ratio
nh size of the reduction basis for a given harmonic order Uh
x frequency of vibration (rad/s)
^
p amplification factor used to compute the reduction basis
q density
u1 ; u2 ; v local coordinate system of the contact patch

In addition, linear DoFs (i.e., ones not on contact surfaces) may be reduced by techniques such as Craig-Bampton compo-
nent mode synthesis (CB-CMS) [9,10]. However, the remaining set of nonlinear equations (involving nonlinear/contact DoFs)
must be solved in full and that requires the highest computational effort. Hence, next generation of model reduction tech-
niques capable of reducing the number of nonlinear and linear equations are an important current need [11–13]. This work
addresses this challenge and presents a novel method to reduce model sizes for computing forced responses of structures
with dry friction contacts.
The proposed Jacobian-based reduction technique uses a set of linear systems with specifically chosen boundary
conditions at the contact over one cycle of vibration. These boundary conditions are applied to compute analytically
a Jacobian of partial derivatives of the contact forces with respect to displacements. The Jacobian is used together
with the (linear) stiffness and the mass matrices of the system to obtain the dominant eigenspace for the frequency
range of interest. A set of basis vectors in this space forms the reduction basis to construct reduced order models
(ROMs).
This technique, here termed Jacobian Projection (JP), involves the solution of linear systems and simple vector
operations. Unlike existing techniques [11–13], the JP uses a multi-harmonic modal framework, as a result it is capable
of intrinsically taking into account the higher-harmonic generation and the static-harmonic influence that friction
contacts introduce. These features enable model reduction for systems studied using multiple harmonics with a coupled
static solution. Simulated responses of full-order models and ROMs are studied under different excitation levels, giving
rise to different conditions of stick, slip and separation at the contacts. The reduction technique is shown to be effective
and to achieve relevant time savings without significant loss of accuracy. A preliminary version of this work has
been presented in [14]. This paper extends the work in [14] in three important directions: the application of the
proposed technique to multi-harmonic balance calculations, the investigation of the influence of the number of
nonlinear degrees of freedom on the technique’s performance and an extensive analysis of the computational savings
introduced.
Finally a method to evaluate the accuracy of the reduced order (RO) solution is proposed. Most reduction techniques (e.g.
[8,11–13]) rely on the comparison between RO and full order (FO) solutions to prove their effectiveness. The comparison
with the FO solution is essential at the validation stage. However, it should not be relied upon for further verifications. In
fact reduction techniques are designed for systems whose FO solutions are not attainable in a feasible amount of time. Usu-
ally users rely upon cumbersome convergence studies, where the size of the reduction basis is progressively increased until
the resulting solutions cease to change significantly. These authors propose an a posteriori error estimate, capable of quan-
tifying the effectiveness of the reduction basis independently from comparisons with the FO or with other RO solutions. As a
result the user will improve the basis and rerun the calculation only if strictly necessary.

2. Methodology

This section describes first the state-of-the-art techniques and methods used to obtain the FO frequency response func-
tion of a frictionally-constrained structure. Next, these same techniques are modified and applied to the novel type of ROMs
obtained using the JP technique described in Section 3.
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 553

2.1. Background equations

The dynamic equations describing the vibration of a frictionally-constrained elastic structure may be represented in the
time domain as:
€ þ Cq_ þ Kq ¼ f E þ f C ðqÞ;
Mq ð1Þ
where, M and K are the ‘‘free” mass and stiffness matrices of the structure with no contact conditions enforced, i.e. the con-
tact nodes are considered not to be in contact when developing these matrices. C is the linear damping matrix of the free
system. Further details on how these matrices were obtained for this specific study are given in Appendix A.1. f E is the vector
of excitation forces applied to the model. The contact conditions are enforced through the vector of contact forces f C which
arise due to Coulomb friction nonlinearities and are determined as a function of displacements using a suitable contact
model (see Appendix A.2).
For cases of periodic excitations of frequency x which lead to periodic responses, the steady state solution of Eq. (1) can
be obtained using a HBM [2,3]. In the HBM, the steady state displacements and nonlinear forces are assumed to be periodic.
Hence, they are expressed as a sum of harmonic terms to obtain a set of algebraic balance equations in the frequency
domain:
h h
Dh qh ¼ f E þ f C ; 0 6 h 6 H; ð2Þ

where the dynamic stiffness matrix Dh can be expressed as:

Dh ¼ ðhxÞ M þ ihxC þ K:
2
ð3Þ
Note that the static and the dynamic governing equations in Eq. (2) are coupled by the Fourier coefficients of the nonlin-
h
ear contact forces. In fact, apart for limit linear cases (full stick) where f C ¼ f ðqh Þ 8h, the nonlinearity induced by friction
h
introduces cross-harmonic coupling, i.e. fC ¼ f ðq0 ; . . . ; qh ; . . . qH Þ.

2.2. Iterative algorithm

Solution of Eq. (2) can be computed using an iterative scheme (e.g., Newton-Raphson, Dogleg Reflective, Levenberg-
Marquardt). Iterative procedures require Eq. (2) to be reformulated as follows:
h h
rh ¼ Dh qh  f E  f C 0 6 h 6 H; ð4Þ
h
where r is the residual vector of the systems of equations.
h h
It should be noted that rh ; qh ; f E and f C are complex-valued vectors for hP1, e.g. qh ¼ qh;R þ iqh;I . Since iterative solvers
cannot handle the minimization of complex quantities, they are split into their real and imaginary components,1 as in:
 T
qh ¼ qh;R ; qh;I ; ð5Þ
As a result Eq. (4) is reformulated as:
0 0
r0 ¼ D0 q0  f E  f C ;
ð6Þ
rh ¼ Dh qh  f hE  f hC 1 6 h 6 H;

where Dh is the dynamic stiffness matrix from Eq. (3), reorganized accordingly:
" #
ðhxÞ M þ K hxC
2
D ¼
h
1 6 h 6 H: ð7Þ
hx C ðhxÞ M þ K
2

The above-mentioned iterative schemes generate approximate solutions which converge toward the solution. As an exam-
ple, consider the Newton-Raphson Method, where the approximate solution at the i-th step is estimated as:

qji ¼ qji1  J1 ji1  rji1 ; ð8Þ

where

 T  T
 qji ¼ q0 ; . . . ; qh ; . . . ; qH ji ¼ q0 ; . . . ; qh;R ; qh;I ; . . . ; qH;R ; qH;I ji is the approximate solution at the i-th iteration;
 Jji is the Jacobian matrix at the i-th iteration. The Jacobian matrix can be expressed as:
^
J ¼ D  K; ð9Þ

1
The complex notation, i.e. qh ¼ qh;R þ iqh;I is here preferred because it is more compact. The notation will be expanded, i.e. see Eq. (5), only when dealing
with partial derivatives.
554 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

with
2 3
D0 0 ... 0
6 7
6 0 D1 ... 0 7
D¼6
6 .. 7 ¼ x2 MJ þ xCJ þ KJ ¼
7
4 . 5
0 0 . . . DH
2 3
0 0 0 ... 0 0
60 M 0 ... 0 7
6 0 7
6 7
60 0 M ... 0 0 7
6 7
¼ x2 6 .. 7þ
6 . 7
6 7 ð10Þ
6 7
4 0 0 0 . . . H2 M 0 5
0 0 0 ... 0 H2 M
2 3
0 0 0 ... 0 0
6 0 0 C . . . 0 7
6 0 7
6 7
60 C 0 ... 0 0 7
6 7
þ x6 .. 7 þ diagð½K; K; K; . . . ; K; KÞ;
6 . 7
6 7
6 7
40 0 0 ... 0 HC 5
0 0 0 . . . HC 0

2 3
^ 0;0 K
K ^ 0;1 ... ^ 0;H
K
6 ^ 1;0 ^ 1;1 ^ 1;H 7
6K ... K 7 0
^ 0;0 ¼ @f C ;
K
^ ¼6
K 7 with K
6 .. 7 @q0
4 . 5
^ H;0 K
K ^ H;1 ... K^ H;H
h 0 0 i
^ ¼
K0;j @f C @f C ; 8j 2 ½1; H;
@qj;R @qj;R
2 h;R 3 ð11Þ
@f C
^ h;0 6 @q0 7
K ¼4 5; 8h 2 ½1; H and
@f h;I
C
@q0
2 3
@f h;R
C
@f h;R
C
^ h;j ¼ 6
K 4
@qj;R @qj;I 7
5; 8h; j 2 ½1; H;
@f h;I
C
@f h;I
C
@qj;R @qj;I

 T  T
 rji ¼ r0 ; . . . ; rh ; . . . ; rH ji ¼ r0 ; . . . ; rh;R ; rh;I ; . . . ; rH;R ; rH;I ji is the vector of residuals at the i-th iteration.

Matrix K b can be considered as a cross-harmonic stiffness matrix since each block-entry Kb h;k quantifies the contribution
that the k-th harmonic component of displacements has on the h-th harmonic component of the contact forces. The struc-
ture of the matrix depends on the degree of nonlinearity introduced by friction. In a perfectly linear case such as full stick,
matrix K b is block-diagonal and no cross-harmonic coupling occurs. If, on the other hand, some slipping occurs, the off-
diagonal blocks of K b will not be empty anymore and cross-harmonic coupling will occur.
Section 3.2 shows that the JP reduction technique is capable of taking into account friction-induced cross-harmonic cou-
pling by exploiting the structure of the Jacobian matrix under different contact conditions.
Iterations end when krk < , for some small value of . It should be noted that, since iterative solvers use partial deriva-
tives (i.e. Jacobian), need to treat real and imaginary part of complex vectors separately. Therefore, the size of the system, as
shown in Eq. (10), goes from NDoF  ðH þ 1Þ (‘‘physical DoFs”) to NDoF  2  H þ 1 since complex force and displacement vectors
have to be separated into their real and imaginary components.
Convergence rates may be greatly improved by using accurate Jacobians, such as the ones obtained using the analytical
Jacobian formulations. This requires the contact model to output the derivatives of the Fourier coefficients of the contact
forces with respect to the Fourier coefficients of the relative displacements as shown in [1].
The Jacobian can be used to approximate the solution. If r  0, then Eq. (2) (and Eq. (16)) can be rewritten in a linearized
version as:

Jq ¼ f E ; ð12Þ
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 555

where
h i
0
f E ¼ f E ; f 1E ; . . . ; f HE : ð13Þ
h i
0
In other words, with reference to Eq. (11), the vector of contact forces f C ¼ f C ; . . . ; f H
C has been substituted with

b
f C ¼ Kq: ð14Þ

2.3. Coordinate reduction

The nonlinearity due to the contact is localized because it involves only a small portion of the structure (i.e., a subset of all
DoFs). Thus, the governing equations can be conveniently re-written as:
" #" # 2 3" #
h
DhN;N DhN;L qhN f E;N h
¼4 5 þ f C;N ; 8h 2 ½0; H; ð15Þ
DhL;N DhL;L qhL h
f E;L 0

where qhN is the vector of nonlinear co-ordinates which contains the displacements of the contact nodes. The rest of the coor-
dinates form the vector of linear coordinates qhL . This partition into nonlinear and linear DoFs has been used before, e.g. in
[15]. It leverages the fact that the linear DoFs do not directly experience the nonlinear contact forces. Therefore, the linear
DoFs are completely determined by the nonlinear DoFs by the relationship:
 1  
h
qhL ¼ DhL;L DhL;N qhN þ f E;L 8h 2 ½0; H: ð16Þ

Using Eqs. (15) and (16) one may describe the dynamics of the system only in terms of the nonlinear DoFs as:
 1
h h h
Hh qhN ¼ DhN;L DhL;L f E;L þ f E;N þ f C;N 8h 2 ½0; H; ð17Þ

with
 1
Hh :¼ DhN;N  DhN;L DhL;L DhL;N ; 8h 2 ½0; H: ð18Þ

This technique is particularly beneficial as the number of nonlinear equations in (17) is proportional to the number of con-
tact nodes (or contact pairs, depending on the application).

3. Model order reduction

In state-of-the-art methods described in the Introduction, the size of nonlinear systems with friction contacts is propor-
tional to the number of contact pairs on the contact surfaces. This is a challenge in cases where a refined mesh is necessary to
model accurately the friction contact behavior or in structures where the multiple contact surfaces are present. In this sec-
tion, a novel reduction technique is proposed to (further) reduce the size of Eq. (2) and obtain nonlinear ROMs, whose size
often does not increase significantly with the number of contact pairs used to model the intermittent contact.

3.1. Reduction strategy

HBM is used to generate the frequency response of the baseline model. The objective is to obtain a reduction basis
Uh ; 8h 2 ½0; H for the reduction of the baseline model (static and dynamic equations alike). The JP procedure is applied to
obtain this reduction basis. The details of the JP approach are explained in Section 3.2. Different types of ROMs are obtained
[13] depending on whether all the DoFs of the reduction basis are employed for reduction or only the nonlinear DoFs.

3.1.1. Full DoF reduction

It is possible to reduce all the DoFs of qh to a set of reduced coordinates ph provided that it is possible to build a reduction
basis Uh , for the reduction of the static terms) such that:

qh  Uh ph ; 8h 2 ½0; H: ð19Þ

Applying the reduction from Eq. (19) into Eq. (2), one obtains:

DhROM ph ¼ ghE þ ghC ; 8h 2 ½0; H; ð20Þ

where
 T
DhROM ¼ Uh Dh Uh ; 8h 2 ½0; H; ð21Þ
556 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

 T h
ghE ¼ Uh f E ; 8h 2 ½0; H; ð22Þ

 T h
ghC ¼ Uh f C ; 8h 2 ½0; H: ð23Þ
We refer to this reduction as the full DoF reduction, abbreviated as FR.

3.1.2. Nonlinear DoF reduction

The reduction basis may also be split into its linear and nonlinear DoFs:
" #
UhN
Uh ¼ ; 8h 2 ½0; H: ð24Þ
UhL
Hence, the reduction may be applied to the nonlinear DoFs and not to the linear DoFs:

qhN  UhN sh ; 8h 2 ½0; H: ð25Þ

where sh is the set of reduced coordinates corresponding to qhN . Eq. (25) may be substituted into Eq. (17) to yield the reduced
equations:
h h
HhROM sh ¼ lE þ lC ; 8h 2 ½0; H; ð26Þ
where
 T
HhROM ¼ UhN Hh UhN ; 8h 2 ½0; H; ð27Þ
  1
h  T h h
lE :¼ UhN DhN;L DhL;L f E;L þ f E;N ; 8h 2 ½0; H; ð28Þ

h  T h
lC ¼ UhN f C ; 8h 2 ½0; H: ð29Þ
We refer to this reduction as the nonlinear DoF reduction, abbreviated as NLR. When nonlinear DoFs are captured by the
reduction basis accurately, the NLR reduction is always accurate while the FR reduction might not be accurate. This is
because the FR reduction reduces the linear DoFs as well as the nonlinear DoFs. Hence, the approximation made in Eq.
(19) is a stronger condition than the approximation made in Eq. (25). However, NLR may also be comparatively slower com-
putationally than FR per iteration as it involves linear solutions of the matrix DhL;L for each investigated harmonic index
h 2 ½1; H. This drawback can be easily compensated if the matrix C is proportional to M and K (i.e. modal damping). In this
case, as pointed out in [16], a spectral expansion of the receptance matrix can be operated and the inversion of matrix DhL;L is
replaced by the trivial inversion of a diagonal matrix together with simple pre and post matrix multiplications.
Another advantageous point of the NLR is its convergence rate, usually superior to the one displayed by the FR technique.

3.2. Calculation of JP modes

The key idea of the JP reduction technique is that nonlinear friction dynamics can be captured by a set of intermediate
linear systems generated by imposing specifically chosen boundary conditions at the contact interfaces. Let us suppose to
know the boundary conditions, i.e. matrix K b associated with the baseline solution. It is then possible to investigate the
autonomous solution associated with Eq. (12):
 
b q ¼ 0:
x2 MJ þ ixCJ þ KJ þ K ð30Þ

A static condensation to express the static DoFs as a function of the dynamic components can be expressed as:
!
 1 X
H
0 0;0 b
q ¼  KJ þ K 0;0 b
K q ;
0;h h
ð31Þ
h¼1

 
x2 MJC þ ixCJC þ KJC qð1HÞ ¼ 0; ð32Þ
with
 T  T
qð1HÞ ¼ q1 ; . . . ; qH ¼ q1;R ; q1;I ; . . . ; qH;R ; qH;I ;
ð1H;1HÞ ð1H;1HÞ
MJC ¼ MJ ;
CJC ¼ CJ ; ð33Þ
 1
ði;jÞ
KJC b i;j þ K  K
¼K i;j b i;0 K þ K
0;0 b 0;0 b 0;j ;
K 1 6 i 6 H; 1 6 j 6 H:
J J
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 557

b is being used, complex value vectors qh ¼ qh;R þ iqh;I have been split into
Note that, since the matrix of partial derivatives K
h
their real and imaginary components (see q definition in Eq. (5)).
The static condensation is a reversible, error-free procedure, necessary to avoid a badly conditioned eigenvalue problem
(see the associated equation, Eq. (30), with M0;0
J ¼ 0 from Eq. (10)). Once the static condensation has been performed, it is
possible to solve the eigenvalue problem associated to Eq. (32) as:
KJC VJC ¼ KJC MJC VJC ; ð34Þ

where KJC ¼ diagð½k1 ; . . . ; k2NDoF H Þ is the diagonal matrix of eigenvalues and VJC ¼ ½VJC1 ; . . . ; VJCð2N  is the matrix of the
DoF HÞ

eigenvectors. The eigenvectors VJC are vectors of size 2  NDoF  H, and there are ð2  NDoF  HÞ eigenvalues since the real and
imaginary component of forces and displacements are treated separately. If the system is fully stuck (i.e. linear case) only
ðNDoF  HÞ eigenvectors are linearly independent. Thus the eigenvectors and corresponding eigenvalues are equal in pairs
(VJCk ¼ VJCkþ1 with k an odd number) since the real and imaginary part of forces and displacements undergo the same trans-
formation. If the system enters slip then VJCk – VJCkþ1 .
The two eigenvectors VJCk and VJCkþ1 whose corresponding eigenvalues fall within the investigated frequency range
kk ; kkþ1 2 KJC , can be used to represent the solution q as:

qh  a1 Uh1 þ a2 Uh2 8h 2 ½1; H ð35Þ

with a1 and a2 generic scalar quantities and

Uhx ¼ Uh;R
x þ iUx ;
h:I
x ¼ 1; 2; ð36Þ
" #
Uh;R
x
Uhx ¼ ; x ¼ 1; 2; ð37Þ
Uh:I
x

and
2 3
U11 U12
6 . .. 7
6 . 7
6 . . 7
6 7  
U1H ¼6
6 U1
h
Uh2 77 ¼ VJCk ; VJCkþ1 : ð38Þ
6 . .. 7
6 . 7
4 . . 5
UH1 UH2

For the static component (h ¼ 0):

X
H
 
q0  bz;1 U02z1 þ bz;2 U02z þ U02Hþ1 ð39Þ
z¼1

with bz;1 and bz;2 generic scalar quantities and

  h i
U0 ¼ U01 ; . . . ; U02Hþ1 ¼ U0dyn ; U0f E

where
 
U0dyn ¼ U01 ; U02 ; . . . ; U02H1 ; U02H ¼
 1 h i
¼ K0;0 þK b 0;0 K b 0;1 U1 ; . . . ; K
b 0;1 U1 ; K b 0;H UH ; K
b 0;H UH
J 1 2 1 2 ð40Þ
 1
U0f 0 ¼ U02Hþ1 ¼ K0;0 þK b 0;0 0
fE
J
E

Specifically U0dyn is obtained through a backward static condensation (see Eq. (31)) and represents the influence of the
dynamics on the static term. Please note that K b 0;h is a rectangular matrix (NDoF  2NDoF ) compatible with the length of U0
(NDoF ) and U (2NDoF )). On the other hand Uf 0 takes into account another important contribution, i.e. the presence of static
h 0
E

loads.
It was verified, under different excitation levels, and with an increasing number of retained harmonics H, that the reduc-
tion basis Uh 8h 2 ½0; H coming from the boundary conditions of the baseline solution is capable of representing the baseline
solution with an error lower than 0:1% for structures such as the one investigated in Section 4. Moreover the method intrin-
sically captures the friction-induced cross-harmonic coupling and the coupling between static and dynamic components.
However the boundary conditions corresponding to the FO solution are usually not available. In fact this reduction technique
was designed for those systems whose baseline solution is not attainable in a feasible amount of time. The following section
558 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

presents a method to predict boundary conditions representative of the contact behavior, not involving the solution of the FO
system.

3.3. Prediction of the boundary conditions

These boundary conditions are generated by exciting the system along a predefined shape, similarly to [13]. Unlike linear
systems, frequency and amplitude of the peak of nonlinear FRFs depend on the excitation level, a. Therefore, friction-damped
systems are usually investigated by calculating optimization curves. These curves plot the frequency and amplitude of the
nonlinear FRF peak as a function of a. The analysis of new systems typically starts from high values of the dimensionless
parameter a ¼ jf E j=jf E j, i.e. close to the linear full-stick case. For a sufficiently high value of this parameter, termed here
0 1

aST , the nonlinear system is fully stuck at all times and its response is the same as the corresponding stuck linear system.
If the analysis starts for a value of a sufficiently close to aST it is reasonable to use the stuck mode shape to produce the
boundary conditions.
The tangential and normal forces at the contact interfaces are calculated by considering an artificial modal displacement
^  WST where WST is the stuck mode in the frequency range of interest and p
p ^ is a dimensionless parameter ranging between
^min and p
p ^max [13]. The static components of this artificial displacement are obtained through a preliminary static analysis.
A practical value of p^min is the highest value of p ^ which still guarantees a completely stuck system. The parameter p ^max
should instead reflect the expected response of the system. The user can estimate it from a variety of sources such as con-
servative linear estimations which depend on material damping (amongst other factors such as contact stiffness) and forcing
 1
b 1;1
^max ¼ j D1 þ K
level, i.e. p b 1;1 is the matrix of partial derivatives of contact forces with respect to dis-
f 1E j=jWST aj where K
ST ST

placements of the fully stuck system.

The range of values between p ^min and p
^max is sampled (e.g., uniformly) n times to create the vector of chosen p
^ values.
^ b
Then, boundary conditions are generated for each chosen value of p, i.e. the matrix K is built and used to generate the JP
modes, as described in Eqs. (36) and (40) to obtain:
h i
Uh ¼ Uhp^min ; . . . ; Uhp^max 8h 2 ½0; H: ð41Þ

A scheme of the algorithm is given in Fig. 2. In it, the vectors Uh 8h 2 ½1; H; U0dyn and U0f0 developed in Eqs. (36) and (40) are
E

represented with an additional subscript (e.g. U0dyn becomes U0dynp^ ) to identify the value of p ^ used to obtain them. As
observed above, one single mode is used to build the basis, therefore superharmonic resonances could not be detected by
the JP as it is. This choice, common to similar techniques (e.g. Amplitude Microslip Projection, [13]), comes from the fact that
the industrial applications the technique is intended for, are very rarely affected by superharmonic resonances.

3.4. Selection of the JP modes

It may be observed that the boundary conditions arising out of two different values of p ^ might be the same. Care must be
taken to eliminate such duplicate conditions and retain only those values of p ^ which generate unique boundary conditions;
otherwise the JP basis may be rank deficient. The boundary conditions can be generated and checked for uniqueness very
fast. Hence, it is computationally easy and inexpensive to initially oversample p ^ and then select only those values corre-
sponding to unique boundary conditions. If several chosen values of p ^ are too close to each other, the boundary conditions
thus generated might be unique but still lead to a basis Uh ; 8h 2 ½0; H which is nearly rank deficient. Thus, even when care is
taken to exclude non-unique boundary conditions, it is possible to obtain contact states which differ only at a few contact
nodes. Such a situation might lead to an ill conditioned basis, which may result in convergence problems. To avoid this unde-
sirable situation the basis is, in most cases, further conditioned by performing a singular value decomposition as follows:
T
Uh Sh ðVh Þ ¼ Uh ; 8h 2 ½0; H; ð42Þ

Uhcond ¼ Uhg ; 8h 2 ½0; H: ð43Þ

Instead of eliminating individual vectors in the basis Uh to improve conditioning, vectors are selected from the left singular
vector matrix Uh . These vectors span the same space as Uh , but are ordered according to the magnitude of their correspond-
ing singular values (in the diagonal matrix Sh ). The conditioning of the matrix refers to the ratio of the maximum singular
value in Uh and the minimum singular value retained. As this ratio (here termed g) decreases the number of conditioned
basis vectors retained also decreases. This ratio is chosen such that it lies below a user defined threshold to ensure good con-
ditioning. Moreover g, can be modulated to include a different number of modes inside each basis, e.g. if g increases more
modes are included in the conditioned basis thus resulting in RO responses with a higher accuracy. The conditioned JP reduc-
tion basis Uhcond may be substituted for Uh in the equations of FR and NLR reductions.
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 559

If the value of the a parameter decreases to a ¼ az , more nodes enter slip and the stuck mode shape may cease to be an
adequate generating shape. In this case the first harmonic component of the reduced solution previously obtained at a higher
a ¼ az1 value can be used as a generating shape:
p1 ðaz1 Þ ¼ a1 U11 þ a2 U12 : ð44Þ
^ are p
In this case a practical values of the amplification factor p ^min ¼ az =az1 . A practical example of this strategy
^min ¼ 1 and p
is provided in Section 4.

3.5. Error predictor metric

This section proposes a method to evaluate the RO solution accuracy independently of the comparison with the FO solu-
tion. In detail, considering a FR and given a reduced system it is possible to compute the vectors:
h h
f C
¼ Dh Uh ph  f E ; 8h 2 ½0; H: ð45Þ

If the reduction leads to no loss of information, then

h h h
f C
¼ f C ¼ f CROM ; ð46Þ
h
where f CROM is the vector of contact forces produced by the reduced displacement Ui pi with i 2 ½0; H which satisfies the
HBM equations of motion:
T T
 
h h
ðUh Þ Dh Uh ph  ðUh Þ f E þ f CROM ; h 2 ½0; H: ð47Þ

Then, the following error estimate can be computed:

h h h
ehfc ¼ kf C
 f CROM k=kf CROM k: ð48Þ

h h
It should be noted that ehfc does not need the FO solution, as it compares two vectors, f C
and f CROM which are obtained using
the reduction basis U , the RO solution ph and two different versions of the equilibrium equations (Eqs. (45) and (47)).
h

This technique is useful in warning the user whenever the reduction basis captures the actual contact conditions inade-
quately. This can happen in two main cases:

 the number of vectors retained in the conditioned basis g is too low;

 the generating shape is inadequate with respect to the excitation level.

In these cases the user should try, at first, to improve the accuracy by including more vectors inside the conditioned basis.
If this is not feasible or if the results do not converge to a solution with an acceptable error estimate, the user should re-
compute the boundary conditions starting from a different generating shape, typically using the RO solution with an a
parameter slightly lower then the one under investigation.
A similar procedure can be applied to a NLR system, considering only the nonlinear DoFs. The error estimate can be com-


puted for any harmonic index, in this study e1fc was used. A practical limit to discriminate between a good and an inadequate
RO solution is 0:05. Further details are provided in Section 4.3.

4. Results

4.1. Baseline model

The JP described in Section 3.2 is applied to study the vibration of a cantilever shrouded blade connected to ground. The
material is steel, with Young’s modulus E ¼ 2:0105 MPa, Poisson coefficient m ¼ 0:3 and density q ¼ 7:800 kg/m3. The FE
model of the structure in shown in Fig. 1, where the contact patch is highlighted. The blade was modeled with linear solid
elements (a total of 21,650 DoFs). The determination of the contact parameters necessary to characterize the friction contact
elements used in this paper is still an open research issue in the friction damping scientific community. Experiments [17] and
theoretical/computational models [18] have been proposed. Here, we are using the rectangular flat punch model of [18]. The
resulting values are: coefficient of friction of l ¼ 0:4, nodal tangential contact stiffness values of
ku1 ¼ 2:0126  104 N=mm; ku2 ¼ 2:0059  104 N=mm and a nodal normal contact stiffness of kv ¼ 2:34152  104 N=mm.
In the following analyses, the JP basis is computed starting from mass and stiffness matrices previously reduced using CB-
CMS, where 50 slave modes (i.e., slave DoFs) are retained. The 50 modes are more than enough to accurately compute the
response of the structure in the frequency range of interest (e.g., natural frequencies, deflection, and mode shapes). Further-
more, the CB-CMS model is used not only to generate the JPMs but also as a reference for the JP results. Thus, the CB-CMS
model and the JP model are consistent with each other.
560 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

Fig. 1. Left: shrouded blade FE model and global coordinate system; right: close up on the right and left contact patches and local coordinate systems.

Fig. 2. Algorithm used to predict boundary conditions and to build the multi-harmonic reduction basis U.

4.2. Investigated cases

The physical DoFs retained after the CB-CMS reduction include 18 contact nodes and 9 nodes along the blade profile used
partly to apply the forcing vectors and partly as output nodes.
0
A vector of static forces f E is applied along the blade profile to simulate the pre-twist load and ensure the contact between
1
shrouds and ground. A first harmonic forcing vector f E is applied on the blade profile to provide a dynamic excitation.
In this work, the first bending mode of the blade is investigated. The frequency range of interest is ½2140  2190 Hz. The
forced responses are analyzed for different excitation levels, i.e. decreasing values of a, to study the response of the system at
different levels of slip. Values of a starting from aST decreasing down to 3  103  aST are investigated. The lower limit of the
interval guarantees that all points of both interfaces enter slip during a cycle of vibration (i.e., gross slip).
Both the baseline and the reduced models are solved with HBM and a coupled static solution. As previously mentioned, a
trust region algorithm is used for solving the algebraic balance equations. The maximum harmonic index is usually set to
H ¼ 1, however additional calculations with H ¼ 3 are performed as well. The output quantities of interest are the ampli-
tudes of the forced response (both its static and harmonic component, here termed X0 and Xh , with 1 6 h 6 3) at the output
node along the x direction (see Fig. 1).

4.3. Results discussion

This section compares the RO forced response with the response of the baseline model. The nonlinear forced response for
frequencies near the 1st bending mode is shown in Fig. 3 for decreasing values of a and for H=1. JPM indicates results
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 561

Fig. 3. First harmonic of the forced response for a: a ¼ aST =10; b: a ¼ aST =3000; static term of the forced response for c: a ¼ aST =10; d: a ¼ aST =3000.

obtained with a JP-reduced model. The JP modes in U0 and U1 are used to reduce the static and the first harmonic component
respectively. In the cases shown in Fig. 3 the first stuck mode was used as a generating shape for the boundary conditions,
and the resulting reduction basis were conditioned using SVD. Parameters n0 and n1 refer to the size of U0 and U1 respec-
tively. One may envision several techniques for the selection of the JP modes to be included in the basis. In this case, as
described in Section 3.4, n0 and n1 are not user-defined, rather the user selects a value of g, the ratio between the highest
and lowest singular value of the vectors retained in each basis. In Fig. 3, g ranges from 102 to 104 , ensuring an increasing
accuracy.
To challenge the JP reduction technique, the same mode is investigated for two very different limit cases. For
a=aST ¼ 1=10 (Fig. 3a and c) only one contact node is slipping, while for a=aST ¼ 1=3000 (Fig. 3b and d) all points belonging
to the contact areas are slipping, i.e. gross-slip. This can be observed in Fig. 4, which shows the energy dissipated by friction
at each contact point for different excitation levels. The dissipated energy was computed according to the formulation in
[19]. It can be seen how a relatively low number of JP modes capture both displacements (see Fig. 3) and dissipated energy
(see Fig. 4) with excellent accuracy.
Fig. 5 plots the relative error between FO and RO displacement at the response node against the number of retained
modes in the reduction basis U0 and U1 . This comparison is performed at different excitation levels. The number of condi-
tioned modes n0 and n1 is generated using as a criterion the level of g. That is why n0 and n1 are not uniformly spaced in the
diagram. It is shown how the g criterion leads to higher values of retained modes n0 and n1 for increasing excitation levels.
This is explained by the fact that the number of spatial correlations (proportional to the number of significantly different
boundary conditions at the contact) increases, over the same frequency range, for higher excitation levels. However, the
number of spatial correlations is significantly lower than the size of the baseline system, therefore the JP still offers signif-
icant time savings.
562 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

Fig. 4. a: Contact points position and labels on the contact surfaces; b-c-d: energy dissipated per contact point for decreasing values of a.

Fig. 5. a: Static term relative error at resonance as a function of the size of the reduction basis U0 ; b: first harmonic term relative error at resonance as a
function of the size of the reduction basis U1 .
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 563

Fig. 6. a: Static term; b: first harmonic term; c: second harmonic term and, d: third harmonic term of the forced response for a ¼ aST =10.

4.3.1. JP for multi-harmonic balance calculations

This section investigates the performance of the JP when multi-HBM calculations are performed. The baseline model is
solved using a multi-HBM technique (H ¼ 3) with coupled static solution. An example of results is shown in Fig. 6 (excitation
level a ¼ aST =10). The JP reduction basis is now composed of Uh ; 8h 2 ½0  3. It should be noted that the additional reduc-
tion bases U2 and U3 are obtained by simply taking the proper portion of the eigenvectors inside VJC , as described in Eqs.
(36)–(38). The procedure is a natural extension of the mono-harmonic case, and involves a negligible additional computa-
tional effort.
Fig. 6 shows that the technique enables one to capture the cross-harmonic coupling introduced by slipping of the contact
surface. All components, from the static to the higher order harmonics, are correctly captured at a fraction of the computa-
tional cost used to solve the full order system.

4.3.2. Error predictor validation

The error-predictor presented in Section 3.5 is applied and validated using on all investigated cases. Fig. 7 a-b shows the


‘‘predicted” error ehfc defined in Eq. (48) vs. the actual one, defined as
h h h
ehfc ¼ kf C  f CROM k=kf CROM k; 8h 2 ½0; 1; ð49Þ
in log-log scale and for different excitation levels. It can be observed that all points fall below the 45° slope line. Therefore,
the a posteriori error estimate is conservative, i.e. the predicted error is always higher than the actual one. Fig. 7c plots the


comparison between the dissipated energy of the baseline (FO) model and that of a ROM with e1fc > 100% (see highlighted
dot in Fig. 7b). Two caveats can be drawn from this figure:

 the error predictor successfully predicts that the RO solution does not effectively capture the actual boundary conditions;
 the actual error e1f c and the error on the displacements (not shown here for the sake of brevity) are one order of magnitude


lower than e1fc . This may be due to the fact that, despite the differing BC at the contact, the RO and FO dissipated energies
have very similar values.
564 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

Fig. 7. Actual vs. predicted error for a: static term; b: first harmonic term; c: energy dissipated per contact point for the case highlighted in Fig. 7b.

Fig. 8. a: First harmonic of the forced response for a ¼ aST =1000 using JP modes obtained using two different generating shapes; b: force based error
estimate for the two cases shown in Fig. 8a.

Finally, Fig. 8a shows that for low values of a ¼ aST =1000 (approaching gross-slip), the stuck mode is not an adequate
shape for the generation of the boundary conditions. If the stuck mode shape is used, the resulting ROM displays a very
low accuracy. The a posteriori error estimate, shown in Fig. 8b warns the user. It is then possible to re-build the basis starting
from a different generating shape (in this case the reduced response previously obtained at a ¼ aST =500). In this case, the
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 565

Fig. 9. a: Size of U1 necessary to obtain an accuracy 6 1% vs. number of nonlinear DoFs (size of the full system) for different excitation levels a; b: different
number of retained contact nodes leading to, respectively, 24, 54 and 150 nonlinear contact DoFs for each harmonic.

right boundary conditions are predicted, and a high accuracy is obtained using a basis with the same size as the previous one.
This result is confirmed both by the comparison with the FO solution (Fig. 8a) and by the error predictor metric (Fig. 8b).

4.3.3. Influence of the number of nonlinear DoFs

This section investigates the performance of the JP reduction technique on systems with an increasing number of nonlin-
ear (contact) DoFs. The baseline system described in Section 4.1 has been reduced using CB-CMS retaining an increasing
number of master (contact) nodes. As shown in Fig. 9b, three cases have been investigated, with respectively 8, 18 and
50 retained contact nodes (case B is the one extensively discussed in Section 4.3). Fig. 9a compares the size of these three
FO systems (solved with the coordinate reduction as in Section 2.3) against that of the corresponding JP-reduced systems.
The comparison is performed for different excitation levels a. In Fig. 9a, the first harmonic of the solution is used as a bench-
mark for comparison: the size of the FO system, i.e. the number of contact DoFs, (x axis) is plotted against the size of the
reduced system n1 (y axis). A similar comparison could be performed on the other harmonic terms. However,
n1 ¼ nh 8h > 1 and n0 6 n1 . The size of the reduced systems is selected to be the smallest that guarantees a relative error
at resonance lower than 1%. Each point on the diagram in Fig. 9a can be used to asses how convenient the reduction is: the
smaller the y=x ratio is, the more convenient the reduction. All points fall well below the 45° slope line. Therefore, the reduc-
tion is reasonably convenient in all investigated cases (y=x 6 0:67). However, it can be observed that the lower the number of
contact DoFs is, the closer are the points to the 45° slope line, and the less convenient the JP reduction is. In fact the size of
the reduced system (n1 ) is proportional to the number of spatial correlations encountered throughout the FRF. The number of
spatial correlations is only marginally influenced by the number of contact DOFs, while the excitation level a has a stronger
influence. It can therefore be concluded that:

 the JP reduction technique is increasingly convenient for higher the number of contact DoFs;
 the size of the JP reduction basis tends to increase with increasing excitation levels.

4.3.4. Computational cost analysis

The purpose of this section is to show that the JP does not only lead to a smaller system, but also to a significant reduction
in computational time.
This analysis is regarded as crucial since the JP implementation, similarly to [12,13], does introduce a few operations
which increase the overall computational cost. In detail:

 building (and conditioning) the reduction basis introduces offline costs;

 switching physical and solution space to evaluate the nonlinear forces increases the computational cost per iteration.

It was extensively shown for the Amplitude Microslip Projection technique in [13] and here remarked for the JP, that
these two aspects are greatly surpassed by the significant improvement in the iterative scheme (e.g. Newton-Raphson) per-
formance when handling the minimization of a smaller (i.e. reduced) system of equations.
To prove this point a comparison between the computational time of the FO vs. RO models has been performed. A typical
30 Hz (centered around the resonance), 60 frequency steps, a ¼ aST =3 case has been chosen as a benchmark. The RO models
used in the analysis guarantee a relative error at resonance lower than 1%. Both solutions (FO and RO) are obtained on the
566 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

Fig. 10. a: Computational time necessary to obtain the FO vs. RO solution for systems with an increasing number of nonlinear DOFs; b: computational time
necessary to obtain the FO vs. RO solution for systems with an increasing number of harmonics retained in the multi-HBM computation; c: ratio between
RO vs. FO overall computational time (18 contact points, H = 1) as a function of the number of analysis performed with the same ROM (offline costs
performed only once).

same machine with 4 GB RAM and an Intel(R) Core(TM) i5-5200U CPU @ 2.20 GHz, 2201 MHz, 2 cores, 4 logical processors.
Each run has been repeated three times and found to be consistent (standard deviation < 1.5 s).
The results are shown in Fig. 10. It can be seen how, despite the non-negligible offline costs, the ROM is still faster than
the FO counterpart in all investigated cases. This can be expressed as the ratio between the RO total computational time and
the FO one, here in the [13–23]% range.
Furthermore, it is observed that time savings increase the larger the size of the FOM. This can be observed by increasing
the number of nonlinear DoFs (Fig. 10a) or by including higher harmonic terms in the multi-HBM computation (Fig. 10b).
This ‘‘increased” advantage is due to the Newton Raphson Scheme, whose efficiency in handling residual minimization dete-
riorates with larger systems (i.e. it becomes a ‘‘bottle-neck”). In the case observed in Fig. 10b at H = 5, the FOM needs an aver-
age of 107.2 iterations per frequency step, resulting in  543 s overall computational time. The same calculation performed
with the ROM is approximately one order of magnitude shorter despite the offline costs and the additional re-projections
performed: the RO model ensures a smaller system of equations to be minimized, thus allowing the iterative scheme to reach
convergence with an average of only 9.4 iterations per frequency step.
This analysis has been performed using a reasonable, typical, benchmark calculation. There are cases where applying the
JP would not be advisable, e.g. medium-small sized systems, to the purpose of computing a one-time-only 5 frequency steps
solution (as would be the case for many other reductions techniques). Conversely, there are cases where repeated calcula-
tions using the same basis make JP increasingly convenient. Fig. 10c shows how the impact of the offline costs decreases with
the number of analysis performed with a given reduction basis. The same basis can be used to safely perform more than one
particular calculation (subject to cross confirmation by the a-posteriori error estimate). Typical examples are parametric
studies in the region of the original set of values, or analysis of the response of the system to limited variations of the forcing
function/preload. The ratio between overall computational cost of the ROM and that of the FOM further decreases the higher
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 567

the number of analysis performed with the same basis (i.e. it goes from 37% for one benchmark calculation to 22% for 10
benchmark calculations).

5. Conclusions

In this paper a Jacobian-based reduction technique for the dynamic analysis of structures with friction contacts is pro-
posed and discussed. The Jacobian Projection (JP) technique is based on simple linear analyses and on the computation of
the Jacobian matrix, a technique already implemented in all state-of-the-art codes for the forced-response calculation of
friction-damped systems. The JP technique is able to capture the response characteristics of the baseline model accurately.
Moreover this technique features several advantages not always present in similar reduction techniques. Specifically,

 cross-harmonic coupling is intrinsically taken into account by the reduction technique;

 all harmonic terms of the solution (static component as well) are represented by the JP;
 the reduction basis not only spans the space of the displacements of the system, but also spans the space of nonlinear
forces as demonstrated by the comparison between FO and RO dissipated energies;
 the JP becomes increasingly convenient the higher the number of nonlinear DoFs: the system is reduced down to 8% of its
original size and solved in down to 13.5% of the computational time necessary for the baseline counterpart;
 the technique is completed by an error-predictor metric to evaluate the quality of a RO solution without the need for
cumbersome convergence studies.

All these features make the JP technique an attractive candidate for the prediction of model dynamics where the full order
model is too cumbersome to solve for validation.
The method was demonstrated for a simple test case, but its application can be easily extended to other classes of prob-
lems with localized friction contacts.

Appendix A. Additional details on the solution of nonlinear dynamic equations of structures with friction contacts

A.1. Preliminary CB-CMS reduction

In this study, the Craig-Bampton CMS method (CB-CMS) [9,10] is chosen to perform a preliminary reduction to facilitate
analysis and model development. The CB-CMS method involves retaining a certain number of DoFs corresponding to specif-
ically chosen master nodes of the full model, while reducing the rest of the DoFs into a set of slave modes. The master nodes
selected for this analysis include the ones on the contact surfaces, and a set of nodes where the structure is forced externally
and a set of nodes where the response is calculated. After the CB-CMS analysis, the mass and stiffness matrices M and K of
the free-system (see Eq. (1)) are extracted. For this study, it is assumed that C is proportional to K .
It is worthwhile noting that the CB-CMS is not necessary to implement the JP method described in this manuscript. The
reduction algorithm described in Section 3 can be applied directly to the full-order FE model. The decision on whether to
implement or by-pass CMS is related to the time necessary to perform modal analysis of the full FE models. In some cases
it is convenient to perform CMS and then apply the JP technique on the CMS-reduced model, in other cases it is faster to
perform JP directly on the FO model.

A.2. Modeling the contact

Contact forces f C in Eq. (1) are computed as a function of input displacements at the contact by means of a contact model.
The first paragraph describes the contact model used within this study. The second paragraph describes the AFT method,
necessary to implement the contact model while solving the equilibrium equations in the frequency domain.

A.2.1. Contact model

The nonlinear forces f C depend on the relative displacements at the contact. A local reference system can be defined at
each contact. Axes u1 and u2 define the local contact plane, while axis v defines the normal direction (as shown in Fig. 1).
Thus, for each contact node, a rotation matrix R can be used to project the absolute displacements of the nth contact node
onto the local reference system.
The contact conditions are enforced at the contacting surfaces by means of contact elements presented in [20] and later
used in [6,13,21]. Each contact element is characterized by a normal stiffness kv and by two tangential stiffness ku1 and ku2
along directions 1 and 2, while the coefficient of friction is l. This contact element can connect, depending on the model and
on the application,

(a) a node belonging to the contact of one body to the ground;

(b) two overlapping nodes, each belonging to one of two bodies mating surfaces (i.e., a node pair).
568 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569

Fig. A.11. Flow chart of the Alternating Frequency Time (AFT) method.

In the simple example analyzed in this paper (i.e., a cantilever blade) the two contact patches are connected to ground
(case (a)). The methodology described in Section 3 can be adapted easily to case (b). In this second case it is convenient
to express the displacements of the contact DoFs in relative coordinates as was done in [12,13]. While the present study
was developed using this specific contact model, the JP technique supports the application of other models as well, e.g. [22].

A.2.2. Alternating frequency time method

The contact model is easy to apply in the time domain. Therefore, to compute the Fourier coefficients of the nonlinear
h
forces f C from the Fourier coefficients of displacements qh , an alternate frequency time (AFT) domain method [2,23] is used.
As shown in Fig. A.11, an inverse fast Fourier transform (IFFT) is used to compute q in the time domain. Next, the periodic
contact forces f C are computed in the time domain. Finally, a fast Fourier transform (FFT) is used to compute the Fourier coef-
ficients of the nonlinear forces. The core step of the AFT method is the second step, where the contact forces are calculated
from the displacements of the nonlinear DoFs.

References

[1] 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, http://dx.doi.org/10.1115/1.4000266.
[2] T. Cameron, J. Griffin, An alternating frequency/time domain method for calculating the steady-state response of nonlinear dynamic system, J. Appl.
Mech. 56 (1) (1989) 149–154, http://dx.doi.org/10.1115/1.3176036.
[3] A. Cardona, T. Coune, A. Lerusse, M. Geradin, A multiharmonic method for nonlinear vibration analysis, Int. J. Numer. Methods Eng. 37 (9) (1994) 1593–
1608.
[4] K. Sanliturk, D. Ewins, Modelling two-dimensional friction contact and its application using harmonic balance method, J. Sound Vib. 193 (2) (1996)
511–523.
[5] D. Laxalde, F. Thouverez, J. Sinou, J.-P. Lombard, Qualitative analysis of forced response of blisks with friction ring dampers, Eur. J. Mech. – A/Solids 26
(4) (2007) 676–687.
[6] C.M. Firrone, S. Zucca, Numerical Analysis: Theory And Application, Intech, 2011, Chapter: Modelling friction contacts in structural dynamics and its
application to turbine bladed disks numerical analysis – theory and application, INTECH (2011) Rijeka, pp. 301–334.
[7] G. Kerschen, J. claude Golinval, A.F. Vakakis, L.A. Bergman, The method of proper orthogonal decomposition for dynamical characterization and order
reduction of mechanical systems: an overview, Nonlinear Dynam. 41 (1–3) (2005) 147–169, http://dx.doi.org/10.1007/s11071-005-2803-2.
[8] M. Krack, L.P. von Scheidt, J. Wallaschek, C. Siewert, A. Hartung, Reduced order modeling based on complex nonlinear modal analysis and its
application to bladed disks with shroud contact, J. Eng. Gas Turbines Power 135 (10) (2013) 102502, http://dx.doi.org/10.1115/1.4025002.
[9] R.R. Craig, M.C.C. Bampton, Coupling of substructures for dynamic analyses, AIAA J. 6 (7) (1968) 1313–1319.
[10] W.C. Hurty, Dynamic analysis of structural systems using component modes, AIAA J. 3 (4) (1965) 678–685.
[11] D.J. Segalman, Model reduction of systems with localized nonlinearities, J. Comput. Nonlinear Dynam. 2 (3) (2007) 249, http://dx.doi.org/10.1115/
1.2727495.
[12] S. Zucca, B.I. Epureanu, Bi-linear reduced-order models of structures with friction intermittent contacts, Nonlinear Dynam. 77 (3) (2014) 1055–1067,
http://dx.doi.org/10.1007/s11071-014-1363-8.
[13] M. Mitra, S. Zucca, B.I. Epureanu, Adaptive microslip projection for reduction of frictional and contact nonlinearities in shrouded blisks, J. Comput.
Nonlinear Dynam. 11 (4) (2016) 041016, http://dx.doi.org/10.1115/1.4033003.
[14] C. Gastaldi, S. Zucca, B.I. Epureanu, Jacobian projection for the reduction of friction induced nonlinearities, in: Proceedings of ISMA 2016, 2016.
[15] O. Marinescu, B.I. Epureanu, M. Banu, Reduced order models of mistuned cracked bladed disks, J. Vib. Acoust. 133 (5) (2011) 051014, http://dx.doi.org/
10.1115/1.4003940.
[16] O. Poudou, C. Pierre, Hybrid frequency-time domain methods for the analysis of complex structural systems with dry friction damping, in: Proceedings
of the 44th AIAA/ASME/ASCE/AHS Structures, Structural Dynamics, and Materials Conference, Norfolk, Virginia, American Institute of Aeronautics and
Astronautics, 2003.
[17] C. Schwingshackl, E. Petrov, D. Ewins, Measured and estimated friction interface parameters in a nonlinear dynamic analysis, Mech. Syst. Signal
Process. 28 (2012) 574–584, http://dx.doi.org/10.1016/j.ymssp.2011.10.005.
[18] M. Allara, A model for the characterization of friction contacts in turbine blades, J. Sound Vib. 320 (3) (2009) 527–544, http://dx.doi.org/10.1016/j.
jsv.2008.08.016.
[19] S. Baek, B. Epureanu, Reduced order modeling of bladed disks with friction ring dampers, in: ASME Turbo Expo, Structures and Dynamics, vol. 7A, ASME
International, 2016, http://dx.doi.org/10.1115/GT2016-57831.
[20] B. Yang, M. Chu, C. Menq, Stick–slip–separation analysis and non-linear stiffness and damping characterization of friction contacts having variable
normal load, J. Sound Vib. 210 (4) (1998) 461–481, http://dx.doi.org/10.1006/jsvi.1997.1305.
 C. Gastaldi et al. / Mechanical Systems and Signal Processing 100 (2018) 550–569 569

[21] E.P. Petrov, D.J. Ewins, Analytical formulation of friction interface elements for analysis of nonlinear multi-harmonic vibrations of bladed disks, J.
Turbomachinery 125 (2) (2003) 364, http://dx.doi.org/10.1115/1.1539868.
[22] D. Segalman, A four-parameter iwan model for lap-type joints, J. Appl. Mech. 72 (5) (2005) 752–760, http://dx.doi.org/10.1115/1.1989354.
[23] S. Yajie, H. Jie, S. Yingchun, Z. Zigen, Forced response analysis of shrouded blades by an alternating frequency/time domain method, in: ASME. Turbo
Expo: Power for Land, Sea, and Air. Marine; Microturbines and Small Turbomachinery; Oil and Gas Applications; Structures and Dynamics, Parts A and
B, vol. 5, ASME International, 2006, pp. 865–872, http://dx.doi.org/10.1115/gt2006-90595.