International Journal of Non-Linear Mechanics 92 (2017) 25–40

Contents lists available at ScienceDirect

International Journal of Non-Linear Mechanics

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

A method to solve the efficiency-accuracy trade-off of multi-harmonic MARK

balance calculation of structures with friction contacts
⁎
Chiara Gastaldi , Teresa M. Berruti
Department of Mechanical and Aerospace Engineering, Politecnico di Torino, Corso Duca degli Abruzzi, 24, 10129 Turin, Italy

A R T I C L E I N F O A B S T R A C T

Keywords: The steady-state nonlinear forced response of systems with frictional damping can be computed in the frequency
Harmonic balance method domain through the Harmonic Balance Method (HBM). A critical point is the selection of the number of
Harmonic selection harmonic terms used to represent the solution. In linear systems, this number is easily determined by the
Jacobian harmonic content of the forcing function (e.g. mono-harmonic). However, if nonlinearities are present, higher
Friction
order harmonics may need to be included to ensure a proper representation of friction forces and displacements,
Underplatform damper
with a detrimental effect on the computational time.
The paper presents a novel method to solve the efficiency-accuracy trade-off of harmonic selection for
nonlinear systems. This method warns the user whenever the number of retained harmonic terms is inadequate.
As a result, it enables the user to run the simulation with a low number of retained harmonics, e.g. designers are
typically interested in the first harmonic of the solution. The calculation is repeated with a larger harmonic
support only when strictly necessary to keep the error below a user-defined threshold.
The method is first compared to existing adaptive HBM techniques, highlighting its novel contributions. It is
then carefully validated against high-order multi-harmonic calculations and against direct time integration. Its
performance in terms of accuracy vs. computational time is highlighted. The method is then implemented in a
state-of-the-art numerical tool for the design of underplatform dampers for turbine blades. Finally, its outcome is
compared with experimental results.

1. Introduction frequency (stiffness contribution) and on the resonant amplitude

Forced vibrations of blades in power turbines still represent one of
the most common causes of failure. The different parts of a turbine are
connected together by interfaces. One of the challenges faced by turbine
designers is to correctly design these interfaces to provide adequate
friction damping to the system. The joints can be optimized in order to
exploit the energy dissipated at the contact to limit the structure
vibrations. The joints can be a part of the bladed disk (blade root joints
[1,2], snubber [3] and shrouds [4]) or can be purposely added to the
system (underplatform dampers [5], ring dampers [6]).
One of the most important parts of the design of systems character-
ized by frictional damping is the numerical prediction of the nonlinear
forced response. Unfortunately this task cannot be performed by
commercial finite element (FE) solvers in a reasonable amount of time.
Rather it requires the development of custom codes capable of
computing the amount of frictional damping for a given excitation
condition. Several authors [7–10] have given their contribution in this
field by taking into account the actual stick-slip, lift off states of the
contact in order to determine the effect of friction on the resonant
(damping contribution). Most of these papers deal with modeling
underplatform dampers (UPDs), metal devices compressed by the
centrifugal force under the blade platforms. The relative motion
between adjacent blades causes the damper to slip, and therefore to
dissipate energy. The common idea is to model both the blade and the
UPD using FE and to introduce contact elements [4] between them.
This paper presents the results of a numerical tool developed to be
used at design stage when the shape of the UPD is still not known. The
tool is required to be very flexible using the minimum number of
equations in order to be run several times at design stage. The method
adopted for the solution of the equations of motion is the harmonic
balance method (HBM) [11]. The number of equations to be solved is
given by the number of contact (potentially nonlinear) degrees of
freedom and the number of harmonic terms retained in the calculation.
A trade-off between accuracy of the solution and speed of computa-
tion exists: the number of retained harmonic terms should guarantee a
proper representation of friction forces and, at the same time, be kept to
a minimum to limit the number of algebraic variables. This trade-off is
typically realized through cumbersome convergence studies [4,9,12],

⁎
Corresponding author.

http://dx.doi.org/10.1016/j.ijnonlinmec.2017.03.010
Received 27 May 2016; Received in revised form 10 February 2017; Accepted 8 March 2017
Available online 10 March 2017
0020-7462/ © 2017 Elsevier Ltd. All rights reserved.
C. Gastaldi, T.M. Berruti International Journal of Non-Linear Mechanics 92 (2017) 25–40

∼
Nomenclature J, J Jacobian matrix, obtained using and respectively q and ∼q
˘
Δq correction applied to vector q by existing harmonic
Abbreviations selection techniques to take into account higher harmo-
nics contribution
AHBM adaptive harmonic balance method (q + Δq) approximation of vector ∼q obtained using J , implemented
AFT alternating frequency-time method by novel technique JAA
DOF(s) degree(s) of freedom T vector of tangential force at the contact
DTI direct time integration t vector of tangential displacement at the contact (perpen-
HBM harmonic balance method dicular to the damper axis)
JAA Jacobian alert algorithm w vector of tangential displacement at the contact (parallel
UPD underplatform damper to the damper axis)
n vector of normal displacement at the contact
Variables id vector identifying a given set of physical degrees of
freedom
Ndof degrees of freedom retained in HBM computation Γ matrix, close to the identity matrix, measure of the
EC evaluation criterion similarity between two matrices
En strain energy
Err error indicator/evaluation criterion Subscripts
t time
m generic frequency step D damper
nH number of harmonics used in the HBM (selected by the B blades
user) C contact
∼
nH number of harmonics used in the HBM (that ensure a CD contact, applied at the damper center of mass
highly accurate representation of friction forces and CB contact, applied at the blades contact nodes
displacements even in severely nonlinear cases) CL centrifugal load
ΔJ% relative norm indicator, to evaluate similarity between EXT external excitation
two matrices HE harmonic excitation
|i iteration i
Vectors and matrices L1 L2 R contact points ID, as shown in Fig. 1b
LN linear, nonlinear
M mass matrix OUT output
C damping matrix p predicted
K stiffness matrix
D dynamic stiffness matrix Superscripts
q generic displacement vector (either in time or in frequency
domain retaining nH harmonics) h, k harmonic indexes
∼
q displacement vector in the frequency domain obtained nH maximum harmonic index retained in nonlinear calcula-
retaining ∼
nH harmonics tion
F generic force vector ∼
nH /nH harmonics in the [0–∼
nH ] interval not retained in nonlinear
r vector of residuals HBM calculation, i.e. [0-nH]
K̂ matrix of partial derivatives of contact forces with respect
to displacements

where the same forced response is computed many times by progres-
sively increasing the number of retained harmonics. More recently,
several harmonic selection algorithms [13–17] have been proposed.
The main goal of this paper is to propose and validate a novel method to
solve this trade-off and compare it to existing methods.
This novel method warns the user whenever the number of retained
harmonic terms is inadequate – i.e. whenever the solution would
significantly change if the number of harmonic terms were to increase.
As a result, it enables the user to run the simulation with a low number
of retained harmonics and to add harmonics only when strictly
necessary. It should be noted that the tool here proposed evaluates
the adequacy of the harmonic terms retained in the non-linear calcula-
tion without the need for repeated iterative calculations.
This novel method is only one of the contributions of this paper
aimed at maximizing the computation speed of the design tool without
sacrificing its accuracy. The complete list of the main features of the
design tool is reported below.
1) Each damper is modelled as a rigid body due to its bulky structure. It
has inertia properties and six degrees of freedom at the center of
mass. This choice has two main benefits:
– the contact points can be conveniently chosen on the blade
platforms, according to an estimated damper geometry even if
at the design stage a finite element model of the damper is not yet
available;
– the displacements and the forces at the contact points are
transferred to the damper center of mass. In this way the degrees
of freedom associated to the damper are always limited to six
whatever the number of chosen contact points.
2) A Craig Bampton reduction [18] of blades mass and stiffness
matrices is performed. In it only the contact DOFs (plus output
and input nodes of interest) are retained, while all other liner DOFs
are approximated with a set of linear normal modes computed by
imposing proper boundary conditions.
3) The resulting blades DOFs (reduced linear plus contact/nonlinear)
are further condensed [10], i.e. linear DOFs are expressed as a
function of contact (potentially nonlinear) DOFs. As a result, non-
contact DOFs are excluded from the equations; full accuracy is
preserved and the size of the system to be solved through HBM is
limited by the harmonic order and the number of contact DOFs only.
4) The Jacobian matrix of the system, required by the iterative scheme,
is calculated analytically as described by [4], rather than numeri-
cally. This decreases enormously the computation time.

26
C. Gastaldi, T.M. Berruti International Journal of Non-Linear Mechanics 92 (2017) 25–40

5) Only the essential number of harmonics is retained in the HB of vector FC different from zero are those corresponding to contact
iterative forced response calculation. For this purpose, a new (nonlinear) DOFs.
efficient method, the Jacobian Alert Algorithm (JAA) has been
developed. At the beginning, the number of harmonic terms are ⎡ D h D h ⎤⎛ q h ⎞ ⎛ 0 ⎞ ⎛ F h ⎞
⎢ LL LN ⎥⎜ L ⎟
⎜ ⎟ = ⎜ h ⎟ + ⎜⎜ hEXT−L ⎟⎟ 0 ≤ h ≤ nH
limited to 0 and 1 with beneficial effects on the computational speed ⎢⎣ DNL DNN ⎥⎦⎝ q ⎠ ⎝ FCN ⎠ ⎝ F EXT− N ⎠
h h h
N (3)
compared to a multi-harmonic calculation (the size of system is
limited and convergence rates are higher). The JAA checks the Therefore, as previously observed by several authors, e.g. [10], the
obtained solution at each frequency and it warns the user when linear DOFs are completely determined by the non-linear ones. It is
more harmonics are needed to effectively represent friction forces therefore possible to solve only the nonlinear DOFs
and displacements. The iterative HBM computation is then repeated
retaining higher harmonic terms only for those frequency steps h
D ⋅qNh = FC + FEXT
h h
0 ≤ h ≤ nH
where it is strictly necessary.
h h h h −1 h
with D = D NN − D NL(D LL ) D LN
−1
The reader will notice how each contribution is aimed at over- hh
FC = FCN
h
FEXT = − D NhL(D LL
h h
) F EXT−L h
+ F EXT−N
coming difficulties introduced by the presence of nonlinearities. Solving
(4)
nonlinear systems through the HBM requires an iterative scheme (e.g.
Newton Raphson): the computational speed is strongly affected by the while the linear ones are computed ‘offline’ using:
size of the system (minimized through 1), 2), 3) and 5)) and further
−1
increased by 4). qLh = (D LL
h
) (D Lh N⋅qNh + F EXT−L
h
) 0 ≤ h ≤ nH (5)
This paper will at first present the governing equations of friction-
damped systems (Section 2), then review and discuss existing harmonic From now on, in the interest of brevity, the subscript N and the ‘bar’
selection techniques when solving non-linear dynamic systems with the sign over all quantities has been dropped, thus Eq. (4) becomes similar,
HBM (Section 3). in form, to Eq. (2).
Section 4 is devoted to the description of the JAA. In this context the Solution of (4) can be computed using the Harmonic Balance
proposed method advantages, novel contributions and differences with Method through an iterative scheme (e.g. Newton-Raphson, Dogleg
respect to existing techniques are clearly highlighted. Reflective, Levenberg-Marquardt). These iterative procedures require
Section 5 presents the numerical tool as a whole and describes the that Eq. (4) is reformulated as follows:
two test cases used to validate the JAA.
Section 6 is devoted to the numerical validation of the JAA. r h = D h⋅qh − FC h(q0−nH) − FEXT h 0 ≤ h ≤ nH (6)
Validation is achieved by comparing solutions obtained with an
increasing number of retained harmonic terms (in this paper up to 5). where r h is the vector of residuals of the systems of equations. The
It is observed that, when running low order harmonic calculation, the above-mentioned iterative schemes generate approximate solutions,
JAA warnings appear if and only if the corresponding high order converging towards the roots of the system. As an example, considering
solution is significantly different. Direct time integration is used as an the Newton-Raphson Method, the approximate solution at the i-th step
additional independent verification that the choice of five harmonic is estimated with the following iterative relationship:
terms is indeed adequate to the representation of friction forces. Finally
q i = q i−1 − J−1 i−1⋅r i−1 (7)
(Section 7), the complete numerical tool is validated against experi-
mental results obtained on a test rig with two real blades and one
where:
underplatform damper.

• q| = {q0, q1, ..., qnH}|i T is the response vector at the i-th iteration
•J
i
2. Governing equations i−1 is the Jacobian matrix at iteration i-1 defined as
δ r i−1 δ FC i−1
J i−1 = δ q i−1
= D i−1 + δ q i−1
Let us consider a multi-DOF system subjected to an external
dynamic forcing and to friction. Its equilibrium equations in the time • r| i = {r , r , ..., r nH}|i T is the residuals vector at the i-th iteration.
0 1

domain are given by:

M⋅q̈ (t) + C⋅q̇ (t) + K⋅q(t) = FEXT(t) − FC(q) (1)
The displacement vector q (and the corresponding matrices) may
contain only the master nodes retained in the application of a Craig-
Bampton reduction technique [18] on the FE model of system.
If the external force is periodic and the user is interested in the
steady state solution only, it is possible to switch to the frequency
domain using the well known Harmonic Balance Method [12]:
D h⋅qh = FC h(q0−nH) + FEXT h 0 ≤ h ≤ nH
with h 2
D = − (hω) M + i(hω)C + K
(2)
where FC is typically modelled by a set of node-to-node contact
elements [1]. Since contact models which compute the nonlinear
contact forces staring from displacements work only in the time
domain, the Alternate Frequency-Time (AFT) procedure [19] switches
between frequency and time domain accordingly using the IFFT and
FFT algorithms (see also Fig. 2).
Furthermore it should be noted that the size of the system in (2) can
be further reduced (as was done in this paper) by partitioning the
equations into linear and nonlinear (i.e. contact) DOFs. The only entries
It should be noted that Eqs. (6) and (7) allow only for purely real
vectors, therefore the size of the system increases from (Ndof)⋅(nH + 1)
to (Ndof)⋅(2⋅nH + 1) since complex force and displacement vectors have
to be separated into their real and imaginary components and the
dynamic stiffness matrices are re-organized accordingly.
One crucial problem when solving nonlinear dynamic systems
through HBM is selecting an adequate number of harmonics. If nH is
kept to a high value, then the size of the system to be solved may be
excessively big. However, if nH is too small neither contact forces nor
displacements will be correctly represented, thus producing misleading
results. For instance, if the system was linear (i.e. FC=0 or described
through simple linear springs) and FEXT was mono-harmonic, then the
solution q would be mono-harmonic as well. However, the presence of
friction makes this dynamic problem nonlinear. As a result, depending
on the level of non-linearity, both the vector of contact forces and the
vector of displacements may have higher harmonic content. The
purpose of Harmonic Selection Techniques is to ensure that the
harmonic support used to solve the equilibrium equations offers an
adequate representation of nonlinear forces and displacements. The
following section offers a review of existing harmonic selection
techniques.

27
C. Gastaldi, T.M. Berruti
3. Review on harmonic selection techniques
There already exists a variety of different approaches for the
implementation of an Adaptive Harmonic Balance Method (AHBM)
[13–17], i.e. HBM with a built-in harmonic selection algorithm. All
these methods can be classified based on the implementation strategy
and on the evaluation criterion used to select nH. It will be shown in
Section 4 that the proposed JAA method differs from existing techni-
ques in the implementation strategy, evaluation criterion and offers
valuable novel contributions with respect to error control.
All existing AHBM techniques aim at finding the “optimal” number
of harmonic terms for each frequency step. Each proposed harmonic
selection algorithm pursues this goal with a different implementation
strategy and evaluation criterion.
Laxalde proposed in [13] to use, as an evaluation criterion, the error
between the nonlinear forces in the time domain (coming from the AFT
procedure) and the corresponding reconstruction of nonlinear forces
from the selected Fourier coefficients:
⎛ nh
2π
ECnH = ∫0 FC(q, q˙ , t ) − ⎜⎜FC0 + ∑ (FCh,C cos(hωt) + FCh,S sin(hωt))⎟⎟dt
⎝ h =1
The maximum harmonic order at the frequency step m+1 is
adapted in order to maintain the error of the frequency step m below
a user-defined threshold.
Jaumouillé et al. in [14] use, as a criterion, the variation of strain
energy with the number of harmonics. At each frequency step the strain
energy is first calculated using a given number of harmonics nH and
then recomputed at nH +1. If the indicator
EnH − EnH +1
ECnH =
EnH
lies above the threshold then the computation is repeated at nH +2 and
so on, otherwise the computation shifts to the following frequency step.
In [15] Maple et al. use, as an evaluation criterion, a fraction of
spectral density of the last retained harmonic:
nH,C 2 n ,S 2
ECnH = (q ) + (q H )
0 2 H n h,C 2 h,S 2 change even if the harmonic order is updated to ∼
(q ) + ∑h=1 ((q ) + (q ) )
If the fraction of spectral density exceeds a given threshold then
additional harmonics are introduced and the solution is computed once
again with the new number of harmonics. The new maximum harmonic
order is retained for the following frequency step.
The main limitation of the strategies presented above is the
potentially high computational cost since, for each iteration of the
selection procedure (i.e. potentially multiple times per a given
frequency step), a complete nonlinear solution must be computed.
Furthermore, as correctly pointed out by Grolet et al. in [16] “if the
increment is not chosen well, it can lead to premature end of the
selection procedure. For instance, in Jaumouillé's method it may
happen that the relative difference between EnH and EnH +1 is zero while
the relative difference between EnH and EnH +2 is larger than the
threshold value, thus leading to a premature stop of the selection
process.”
The last two methods present in literature [16,17] and here
reviewed overcome this limitation by setting a reference value of
retained harmonics, here termed ∼ nH which ensures an adequate
representation of forces and displacements. Furthermore, the evalua-
tion criteria are not based on the comparison between two full non-
linear solutions, thus further reducing computational costs.
The criterion adopted by Grolet et al. in [16] is again the fraction of
spectral energy, however the vector used in the computation is not the
solution at a given frequency step m, it is rather a prediction of the
solution at the following frequency step m+1. This prediction q p is
International Journal of Non-Linear Mechanics 92 (2017) 25–40
given by:
q p = q + Δq ˘ (11)
where q is the solution at frequency step m obtained with a given
harmonic content (e.g. 0-nH ), padded with zeros up to ∼
nH . The vector
Δ q̆ does not require nonlinear calculations, it is computed using a
tangent predictor:
∂r ∂r
∂q
(q, ω)Δ q = − ˘
∂ω
(q, ω)Δω
(12)
This prediction is based on the assumption that both the equilibrium
and the harmonic support will not change significantly from frequency
step m to m+1. Another fundamental difference with respect to the
techniques described above is the harmonic selection procedure. Rather
than adding harmonics using a fixed increment, the harmonics for the
computation of the following frequency step are selected depending on
the harmonic support of qp . The criterion (i.e. fraction of spectral
energy) is computed for all harmonics 0 < h < ∼ nH : the following
⎞ frequency step will be computed retaining only those harmonics whose
indicator ECh is above a given threshold.
⎠ Süß et al. in [17] proposed two methods. The first method is
(8) somewhat similar to [16]. Like Grolet's it aims at selecting which
harmonics should be retained in the calculation, rather than updating
the harmonic selection by a fixed increment. The selection procedure is
once again based on a vector of displacements:
qp = q + Δ q ˘ (13)
where q is once again the solution at frequency step m obtained with a
given harmonic content (e.g. 0–nH ), padded with zeros up to ∼
nH . In this
case however Δ q̆ is a vector padded with zeros in correspondence of
the harmonics already computed (e.g. 0–nH) with the nonlinear
(9) solution, while the missing harmonics contribution is given by follow-
ing:
∼
nH /nH ∼ ∼ −1 ∼ ∼
nH /nH
Δ q̆ = (D nH /nH,nH /nH) (F HE + FCnH /nH) (14)
The ‘missing harmonics’ vector Δ q̆ is computed based on the
assumption that the contact forces vector FC, computed using a
displacement vector limited to harmonic order nH will not significantly
nH . The evaluation
(10)
criterion used by Süß et al. [17] is once again the fraction of spectral
energy for each harmonic 0 < h < ∼ nH : the frequency step m+1 is
computed retaining only those harmonics whose indicator ECh is above
a given threshold.
The second method proposed by Süß et al. [17] selects the harmonic
support at frequency step m+1, by evaluating the norm of portions of
the Jacobian of partial derivatives of contact forces with respect to
displacements. The influence of one harmonic on other harmonics is
evaluated column-wise: if the norm of column z of the Jacobian at
frequency step m is above a given tolerance, then the harmonic term
0 < z < nH is retained at frequency step m+1.
The following section is devoted to the description of the Jacobian
Alert Algorithm, the novel contributions it introduces and the simila-
rities and differences with respect to the existing techniques described
above.
4. A novel method for harmonic verification: the Jacobian Alert
Algorithm – description and comparison with existing techniques
Similarly to existing techniques the JAA too is based on the
definition of an evaluation criterion or error metric. The purpose of
this section is to guide the reader through the definition of the
evaluation criterion, starting from the equilibrium equations (Sections
4.1–4.3). Namely, the equilibrium equation will be, at first, re-written
in a linearized form using the Jacobian matrix (Section 4.1). Then the
linearized system of equations will be used to quantify the level of cross
28
C. Gastaldi, T.M. Berruti International Journal of Non-Linear Mechanics 92 (2017) 25–40

harmonic coupling associated to the presence of friction-induced interested in the corresponding harmonic components of the response
nonlinearities (Section 4.2). Finally an intuitive and practical error it is perfectly sufficient to solve the equation:
metric will be defined (Section 4.3).
ˆ 1,1)⋅q1 = J1,1⋅q1 = F1
(D1,1 + K (18)
Section 4.4 presents a step-by-step description of the algorithm EXT
while Section 4.5 compares the proposed algorithm to existing techni- h,k
If, on the other hand, some slipping occurs, the blocks K̂ with
ques reviewed in Section 3.
h ≠ k will not be empty anymore as shown in Fig. 1b and cross-
h,k
4.1. Linearization of the equilibrium equations harmonic coupling will occur. Blocks K̂ with h≠k display non-zero
entries only for contact DOFs. Reducing Eq. (15) to (18) will introduce
With reference to Eq. (6), if r i ≅ 0 then the equilibrium is satisfied an error whose magnitude will depend on the degree of cross-harmonic
and Eq. (4) can be rewritten, in its linearized version, as coupling.

J⋅q = FEXT (15) 4.2. Quantifying cross-harmonic coupling

where:
The purpose of this section is to illustrate the procedure to quantify
• ⌢
J = D + K with the error introduced by neglecting higher harmonic terms without
performing additional iterative calculations.
⎡ D0 0 0 0 ⎤ ⎡ ⌢0,0 ⌢0,1 ⌢0,nH ⎤
⎢K K ⋯ K ⎥ Cross-harmonic coupling can be quantified by using perturbation
⎢ ⎥
1 ⌢ ⎢⌢ 1,0 ⌢1,1 ⌢1,nH ⎥ analysis on Eq. (15). As an example let us assume that for a given case
D=⎢0 D 0 0 ⎥ K =⎢K K ⋯ K ⎥ ∼
⎢0 0 ⋱ 0 ⎥ ⎢ ⋮ ⋮ ⋱ ⋮ ⎥ nH = 3 is enough to perfectly represent the system dynamics. However,
⎢⎣ 0 0 0 D nH ⎥⎦ ⎢⎣⌢ nH,0 ⌢nH,1 ⌢nH,nH ⎥⎦ suppose that the user decides to solve the system retaining only the
K K ⋯ K
static term and the 1st harmonic (nH = 1).
Still, it is possible to write the Jacobian matrix up to ∼
h,k h
and K̂ =
∂FC nH = 3, since
∂qk the contact forces and their derivatives are reconstructed in time inside
the contact model (AFT procedure). Therefore, only the length of the
•F = {F 0EXT , F1EXT , ..., F nEXT
EXT
T
H } , a typical example vector of external
user-defined time step limits the number of harmonic terms available. It
forces in a typical bladed system is is then possible to retain any number of harmonics when computing the
T
FEXT = {F 0EXT , F1EXT , [0]1x2⋅(nH −1)⋅(Ndof) } Jacobian matrix.
⎡ 0 ˆ 0,0 ˆ 0,1 ˆ 0,2 ˆ 0,3 ⎤
Eq. (15) is a linearized version of Eq. (4), where the vector of ⎢D + K K K K ⎥
⌢
contact forces has been substituted with the product K (q)⋅q . ⎢K
ˆ 1,0
D1 + Kˆ 1,1
ˆ 1,2
K ˆ 1,3
K ⎥
J(ω, q) = ⎢ 2,0 2,1
⎥
⎧ F 0 ⎫ ⎡ ˆ 0,0 ˆ 0,1
ˆ0,nH ⎤⎧ 0⎫ ⎢K
ˆ ˆ
K D2 + Kˆ 2,2 ˆ
K
2,3
⎥
⎪ C⎪ ⎢K K ⋯ K ⎥⎪q ⎪ ⎢ 3,0 3,3 ⎥
⎪ F1 ⎪ ⎢ ˆ 1,0 ˆ 1,1 ˆ 1,nH ⎥ ⎪ 1 ⎪ ⎣K
ˆ ˆ 3,1
K ˆ 3,2
K 3
D +K ˆ ⎦ (19)
⎨ C⎬ = ⎢ K K ⋯ K ⋅⎨ q ⎬
⎪ ⋮ ⎪ ⎢ ⋮ ⋮ ⋱ ⋮ ⎥⎥ ⎪ ⋮ ⎪
The solution obtained by retaining only the static term and the first
ˆ nH,nH ⎦⎥ ⎪
⎪ F nH ⎪ ⎢ ˆ nH,0 n ⎪
⎩ C ⎭ ⎣K ˆ
K
nH,1
⋯ K ⎩q H ⎭ (16) harmonic, here termed q, can be seen as the solution of the perturbed
⌢ system:
Matrix K (q)can be considered as a cross-harmonic stiffness matrix
⌢h,k
since each block-entry K quantifies the contribution that the k-th (J − ΔJ)⋅q = FEXT
harmonic component of displacements qk has on the h-th harmonic ⎡ ˆ 0,2 ˆ 0,3 ⎤
component of the contact forces FCh . The structure of matrix K (q)
⌢ ⎢0 0 K K ⎥
⎢ ˆ 1,2 ˆ 1,3 ⎥
defines the degree of non-linearity introduced by friction. In a perfectly ΔJ = ⎢ 0 2,0 0 2,1 K K
⎥
linear case such as full stick it holds: ⎢Kˆ ˆ
K 0 ˆ 2,3 ⎥
K
⎢ 3,0 ⎥
h,k ⎣Kˆ ˆ 3,1 K3,2
K 0 ⎦
K̂ =0 ∀h≠k (17) T
FEXT = {F 0 EXT, F1EXT , 0 , 0} (20)
In this special case the harmonic components of forces and
displacements are perfectly uncoupled, as shown in Fig. 1a. If the user where ΔJ represents the coupling with the 2nd and the 3rd harmonics,
is providing the system with a mono-harmonic excitation and is neglected when solving the system using only the static and the 1st

Fig. 1. a) Jacobian Matrix in a full-stick case, harmonics from 0 to 5 b) Jacobian Matrix in a non-linear case with low cross-harmonic coupling, harmonics from 0 to 5 c) Matrix ΔJ in a
non-linear case with strong cross-harmonic coupling, harmonics from 0 to 5 (dashed-line rounded rectangles signal entries with higher magnitude).

29
C. Gastaldi, T.M. Berruti
harmonic (nH = 1) terms.
In Eq. (18) it should be noted that, since the 2nd and 3rd harmonic
are uncoupled from the lower terms and the exciting forces are
concentrated on the 0th and 1st harmonic, it holds: qh = 0 h > 1.
Let us now consider the system with the complete Jacobian matrix J
(up to ∼
nH = 3):
J⋅(q + Δq) = FEXT (21)
It should be noted that q, although equal to the solution of the linear
system in Eq. (20), has been obtained through the HBM Newton-
Raphson algorithm, while (q + Δq) has been obtained by solving the
linear system in Eq. (21).
Let us now analyze the differences between the solutions of Eqs.
(20) and (21). If the system is fully stuck then:
• ΔJ = 0;
• the difference between the solutions of Eqs. (20) and (21) Δq is null;
• the harmonic terms higher than one are null: (q + Δq) =0 h > 1.
h
If cross-harmonic coupling is weak:
• the terms contained in the matrix ΔJ are small with respect to
J − ΔJ
• the difference between the solutions of Eqs. (20) and (21) Δq will be
negligible with respect to q;
• the harmonic terms (q + Δq) with h > 1 are negligible with respect
h A similar error indicator can be produced considering the static
to (q + Δq)1 and q1.
Let us define vector ∼ q , the solution of (4) that would have been
obtained using the HBM retaining all the harmonics up to the 3rd
(nH = ∼ nH = 3).
It should be noted that, apart from the full stick case, (q + Δq) is not
the solution ∼ q . Similarly, matrix J (Eq. (19)) is not the Jacobian matrix
∼
J , which would have been obtained if ∼ nH = 3 had been retained in the
calculation.
However it will be shown below that (q + Δq) can be used to
evaluate how far q is from the real solution ∼ q and how different J is
∼
from the real Jacobian matrix J .
In detail if (q + Δq)1 ≅ q1:
– the contact forces identified by J display a low influence of
harmonics > 1 on harmonic terms 0 and 1 (i.e. ΔJ negligible);
– the configuration of the Jacobian matrix J is very similar to that of
∼
J , as demonstrated in Section 6.2.
– as a consequence including harmonics > 1 in the iterative solution
∼
would have led to a set of contact forces identified by J quite similar
to the set identified by J;
– it can therefore be concluded that including harmonics > 1 in the
iterative solution would not have significantly modified the output
∼
q1 ≅ q1 ≅ (q + Δq)1.
On the other hand if the (q + Δq)1 is markedly different from q1:
– the contact forces identified by J display a strong influence of
harmonics > 1 on harmonic terms 0 and 1 (i.e. ΔJ non negligible as
shown in Fig. 1c);
∼
– the configuration of the Jacobian matrix J is different from that of J.
– as a consequence including harmonics > 1 in the iterative solution
∼
would have led to a set contact forces identified by J quite different
from the set identified by J;
– it can therefore be concluded that including harmonics > 1 in the
iterative solution would have significantly modified the output. The
iterative solution needs then to be repeated, including all necessary
harmonics in order to determine the true value of
∼
q1 ≠ q1 ≠ (q + Δq)1.
30
International Journal of Non-Linear Mechanics 92 (2017) 25–40
It should be stressed that, although J is not the ‘true’ Jacobian
∼
matrix J , it is still possible to use it to compute the vector (q + Δq)1
answer the question: “is the harmonic support nH adequate to represent
the contact forces and displacements identified by J?” If the answer is
“YES” then it means the presence (or absence) of higher order
harmonics will not modify the Jacobian configuration, nor the contact
forces or the resulting displacement. If the answer is “NO”, then the
vector of displacements produced by J is not close to the true solution
∼
(produced by J ), and the computation must be repeated with an
increased harmonic support.
4.3. Error metric
In order to compare (q + Δq)1 with q1 it is necessary to define a
similarity or error metric. The maximum error over the output degrees
of freedom idout was found to be an effective indicator (i.e. evaluation
criterion):
⎛ Δq1(id ) ⎞
Err1% = max⎜⎜ 1 out
⎟⎟⋅100
⎝ q (id out) ⎠ (22)
If the discrepancy between q1 and (q + Δq)1 reaches a critical limit,
in this paper set to Err1% = 5%, it is advisable to repeat the iterative
calculation with a higher number of retained harmonics ∼
tolerance will be further discussed in Section 6.2.
nH . The 5%
term:
⎛ Δq0(id ) ⎞
Err%0 = max⎜⎜ 0 out
⎟⎟⋅100
⎝ q (id out) ⎠ (23)
however it has a very similar trend with respect to Err1%,
but it was
found to be less sensitive to discrepancies between q and (q + Δq), Err1%
was therefore preferred as an evaluation criterion.
It should also be stressed that the choice of the set of degrees of
freedom to involve in the computation of the error metric (vector idout)
can and should be tailored to the application of interest. In turboma-
chinery applications it may coincide with a given design parameter (e.g.
motion of a high-stress point(s) along a predetermined direction
connected to the mode under investigation). These authors propose
the output (linear) degrees of freedom on practical grounds, however
other choices are possible, e.g. nonlinear (contact) degrees of freedom.
Although the absolute value of the displacements at the contact is
considerably lower, the relative error as defined in Eq. (22) (and the
actual one, as defined in Eq. (29)) were found to be quite consistent
with those computed considering the output degrees of freedom, e.g. for
a 3.5% error at the tip of the blade (linear output degree of freedom)
the error at the nonlinear degrees of freedom ranges between [2.6 and
4.1]%.
Furthermore, in other applications additional set of points idout can
be chosen according to specific needs, e.g. different nodes along a
specific direction or the motion of a specific node along all its degrees of
freedom.
4.4. The Jacobian Alert Algorithm (JAA) – step-by-step description
The calculation starts by assuming that the static and the 1st
harmonic terms are an adequate choice to capture forces and displace-
ments.
It was extensively shown in [20] that the static term should always
be retained to avoid strong assumptions on the mean component of
forces and displacements. Specifically it should be noted that, while
excluding higher order harmonics implies that their value is zero (and
in most cases it is indeed negligible), excluding the static term forces
the user to make assumptions on the mean value of the force
components. This is a dangerous practice because it neglects the
static-dynamic coupling, thus producing results which are potentially
C. Gastaldi, T.M. Berruti
Fig. 2. HBM iterative procedure plus JAA sample implementation for a given frequency step.
wrong both in their dynamic and static components. This was demon-
strated in [20] where the authors showed that, since in the DTI solution
static and dynamic loads are accounted for simultaneously, a good
correlation between HBM and DTI can be obtained if and only if the
HBM equations include the static term.
This is confirmed in this paper by the structure of the Jacobian
matrix in Fig. 1b and c where it is shown that the coupling of the static
term with the harmonic components is non-negligible.
In principle, one could apply the JAA by neglecting the static term
and verifying its influence a-posteriori. In practice, however, apart from
the full-stick case, the influence of the static term is rarely negligible. It
can therefore be concluded that performing a forced response calcula-
tion based on assumed values of the static component of friction forces
is potentially a dangerous practice and the most practical choice is to
always include the static terms in the system equations.
The JAA is implemented in order to warn the user when more
harmonics are required.
The JAA, schematized in Fig. 2, can be summed up in the following
basic steps for a given frequency value.
1. Initial hypothesis: the static and the 1st harmonic terms are an
adequate choice to capture forces and displacements.
2. Compute the solution q1 through the HBM iterative procedure with
nH = 1 and the corresponding set of contact forces.
3. The alternating frequency-time procedure allows investigating all
harmonics of the set of contact forces identified by the equilibrium
at bullet point 2. Are higher harmonic terms relevant?
4. Build the system in Eq. (21) and compute (q + Δq)1.
Remark: The computational cost of solving the linear system in
(21) is very low for medium size systems. If inverting the complete
Jacobian matrix becomes too cumbersome due to the size of the
system it is possible to rely on partitioned matrix inversion
techniques and compute only the portions of inverse actually needed
to obtain (q + Δq)1.
5. Verify the assumption at bullet point 1 by analyzing the difference
between q1 and (q + Δq)1 quantified through the error indicator
Err1% of Eq. (22). In detail:
1. if Err1% < 5% then the hypothesis at point 1 is verified and the
equilibrium is compatible with the initial assumption
31
International Journal of Non-Linear Mechanics 92 (2017) 25–40
2. if Err1% ≥ 5% then the hypothesis at point 1 is not verified, the
equilibrium at point 2 is not compatible with the initial assump-
tion, the calculation is then repeated by considering more
harmonic terms: nH = ∼ nH .
The procedure proposed here can be defined a "two passages"
strategy because the solution is first computed with nH = 1 and then
recomputed with nH = ∼ nH if the error indicator is above tolerance. A
‘single passage’ strategy where the error metric is computed online and
the harmonic support is increased or decreased accordingly is also
possible. Both have been tried out, however the ‘double passage’
strategy was preferred for several reasons:
– it enables the user to first compute a ‘quick-fix’ version of the FRF
and then improve the accuracy of the solution only for the frequency
steps of interest (i.e. close to resonance);
– it avoids assumptions on the harmonic support of frequency step m
+1, based on information gathered at frequency step m, thus
avoiding limits on the frequency step size and possible rattling
effects (see also Section 4.5);
– considering a given frequency step m+1, the overall computational
time of the ‘double passage’ strategy (nH = 1 solution plus that of the
nH = ∼nH solution computed using (q + Δq) as an initial guess) has
been found, in the investigated cases, lower or equal to that of the
single passage strategy (directly with nH = ∼nH ) computed using the
solution at m as an initial guess.
A crucial part of the algorithm resides in a proper selection of ∼
nH , i.e.
the number of harmonics which ensures a good accuracy even for
strongly non-linear cases. This aspect is carefully considered in Section
6.1.
4.5. Novel contributions of the JAA and comparison with existing harmonic
selection techniques
The proposed algorithm, JAA, was developed independently from
existing techniques. Unsurprisingly it shares with them some simila-
rities, but also profound differences. Both similarities and novel
contributions are outlined below.
C. Gastaldi, T.M. Berruti
4.5.1. Similarities with respect to existing techniques
Out of all existing AHBM methods [13–17], the JAA shares a few
characteristics with the most recent and advanced techniques [16,17].
Like [16,17], it uses a reference number of harmonic terms ∼ nH to
overcome the risk of premature end of the selection procedure present
in incremental strategies such as [14].
Like [16,17] the computation of the evaluation criterion does not
involve non-linear calculations and, unsurprisingly, uses the same set of
quantities (e.g. jacobian of partial derivatives of residuals with respect
to displacements). However, the evaluation criterion formulation
proposed in this paper is profoundly different from all existing ones;
this statement will be proven in Section 4.5.4.
Furthermore, the JAA presents completely novel contributions in
terms of implementation strategy and error control, both of which are
thoroughly discussed below.
4.5.2. Novel contribution with respect to implementation strategy
To sum up all existing techniques compute a non-linear solution at
frequency step m and try to predict the optimal number of harmonics at
step m+1. To reach this goal all techniques use an evaluation criterion,
which can be more or less refined, however all are computed using
information obtained at frequency step m. It is therefore evident that all
techniques are based on the assumption that the harmonic support of
the solution at frequency step m+1 will not change significantly from
that obtained at frequency step m. This assumption is not necessarily
correct and leads to two main drawbacks:
– the frequency step size may need to be reduced excessively to
accommodate the assumption detailed above, as observed in [16];
– the changing conditions and amount of non-linearity from one step
to the other may induce ‘rattling’ effects, i.e. single harmonics are
put in and out alternatively, as observed in [17].
To overcome this limitation these authors propose a novel method
with a new implementation strategy. It abandons the idea of finding the
“optimal” number of harmonics for each frequency step. Rather, it
achieves the trade-off between accuracy and computational power by
switching between two harmonic order selections:
– 0–nH which is the minimum number of harmonics (i.e. in this paper
nH=1 to match the harmonic support of the external force FHE).
– 0–∼
nH which is the number of harmonics capable of adequately
representing non-linear forces and displacements even in highly
nonlinear cases.
This idea is based on the authors' experience with systems with dry
friction contacts and mono-harmonic excitation: it was observed that
∼
nH =5 offers an adequate representation of forces and displacements
even in highly nonlinear cases (e.g. partial detachment mixed with
slip). Conversely, the static term and the first harmonic term (0–nH) are
a perfectly adequate choice for most weakly non-linear cases (i.e.
partial slip or gross slip without detachment). By adequate these
authors mean an error lower than 1% with respect to the Direct Time
Integration FRF results, used in this paper as a benchmark.
The complete non-linear calculation (at all frequency steps) is first
performed using nH and it is repeated using ∼ nH only for those frequency
steps where the proposed algorithm signals an inadequate representation.
In other words, the authors prefer a ‘verification strategy’, rather
than the ‘predictive strategy’ proposed by existing algorithms. By
verification strategy these authors mean that the only assumption made
by the algorithm (i.e. increasing the harmonic order does not modify
the contact state) is not relied upon, rather the JAA checks its validity
and acts accordingly. Furthermore, it does not rely on the assumptions
of similarity between frequency steps. As a result, the size of the
frequency step can be set according to the user needs without any
limitations imposed by the algorithm (i.e. each frequency step is treated
32
International Journal of Non-Linear Mechanics 92 (2017) 25–40
independently).
The algorithm is obviously not affected by ‘rattling’ since only two
options are available (maximum harmonic order can be either set to
nH=1 or ∼ nH = 5). There are indeed cases where one may achieve a good
enough accuracy by employing, for instance, only three harmonics
rather than five, and these authors did consider the idea of finding those
cases. However the idea was finally rejected to avoid costly incremental
strategies such as [14]. In the JAA two non-linear calculations at most
will be performed for each frequency step. In other words, the
computational effort involved in finding the optimal number of
harmonics and verifying its adequacy (see error control in the following
paragraph) may be higher than that of directly computing the solution
using ∼nH if the contact states are suspected to be highly nonlinear.
4.5.2.1. Remark on local vs. global harmonic selection techniques. Maple
[15] and Grolet [16] propose to select a different number of harmonics
not only with different frequency steps, but also depending on the DOF,
since not all DOFs are exposed to the same level of nonlinearity. This is
certainly true and it is quite convenient in some applications. However,
the focus of these authors is on systems with localized nonlinearities
(i.e. friction contacts). In these cases the number of nonlinear DOFs is
typically much lower than the number of linear DOFs, therefore the size
of the system is already reduced by condensation as in [10]. As a result,
the size of the system to be solved is proportional to the contact DOFs
only (see also Eq. (4)). All contact DOFs are potentially nonlinear,
depending on the level of excitation and on the structure kinematics.
While the idea of tailoring the harmonic support depending on the DOF
is certainly quite valuable in other applications (i.e. cubic springs, CFD
applications etc), these authors prefer a global approach (all DOFs are
given the same harmonic support) for mainly two reasons:
– the authors want to avoid a predictive strategy (to avoid assump-
tions on the similarity between the harmonic support of different
frequency steps, both globally and in a DOF-specific fashion)
– in the authors' specific application (localized nonlinearities) the
advantage of this local selection procedure may not be worth the
computational effort of verifying the validity of the assumption.
In other words these authors find more convenient switching
between a very small system and one complete reference system,
whose size is defined by harmonic order ∼
nH and the number of contact
DOFs only.
4.5.3. Novel contributions with respect to error control
In these authors opinion a strict control over the error at each
frequency step can only be achieved by algorithms implementing:
– a number of harmonic terms and of retained DOFs that ensures a
faithful representation of forces and displacements to be used as a
benchmark (i.e. ∼nH and all potentially nonlinear DOFs).
– the capability of verifying the predicted harmonic terms used for a
non-linear calculation.
Out of all existing techniques only Jamouillé [14] and Maple [15]
accept to repeat frequency step m if the evaluation criterion exceeds the
threshold. However, since their strategies are incremental (there is no
reference number of harmonic terms ∼ nH ), the selection process may stop
prematurely, as correctly pointed out by Grolet in [16].
All other algorithms [13,16,17] update the vector of retained
harmonics at step m+1 but fail to repeat the nonlinear calculation at
step m. This practice is potentially dangerous:
– the solution at frequency step m displays a predicted harmonic
support which exceeds the one that was actually included in the
computation, therefore the result is potentially far from the truth;
– this potentially wrong result is used to predict the harmonic support
C. Gastaldi, T.M. Berruti International Journal of Non-Linear Mechanics 92 (2017) 25–40

at frequency step m+1. Like these authors, Süß (see Eq. (14)) starts from the hypothesis that
the contact states do not change with increasing harmonic order. To
For instance, if one were to start even with a high number of better illustrate the difference between the two techniques let us
harmonic terms (i.e. nH =5) far from resonance (i.e. full stick condition assume that the original computation has been performed with nH = 1
– linear case) all algorithms would drastically reduce the number of and that ∼nH = 3 ensures an optimal representation of friction forces. For
retained harmonics. If the structure were to switch between a fully both techniques (the one presented in this paper and the one proposed
stuck contact state and a mixed contact state where partial detachment by Süß) the result of the nonlinear computation in its linearized version
and slip are both present (i.e a highly nonlinear case) the system, can be expressed as:
running with a low number of harmonics may not catch this transition
⎧ F0 ⎪
⎡ J 0,0 J 0,1⎤⎧ q0 ⎫ ⎪
⎪ ⎪
⎫
correctly. For instance, it may predict an equilibrium far from the truth ⎢ 1,0 1,1 ⎥⎨ ⎬ = ⎨ HE ⎬
where contact states are not correctly represented. Consequently, it ⎣J J ⎦⎩ q ⎭ ⎩ F HE ⎭
1⎪ ⎪⎪ 1 ⎪
(24)
may take several frequency steps before the algorithm “caught up” with
the correct solution. The effects of this potential error could be These authors propose to determine the predicted vector (q + Δq)
overcome by repeating the non-linear calculation until the foreseen by solving the new linear system:
harmonic support matches the one actually retained during residual ⎡ J 0,0 ⎧ 0⎫
J 0,1 J 0,2 J 0,3 ⎤⎪ (q + Δq) ⎪ ⎧ F 0HE ⎫
minimization, however this practice may have a detrimental effect on ⎢ 1,0 ⎥ ⎪ ⎪
J1,3 ⎥⎪ (q + Δq) ⎪ ⎪ F1 ⎪
1
the efficiency of existing algorithms. For this reason, these authors find ⎢J J1,1 J1,2
⎢ 2,0 ⎥ ⎨ ⎬ = ⎨ HE ⎬
switching between nH=1 and ∼ nH = 5 the best compromise between ⎢J J2,1 J2,2 J2,3 ⎥⎪ (q + Δq)2 ⎪ ⎪ 0 ⎪
⎢⎣ J3,0 3,3 ⎥⎪ ⎪ ⎪ ⎪
efficiency and a strict error control. J3,1 J3,2 J ⎦⎩ (q + Δq)3 ⎭ ⎩ 0 ⎭ (25)
The overall level of accuracy ensured by the JAA is given by the
error with respect to DTI results (< 1%) plus the tolerance on the Süß, on the other hand, focuses on finding an approximated version
evaluation criterion set, in this paper, to 5%. These authors deem this of the missing harmonics, thus solving the smaller system.
level of accuracy perfectly sufficient considering all other sources of ⎧ 0 0⎫
uncertainty affecting numerical simulation (e.g. imperfect knowledge ⎪ q + Δq ⎪ ⎧ q0 ⎫ ⎧ 0 ⎫
˘ ⎪ ⎪
of contact parameters, level of excitation etc) and experimental tests
⎪ ⎪ ⎪
⎪ q1 + Δq 1 ⎪
˘ ⎪ 1⎪
⎪ ⎪ 0 ⎪ ⎡ J2,2 J2,3 ⎤ ⎧ 2⎫
˘
⎪ Δq ⎪
⎨ ⎬ = ⎨q ⎬ + ⎨ 2 ⎬ with ⎢ ⎥ ⎨ ⎬
used for validation purposes (e.g. measurement error, repeatability etc). ⎪ q2 + Δq ⎪ ⎪ 0 ⎪ ⎪
2
˘
Δq ⎪ ˘
⎣ J3,2 J3,3 ⎦ ⎪ 3⎪
˘
⎩ Δq ⎭
⎪ 3 ⎪ ⎪ ⎪ ⎪ 3⎪
It should be noted that this strategy is tailored for systems with ⎪ q + Δq 3 ⎪ ⎩ 0 ⎭ ⎩ Δq ⎭ ˘
friction contacts of industrial application: different limits and levels of
⎩ ˘
⎭
accuracy may be found for different applications. ⎧−J2,0q0 − J2,1q1 ⎫
⎪ ⎪
= ⎨ 3,0 0 ⎪ 3,1 1
⎬ ⎪
⎩− J q − J q ⎭ (26)
4.5.4. Novel contribution with respect to evaluation criterion
However, in doing so, it fails to update the values of q1 and q0 , and
Out of all evaluation criteria used by existing Harmonic Selection 2 3
Algorithms, only those proposed by Grolet [16] and Süß [17] use the ˘ ˘
consequently those of Δq and Δq . Eq. (26) does not record the
Jacobian of partial derivatives to avoid repeated nonlinear calculations potential effect that including higher order harmonics could have on
(as proposed by these authors). the quantity of interest q1. This choice is directly connected to another
Both are different from the technique proposed in this paper in two fundamental difference between the two techniques. As already dis-
aspects: the formulation of the error metric (or evaluation criterion) cussed in Section 3, existing algorithms implement a predictive
and in the way in which the prediction (q + Δq) is obtained. strategy, under the assumption that the harmonic support at frequency
The error metric proposed by Grolet and Süß is the fraction of step m+1 will be the same as that of the predicted vector q + Δ q . ˘
spectral energy of all harmonics (each evaluated separately). Since their These authors implement instead a verification strategy, therefore they
technique is predictive it aims at selecting which harmonics to include need two ‘homogenous’ quantities to compare: the predicted vector of
at the following frequency step, as a result their evaluation criterion displacements (q + Δq)1 and the result of the nonlinear calculation q1 at
must span multiple harmonics. These authors propose instead a the same frequency step m.
different concept based on practical considerations on the engineering
significance of results: in case of mono-harmonic excitation the first 5. Test case
harmonic of the response q1 is typically orders of magnitude higher than
all others (static term aside, see Section 4). For this reason, from a The test case used in this paper consists in the two blades with a
practical point of view, only q1 is of interest (frequently the only one damper between them since experimental results of forced response in
shown in FRFs). In this context, the influence of higher order harmonics this configuration are available to validate the numerical results (see
is relevant only in its capability of producing variations on q1, especially Section 7).
along the degrees of freedom aligned with the mode shape under The test rig for the experimental measurements of the forced
investigation. For this reason, these authors propose to compute a response is described in detail in [21] and is similar to the one sketched
‘predicted’ vector of displacements from scratch (q + Δq)1 and compare in Fig. 3a. In detail the test rig and the experimental set-up is the same,
it with q1(see Eq. (22)) as an evaluation criterion. while the blades tested in Section 7 are real blades of a gas turbine for
The second fundamental difference is in the computation of the power generation and are here substituted by simple mock-ups in
predicted vector of displacements. Both Grolet [16] and Süß [17] start Figs. 3a and Fig. 8 for confidentiality reasons. The centrifugal force
from the nonlinear solution q (padded with zeros throughout the (FCL) is applied only to the damper, to mimic the deadweights used in
missing harmonics to obtain q̆ ) and add a vector Δ q̆ , while these the laboratory setting. As in the test rig the mono-harmonic excitation
authors compute an independent new quantity (q + Δq). force (FHE) is applied on one of the two blades as shown in Fig. 3a.
˘ ˘
Grolet's q + Δ q (see Eq. (12)) is clearly different from the one
˘
proposed by these authors. In it, the vector Δq is computed to keep the 5.1. Damper model and contact characterization
residual to zero even if the frequency shifts by Δω (to go from frequency
step m to frequency step m+1). This assumption which may not The damper is considered as a rigid body with six degrees of
necessarily be representative of reality, is perfectly coherent with the freedom at its center of mass. The rigid body assumption is well
authors’ intention to use the information at frequency step m to predict justified by the bulkiness of “solid-bar” underplatform dampers, and
the harmonic support at frequency step m+1. has already been validated against specific experimental evidence on

33
C. Gastaldi, T.M. Berruti International Journal of Non-Linear Mechanics 92 (2017) 25–40

a) b)

FCL

FHE

c)

Fig. 3. a) Model of two blades. b) Sketch of a typical asymmetric flat-curved underplatform damper. c) Sketch of the contact model.

underplatform dampers [22–24]. The number of contact points is are turned in the frequency domain through harmonic balance method
chosen according to the damper geometry. In the example of Fig. 3a (see Eqs. (2) and (4))
and b, the damper is asymmetric (flat-curved), nine contact points are
D Dh ⋅qDh = FCD h(qD 0−nH , qB0−nH) + FEXT − D h 0 ≤ h ≤ nH
chosen, three contact points for each of the three damper sections along
the damper axis. Each ‘section’ of a curved-flat damper can be D Bh ⋅qBh h
= FCB (qD 0−nH
, qB 0−nH
) + FEXT − B h
0 ≤ h ≤ nH (28)
adequately represented using three contact points (two on the flat side
and one on the cylindrical side). This has been extensively demon- where
strated in [22–24] where numerical results have been compared with D Dh = − (h⋅ω)2 ⋅MD
direct measurements on dampers. A convergence study has been carried
D Bh = − (h⋅ω)2 ⋅MB + i⋅(h⋅ω)⋅CB + KB
out to choose the appropriate number of sections. The results with two,
three and five sections were investigated and no significant change on F 0EXT − D = FCL , h
F EXT −D = 0 ∀h≠0
the FRF has been detected, provided that the stiffness assigned to each F1EXT − B = FHE , h
F EXT −B = 0 ∀h≠1
contact node is scaled accordingly. This lack of sensitivity to the
number of sections is consistent with fact that the mode under Finally, as shown in Eq. (15), the equilibrium equations can be
investigation is a bending mode, where all “sections” of the blade written in a linearized form (J⋅q = FEXT ) with:
behave in the same way. T
The contact forces between damper and platforms are modelled by a – q = {q0D, q0B, q1D, q1B, ..., qnDH , qnBH} is the response vector of the
set of two 2D node-to-node contact elements [1]. In order to model 2D blades-damper system;
tangential forces, two mutually orthogonal elements are used at each – the Jacobian matrix J is composed of the dynamic stiffness matrix D
⌢
contact pair in tangential (tD) and axial damper (wD) directions as and the cross-harmonic stiffness matrix K
shown in the sketch of the contact model of Fig. 1c. In the examples ⎡ D0 0 0 0 0 0 0 ⎤
shown in this paper (see Fig. 3b and Sections 6–7 for results) nine ⎢ D ⎥
contact points have been used to model the damper-platforms contact, ⎢ 0 D0B 0 0 0 0 0 ⎥
⎢ 1
⎥
therefore nine non-linear elements (as the one shown in Fig. 3c) have ⎢ 0 0 DD 0 0 0 0 ⎥
been applied to each damper-platforms contact pair. Each contact D = ⎢ 0 0 0 D1 ⎥⌢
⎢ B 0 0 0 ⎥K
element requires calibration (normal and tangential contact stiffness ⎢ 0 0 0 0 ⋱ 0 0 ⎥
values, friction coefficients): these values have been drawn from ⎢ 0 0 0 0 0 D nDH 0 ⎥
experimental evidence (see Section 7 for further details), and have ⎢ ⎥
⎢⎣ 0 0 0 0 0 0 D nBH ⎦⎥
been used to produce the results shown in Sections 6 and 7.
⎡ ˆ 0,0 ˆ 0,0 ˆ 0,1 ˆ 0,1 ˆ 0,nH ˆ 0,nH ⎤
⎢ K DD −K DB K DD −K DB ⋯ K DD −K DB ⎥
5.2. Governing equations for a damper-blades system ⎢ ˆ 0,0 ˆ 0,0 −K ˆ 0,1 ˆ 0,1 ˆ 0,nH
ˆ 0,nH ⎥
⎢ −K BD K BB BD K BB ⋮ −K BD K BB ⎥
⎢ ˆ 1,0 ˆ 1,0 ˆ 1,1 ˆ 1,1 ˆ 1,nH ˆ 1,nH ⎥
K
⎢ DD K DB K DD − K DB ⋮ K DD −K DB ⎥
The general governing equations from Section 2 can be tailored to
= ⎢ ˆ 0,1 ˆ 1,0 ˆ 1,1 ˆ 1,1 ˆ 1,nH ˆ 1,nH ⎥
the specific case of a bladed system with friction dampers. In detail, the ⎢ K BD K BB − K BD K BB ⋮ −K BD K BB ⎥
governing equations in the time domain (see Eq. (1)), ⎢ ⋮ ⋯ ⋯ ⋯ ⋱ ⋯ ⋮ ⎥
⎢ nH,0 ⎥
⎢K ˆ − ˆ nH,0 K
K ˆ nH,1 −K ˆ nH,1 ˆ nH,nH
⋮ K −Kˆ nH,nH ⎥
MD⋅q̈ D(t) = FCD(qD, qB) + FCL DD DB DD DB DD DB
⎢ ˆ nH,0 ˆ nH,0 ˆ nH,1 Kˆ nH,1 ˆ nH,nH
ˆ nH,nH ⎥⎦
MB⋅q̈ B(t) + C B⋅q̇ B(t) + KB⋅qB(t) = FHE(t) − FCB(qD, qB) (27) ⎣−K BD K BB −K BD BB … −K BD K BB

34
C. Gastaldi, T.M. Berruti
h h h h
ˆ h,k =
and K
∂FCD ˆ h,k =
,K
∂FCB ˆ h,k =
,K
∂FCD ˆ h,k =
,K
DD k BD k DB k BB
∂qD ∂qD ∂qB
– the vector of external forces
FEXT ={FCL , [0]1xNdof B , [0]1x 2⋅Ndof D , FHE , [0]1x2⋅(nH −1)⋅(Ndof B+Ndof D) }T .
The blades displacement vector contains only the master nodes
retained in the application of a Craig-Bampton reduction technique
[18] on the FE model of the blades. In detail 13 master nodes have been
retained (nine contact nodes, one force node and three observation/
output nodes) and 25 modes have been used in the reduction. However,
as detailed in Eq. (4), only the nonlinear degrees of freedom are solved
using HBM, therefore the overall size of the system reduced to
(NdofB + NdofD)⋅(2⋅nH + 1) with NdofB=27 and NdofB=6.
The contact forces vectors FCD and FCB refer to the same set of
forces, they differ only for the coordinate system in which they are
expressed: FCD is the vector of the six generalized forces (forces and
moments) applied on the damper center of mass, resulting from the
combination of the contact forces at the contact points, while FCB is the
vector of the contact forces at the nodes on the blades platforms. It is
possible to switch from one reference system to the other through a set
of transformation matrices.
It will be shown how the lack of flexibility of the damper, combined
with the platform/damper kinematics and damper/platform angles may
induce a highly nonlinear behavior. For instance the contact patch may
partially detach (i.e. contact point lift-off [22,23]) thus triggering a
severe damper rotation and the presence of higher order harmonics (see
Sections 5.1 and 7). In this case the JAA increases the retained
harmonics close to resonance to ensure the user-defined level of
accuracy (see Fig. 4b). There exist other blades/damper systems, such
as the one available for experimental validation (see Sections 5.2 and
8), where the combination of blade platforms kinematics and damper
configuration prevents reaching highly nonlinear contact states. In this
case the JAA allows a fast yet accurate computation with nH = 1 (see
Fig. 10).
5.3. Test case for numerical validation
The existing damper is asymmetric with a flat surface on one side
and curved on the other. A first calculation was performed supposing to
be at design stage having decided the shape and the scale of the damper
but with unknown values of the angles of the damper and platforms.
For the first attempt an angle of 45° of the blade platforms (as in
Fig. 3b) was chosen with a damper mass and inertia very close to those
of already existing dampers. As sketched in Fig. 3b nine contact points
are chosen: the points L1 are those on the flat side close to the damper
Fig. 4. a) Calculated forced response with 1 (black) and 5 (red) harmonic terms. b) Calculated forced response (marker) by using the JAA. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article).
International Journal of Non-Linear Mechanics 92 (2017) 25–40
∂FCB
; apex, points L2 those on the flat side close to the damper base, points R
k
∂qB those on the curved side. The calculation of the forced response of the
excited blade (at its tip) was performed for a range of frequencies
around the natural frequency of the first bending mode. The results are
reported in Section 6 and are used for the numerical validation of the
JAA.
5.4. Test case for experimental validation
A test campaign described in detail in [21], was performed to
measure the forced response of two blades with a damper between them
as shown in Fig. 3a. In this case the damper has already been
manufactured with angles different from those in Section 5.1. Calcula-
tions and experiments were performed for a range of frequencies
around the natural frequency of the second bending mode. The results
are reported in Section 7 and are used for the experimental validation of
the design tool as a whole.
6. Numerical validation of the Jacobian Alert Algorithm
The JAA was tested after computing the forced response of the two
blades with the damper in two cases:
– nH = 1 (one harmonic term for forces and displacements)
– nH = 5 (five harmonic terms for forces and displacements)
In both cases, the static term (harmonic zero) is always retained.
The second calculation is of course time consuming and it can suffer
from convergence problems.
The calculated frequency response functions (FRFs) in the two cases
(nH=1 and nH=5) are those shown in Fig. 4a as solid lines, dashed
black and solid red. It can be observed that, for some frequencies values
(especially far from resonance), the dashed black and solid red curves
overlap. This means that for those specific frequency values and for that
specific excitation level, the same results can be obtained either with
nH = 1 or nH = ∼nH = 5.
In Fig. 4b the green markers represent the calculation performed
using the JAA to check whether one harmonic term is enough or if five
harmonic terms are needed. It can be observed that the green markers
follow the black curve when q1 ≅ ∼ q1 and switch to the red curve when
five harmonic terms are needed (i.e. q1 ≠ ∼ q1). In other words the JAA
prompts the user to perform the calculation with five harmonic terms
only when and if a standard one-harmonic calculation is not accurate
enough.
The JAA described above selects the solution with the highest
35
C. Gastaldi, T.M. Berruti
accuracy-computational time ratio out of two possible choices. One is
obtained with nH = 1, it is readily computable but potentially inaccu-
rate. The second one is obtained with nH = ∼ nH , it requires the solution of
a bigger system but ensures a higher accuracy.
The comparison between the JAA predictions and the two HBM
cases (nH=1 and nH=5) is sufficient to validate the algorithm.
However, in order for the algorithm to give valuable predictions, the
input parameter ∼ nH has to be carefully chosen.
In other words the parameter ∼ nH has to ensure a proper representa-
tion of friction forces even for highly non-linear cases. For systems
characterized by mono-harmonic excitation a practical choice, in the
authors’ experience, is ∼nH = 5. This choice is confirmed in the following
Section 6.1 by the comparison of the solution obtained with nH = 5 and
the results of direct time integration (DTI). This parameter can however
be tailored to specific systems, and additional runs of the JAA for higher
values of ∼
nH are possible for highly non-linear cases.
6.1. Comparison with direct time integration
In order to justify the assumption of five harmonic terms as a proper
choice for the parameter ∼ nH , a comparison was made between the
results obtained by HBM (five harmonics) and the DTI in terms of
contact forces and damper displacements. The comparison shown here
for strongly nonlinear cases (i.e. FCL/FCE=2 and [168–170] Hz inter-
val) guarantees a certain amount of non-linearity (contact points reach
slip and lift-off). As shown in Fig. 4a using five harmonics leads to the
correct solution which coincides with the DTI solution (blue markers).
Using only one harmonic leads to an underestimation of the amplitude
and overestimation of the resonance frequency. In order to understand
better this difference, contact forces and displacements at the damper
center of mass are plotted in Fig. 5. Displacements and forces are
reconstructed versus time, keeping the different harmonic terms. They
are then compared to the DTI results. It can be seen that the calculation
with five harmonics merges with the DTI calculation, while the
calculation with one harmonic does not represent properly the damper
displacements and contact states. One harmonic is not sufficient to
faithfully represent the complex situation undergone by the damper. In
this case higher harmonics do have an influence and cross-harmonic
coupling occurs. If higher harmonics are not retained in the iterative
calculation the HBM algorithm finds a different equilibrium with
different contact states (i.e. different Jacobian matrix).
Both the calculated displacement and the friction forces are not well
estimated with one harmonic as shown in the hysteresis cycles of
Fig. 5b and c. This difference leads to the discrepancy obtained in Fig. 4
on the amplitude and frequency of the FRF maximum.
It should be noted, as discussed above and shown in Fig. 4b, that the
JAA is indeed capable of warning the user that the solution obtained
with nH = 1 is not general in the [168–173] Hz interval.
Fig. 5. Comparison of solutions obtained by HBM and DTI. a) Damper rotation, b) and c) hysteresis cycles at different damper contact points.
36
International Journal of Non-Linear Mechanics 92 (2017) 25–40
6.2. Testing the JAA error metric
The purpose of this section is to compare the error predicted by the
JAA (Err1%) with the actual discrepancies between the solution com-
puted with nH = 1 (i.e. q1) and the one computed with nH = ∼ nH = 5(i.e.
∼
q1). For this purpose let us define as the “actual” error the quantity:
⎛ |q1(id ) − ∼ q1(id out)| ⎞
Err1%−REAL = max⎜⎜ out
⎟⎟⋅100
⎝
1
|q (id out)| ⎠ (29)
Fig. 6 plots the “predicted” error Err1% and the actual one Err1%−REAL
against frequency for two different excitation levels. Fig. 6a investigates
a weakly non-linear case (FCL/FCE=20): as a result, one harmonic is
sufficient to capture forces and displacements. In fact
Err1% = Err1%−REAL = 0 for those frequencies where the contacts are in
full stick (i.e. [178,186:188] Hz) and both the actual and the predicted
errors are below tolerance (5%) for frequencies in the interval
[179:185] Hz. Consequently, as shown in Fig. 4b the forced response
calculation can be performed entirely at nH = 1.
Fig. 6b investigates a case with stronger nonlinearities (FCL/
FCE=5). Both Err1%−REAL and Err1% are below the 5% tolerance for
frequencies in the [168:170] Hz and [176:188] Hz intervals, while
computation at nH = 5 is required in the [171:177] Hz interval.
It should be noted that in both cases, if Err1% < 5%, then
Err %−REAL < 5% and, in most cases, Err1%−REAL < Err1%.
1
This concept can be fully appreciated by analyzing Fig. 7a, which
plots the real error against the predicted error in log-log scale. It can be
observed that Err1%−REAL > Err1% only for Err1%−REAL < 0.05%. If the errors
are in the [0.5–5]% range, all points fall on the 45° slope line, with
Err1%−REAL slightly lower than Err1%. This ‘proportionality’ is lost for Err1%
above tolerance, when the two different equilibria, identified respec-
tively by nH = 1 and nH = 5 lead to markedly different contact states, as
shown in Fig. 5. This is in no way a limit since, if Err1% > 5%, the
calculation must be repeated.
The threshold can be chosen anywhere in the interval of propor-
tionality between predicted and actual error. This interval was tested
using different dampers and was found to be independent of the specific
configuration. This is to be expected since actual and predicted error
are bound to be similar for weakly nonlinear solutions (this should hold
for different applications as well, such as shrouded blades, ring dampers
etc). These authors have chosen 5% as a threshold considering all
sources of uncertainty affecting numerical simulation (e.g. imperfect
knowledge of contact parameters, level of excitation etc) and experi-
mental tests used for validation purposes (e.g. measurement error,
repeatability etc.). However lower values can be selected if desired.
Another way to demonstrate the effectiveness of the technique is to
∼
compare the Jacobian matrices J and J for a given frequency step.
Matrix J has been obtained using a vector of displacements q , padded
∼
with zeros for all h > nH . While J is the Jacobian matrix associated with
C. Gastaldi, T.M. Berruti International Journal of Non-Linear Mechanics 92 (2017) 25–40

Fig. 6. Predicted error and actual error for: a) weakly non-linear case (FCL/FHE=20), b) strong nonlinear case (FCL/FHE=2).

the ∼nH nonlinear calculation. In this paper the similarity between
matrices has been assessed in two ways. At first the relative norm
indicator is computed
∼
J−J
ΔJ% = ∼ ⋅100
J (30)
Another way to assess whether two matrices are close to identical is
to evaluate the matrix:
∼−1
Γ = J ⋅J (31)
∼ the 5% threshold assigned to Err1% has an immediate physical meaning.
In a full stick case J and J are identical, then ΔJ%=0 and Γ is equal
to the identity matrix. In a weakly nonlinear case (e.g., FCL/FHE=5,
frequency 182 Hz, with Err1% < 5%), ΔJ%=3.2% and Γ the Γ matrix is
quite similar to the identity matrix. Specifically:
– the trace of Γ = 363.002 where 363 is the size of matrix Γ ;
– the mean of the diagonal of Γ is 0.9995;
– 99.7% of the off-diagonal terms is equal to zero.
Conversely, if the error indicator is above tolerance (e.g. FCL/
FHE=5, frequency 177 Hz), then ΔJ% > 4% and Γ is less and less similar
to the identity matrix, specifically:
– the trace of Γ = 362.9 where 363 is the size of matrix Γ ;
– the mean of the diagonal of Γ is 0.994;
– 96.8% of the off-diagonal terms is equal to zero.
All the above is a further proof, together with the comparison of
displacement vectors, FRF curves (e.g. Fig. 4) and hysteresis cycles (e.g.
Fig. 5) that the error indicator is an effective harmonic verification tool.
It should be noted that the variations of the indicator ΔJ%, although
very computationally inexpensive, cannot substitute the proposed error
indicator Err1%. Its variations are too subtle to guarantee a proper
discrimination between acceptable and non-acceptable cases. Moreover
Setting a threshold for ΔJ% is more complex and its link to the accuracy
of the FRF may be application-dependent
6.3. Computational efficiency of the HBM calculation with JAA
The maximum error at resonance is here defined as the error of the
HBM calculated forced response amplitude at resonance with respect to
the amplitude calculated by DTI (considered as the benchmark solu-
tion). This error is plotted against the computational time needed to
solve a 20 frequency steps forced response (see Fig. 7b). Each marker on
each line corresponds to a different calculation of the forced response

Fig. 7. a) Actual error against predicted error. b) Maximum error at resonance against computational time over a 20 frequency steps forced response HBM computation, each dot
corresponds to a given FCL/FHE ratio. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article).

37
C. Gastaldi, T.M. Berruti
Fig. 8. a) Blades on the test rig. b) Asymmetric damper.
for a different FCL/FHE ratio. The points from right to left correspond to
increasing values of FCL/FHE, that is moving closer and closer to the
fully stuck case (linear case).
This plot highlights the efficiency introduced by implementing the
JAA during HBM calculation: the JAA selects the solution with the
highest accuracy-computational time ratio out of two possible choices
(nH=1 and nH=5).
In detail, it can be seen that in the case of the three left-most points
(corresponding to high FCL/FHE ratios) the HBM +JAA (green curve)
has the same performance (same error and same computational time) of
the standard HBM with nH = 1 (black curve) since there is no need to
introduce higher harmonic terms. From point four on, the nonlinearity
in the system increases and the calculation with standard HBM with
nH = 1 (dashed black) gives errors higher than the prescribed tolerance
(in this case 5%). From this point on the JAA starts to prompt the user
to repeat the calculation at a given frequency step with ∼ nH = nH = 5.
The maximum error (green curve with filled markers) drops to the same
values ensured by an HBM calculation performed with nH = 5 (solid red
curve). It can be observed comparing the green and the solid red curves
that the HBM computation with the JAA ensures errors (0.8%) equal to
those obtained with HBM with nH = 5, but in a fraction of the
computational time (43 min against 100 min).
This evaluation has been performed with a system whose size is
limited to 33⋅(2⋅nH + 1), where 33 is the number of nonlinear degrees of
Fig. 9. a) Amplitude optimization curve. b) Frequency optimization curve for |FCL|=442 N=40 kg. (For interpretation of the references to color in this figure, the reader is referred to the
web version of this article).
38
International Journal of Non-Linear Mechanics 92 (2017) 25–40
freedom retained in the HBM calculation (NdofB+NdofD). It is expected
that the advantage the JAA introduces in terms of computational time
will increase even more for larger systems.
7. Comparison with experimental results
As described in detail in [21], a test campaign was performed to
measure the forced response of two real blades with a damper between
them on the test rig shown in Fig. 8a (here showing dummy blades).
The damper, currently used on real turbines, is asymmetric as that in
Fig. 8b but the angles of the damper and the corresponding platforms
are different from the case in Section 6. An experimental plan was set
up for different pulling forces on the damper (centrifugal force FCL) and
different excitation forces (FHE).
In Fig. 9 the amplitude and frequency of the FRFs for Mode 2 are
plotted against versus the FCL/FHE ratio in the form of “damper
optimization curves”. The blue curves of Fig. 9 represent the experi-
mental values. Each measurement was repeated four times after
dismounting and repositioning the damper. The error bars represent
the standard deviation due to repeatability. This lack of repeatability
may be partially due to the different initial conditions (the way the
damper is loaded before the start of the experiment can lead to different
steady state cycles). All values included within the error bars are
equally probable and correct. Numerical investigation starts from one
C. Gastaldi, T.M. Berruti International Journal of Non-Linear Mechanics 92 (2017) 25–40

Table 1 different blue curves are shown to account for measurements repeat-
Contact stiffness values (for the complete interface) tuned against the experimental ability. The two cases in Fig. 10 were chosen since they correspond to
benchmark.
values of FCL/FHE < 30 where at least some of the contact points reach
Interface Normal Tangential contact Tangential contact slip. It can be noted that the two calculated curves (nH = 1 and nH = 5)
contact stiffness along tD – stiffness along wD – overlap and they match well with the experimental ones. The JAA
stiffness (N/ (N/µm) (N/µm) never suggested during the calculation the use of harmonic terms
µm) higher than one. This is a typical case where starting the forced
Cylinder on 25 20 450 response calculation with nH = 5 would have resulted in an overkill,
flat since, in this case, the contribution of high harmonic terms in the
Flat on Flat 520 450 450 damper equilibrium solution is negligible. This behavior can be
explained by the fact that, differently from the previous numerical
case in Section 6, this real damper has a geometry (i.e. combination of
specific set of initial conditions: resulting steady state solutions falling platform angles) which does not allow the flat side to slip [24] even for
within the error band should therefore be considered acceptable even if high forcing amplitudes. The behavior of the damper cannot be
far from the mean value. changed by increasing the forcing amplitude, the flat side of the
The contact stiffness values to be fed to the code are chosen with a damper will remain ‘glued’ to the corresponding platform for very
tuning process in order to match the experimental frequencies of the low FCL/FHE ratios (even lower than those which can be imposed to the
first three modes in the case where the damper can be considered fully test rig). As a consequence, the damper displacements at the contacts
stuck between the two blades (FCL/FHE > 230). The resulting values are are limited and the damper dynamic is quite close to the full stick case
reported in Table 1. Friction coefficient values were drawn from (linear case). This is confirmed by Fig. 11, where the hysteresis cycles at
independent experimental investigation on the same kind of asym- the different contact points are plotted for the case FCL/FHE=11.7 at
metric underplatform dampers [24] and have been set to 0.6. The green resonance frequency. The cycles calculated with one and five harmonics
curves in Fig. 10 are obtained by the HBM numerical calculation using overlap, proving that the contribution of harmonic terms higher than
the JAA. In the investigated cases the JAA confirms the results obtained one is so small that it does not influence the damper equilibrium
with one harmonic. position with respect to the blade. As a consequence the first harmonic
The calculation was also repeated using nH = 5, to validate the term of the contact forces given to the blades will be equal in the two
prediction of the JAA. Both calculated FRFs are compared with the cases (nH = 1 and nH = 5) thus producing the same FRF. It is therefore to
experimental ones in Fig. 10, where two (Fig. 10a) and three (Fig. 10b) be expected that one harmonic is enough to capture the damper

Fig. 10. Comparison of calculated and experimental FRFs. a) FCL/FHE = 23.4 . b) FCL/FHE = 11.7 . (For interpretation of the references to color in this figure, the reader is referred to the
web version of this article).

Fig. 11. Hysteresis cycles in the different damper contact points, at resonance frequency FCL/FHE=11.7.

39
C. Gastaldi, T.M. Berruti
behavior. This is not a desirable feature for a damper which should
display a much more non-linear behavior to be effective.
Despite the lack of a highly nonlinear behavior, this experimental
test case is quite relevant in the JAA validation. The capability of the
JAA to avoid ‘false alarms’ is as important as its capability of warning
the user whenever nH = 1 is not high enough. Specifically if the JAA
‘over-corrected’ the results, thus prompting the user to repeat the
calculation with nH = 5 even if not necessary, the resulting computa-
tional time would drastically increase. The risk of over-correction is
quite detrimental to the effectiveness of a harmonic selection/verifica-
tion technique and validating the technique against over-correction
should not be overlooked.
8. Conclusions
The paper presents an effective method of calculation of the forced
response of blades with damper to be used at design stage. The presence
of friction makes this dynamic problem nonlinear: for this reason, the
harmonic support of the solution depends on the contact state at each
frequency step. One of the major concerns is to perform HBM
calculations which are simultaneously accurate and computationally
efficient. The main novelty is the purposely developed Jacobian Alert
Algorithm (JAA), which checks the solution at each frequency step and
warns the user whenever the number of retained harmonic terms is
inadequate - i.e. whenever the solution would significantly change if
the number of harmonic terms were to increase. As a result, it enables
the user to run the simulation with a low number of retained harmonics
and to repeat the calculation only when strictly necessary.
An extensive review of existing harmonic selection algorithms
(Adaptive HBM) has been performed and all novel contributions and
advantages introduced by the JAA have been highlighted. Specifically:
– Unlike existing algorithms (which implement a predictive strategy),
the JAA is not based on the assumption that the harmonic support
will not change significantly from frequency step m to m+1.
– The JAA implements a ‘verification strategy’ where the only
assumption is that increasing the harmonic order will not modify
the contact state at the same frequency step. However, this
assumption is not relied upon, rather the JAA checks its validity
and acts accordingly.
– The JAA verifies the adequacy of the harmonic support at each
frequency step and repeats, if needed, the nonlinear calculation with
a number of harmonics fit to represent highly nonlinear cases. This
feature ensures an unprecedented strict error control which can be
tailored to the user needs through a user-defined threshold.
– The JAA implements a novel way to compute a prediction of what
the vector of displacements would look like if a large number of
harmonic terms had been retained in the HBM calculation.
This formulation is based on the solution of a simple linear
system.
– The JAA implements a novel evaluation criterion based on quan-
tities of engineering significance (i.e. the variation of the first
harmonic of the response with increasing harmonic order).
Validation of the JAA has been performed both against high order
HBM calculations and Direct Time Integration.
The computational cost of the JAA is low since it is based on the
analysis of the Jacobian matrix, already computed analytically as part
of the HBM iterative schemes and requires only the solution of a simple
linear system. The low computational cost and high accuracy make the
40
International Journal of Non-Linear Mechanics 92 (2017) 25–40
JAA a precious tool fit to be integrated in all HBM-based forced
response calculation routines.
Acknowledgements
The experimental results described in this paper have been obtained
within a research contract supported by Ansaldo Energia s.p.a.
References
[1] S. Zucca, C.M. Firrone, M.M. Gola, Numerical assessment of friction damping at
turbine blade root joints by simultaneous calculation of the static and dynamic
contact loads, Nonlinear Dyn. 67 (2011) 1943–1955.
[2] A. Kunio, S. Sakurai, T. Kudo, N. Ozawa, T. Ikeda, Evaluation of friction damping in
dovetail root joints based on dissipation energy on contact surfaces, Proc. ASME
Turbo Expo (2009) (GT2009-59508).
[3] P. Pennacchi, S. Chatterton, N. Bachschmid, E. Pesatori, G. Turozzi, A model to
study the reduction of turbine blade vibration using the snubbing mechanism,
Mech. Syst. Signal Process. 25 (2011) 1260–1275, http://dx.doi.org/10.1016/j.
ymssp.2010.10.006.
[4] 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 (2010) 082501.
[5] C.M. Firrone, S. Zucca, M.M. Gola, The effect of underplatform dampers on the
forced response of bladed disks by a coupled static/dynamic harmonic balance
method, Int. J. Non-Linear Mech. 46 (2011) 363–375.
[6] D. Laxalde, F. Thouverez, J.P. Lombard, Forced response analysis of integrally
bladed disks with friction ring dampers, J. Vib. Acoust. 132 (2010) 011013, http://
dx.doi.org/10.1115/1.4000763.
[7] E.P. Petrov, D.J. Ewins, Advanced modelling of underplatform friction dampers for
analysis of bladed disk vibration, J. Turbomach. 129 (1) (2007).
[8] L. Panning, W. Sextro, K. Popp, Spatial dynamics of tuned and mistuned bladed
disks with cylindrical and wedge-shaped friction dampers, Int. J. Rotat. Mach. 9 (3)
(2003) 219–228.
[9] C.M. Firrone, D. Botto, M.M. Gola, Modelling a friction damper: analysis of the
experimental data and comparison with numerical results, in: Proceedings of the
ESDA2006, Torino, Italy, ESDA2006-95605, July 4–7, 2006.
[10] E.P. Petrov, Explicit finite element models of friction dampers in forced response
analysis of bladed discs, Proc. ASME Turbo Expo (2007) (GT2007-27980).
[11] 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.
[12] M. Mitra, S. Zucca, B. Epureanu, Adaptive microslip projection (amp) for reduction
of frictional and contact non-linearities in shrouded blisks, J. Comput. Nonlinear
Dyn. 11 (4) (2016).
[13] D. Laxalde, Etude d’ammortisseurs non-linéaires appliqués aux roues aubagées et
aux systems multi-étages (Ph.D. thesis), Ecole Centrale de Lyon, Lyon, 2007.
[14] V. Jaumouille, J.J. Sinou, B. Petitjean, An adaptive harmonic balance method for
predicting the nonlinear dynamic responses of mechanical systems — application to
bolted structures, J. Sound Vib. 329 (19) (2010) 4048–4067.
[15] R.C. Maple, P.I. King, P.D. Orkiwis, J.M. Wolff, Adaptive harmonic balance method
for nonlinear time periodic flows, J. Comput. Phys. 193 (2004) 620–641.
[16] A. Grolet, F. Thouverez, On a new harmonic selection technique for harmonic
balance method, Mech. Syst. Signal Process. 30 (2012) 43–60.
[17] D. Süß, M. Jerschl, K. Willner, Adaptive harmonic balance analysis of dry friction
damped systems, in: Proceedings of the 34th IMAC, A Conference and Exposition on
Structural Dynamics, 2016, pp. 405–414.
[18] R.R. Craig, M.C.C. Bampton, Coupling of substructures for dynamic analyses, AIAA
J. 6 (7) (1968) 1313–1319.
[19] T.M. Cameron, J.H. Griffin, An alternating frequency/time domain method for
calculating the steady-state response of nonlinear dynamic systems, J. Appl. Mech.
56 (149) (1989).
[20] S. Zucca, C.M. Firrone, Nonlinear dynamics of mechanical systems with friction
contacts: coupled static and dynamic multi-harmonic balance method and multiple
solution, J. Sound Vib. 333 (2014) (916–926 91).
[21] A. Bessone, F. Toso, T. Berruti, Investigation on the dynamic response of blades with
asymmetric under platform dampers, Proc. ASME Turbo Expo (2015) (GT2015-
42597).
[22] M.M. Gola, C. Gastaldi, Understanding complexities in underplatform damper
mechanics, Proc. ASME Turbo Expo (2014) (GT2014-25240).
[23] M.M. Gola, T. Liu, A direct experimental-numerical method for investigations of a
laboratory under-platform damper behavior, Int J. Solids Struct. 51 (25–26) (2014)
4245–4259.
[24] C. Gastaldi, M.M. Gola, Pre-optimization of asymmetrical underplatform dampers,
J. Eng. Gas Turbines Power (2016), http://dx.doi.org/10.1115/1.403419- ISSN
0742-4795.