See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/256156971

Comparison of model reduction techniques for large mechanical systems ---

A study on an elastic rod

Article  in  Multibody System Dynamics · August 2008

DOI: 10.1007/s11044-008-9116-4

CITATIONS READS

69 379

2 authors:

Panagiotis Koutsovasilis Michael Beitelschmidt

BorgWarner Corporate Technische Universität Dresden
34 PUBLICATIONS   214 CITATIONS    96 PUBLICATIONS   222 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Book "Maschinenbau" View project

Multi Body Dynamics View project

All content following this page was uploaded by Panagiotis Koutsovasilis on 02 March 2015.

The user has requested enhancement of the downloaded file.

Multibody Syst Dyn (2008) 20: 111–128
DOI 10.1007/s11044-008-9116-4

Comparison of model reduction techniques for large

mechanical systems
A study on an elastic rod

P. Koutsovasilis · M. Beitelschmidt

Received: 30 November 2007 / Accepted: 15 May 2008 / Published online: 23 July 2008
© Springer Science+Business Media B.V. 2008

Abstract Model reduction is a necessary procedure for simulating large elastic systems,
which are mostly modeled by the Finite Element Method (FEM). In order to reduce the
system’s large dimension, various techniques have been developed during the last decades,
many of which share some common characteristics (Guyan, Dynamic, CMS, IRS, SEREP).
A fact remains that many reduction approaches do not succeed in reducing the system’s
dimension without damaging the dynamical properties of the model. The mathematical
field of control theory offers alternative reduction methods, which can be applied to sec-
ond order Ordinary Differential Equations (ODEs), derived by the FE-discretization of large
elastic Multi Body Systems (MBS), e.g., Krylov subspace method or balanced truncation.
In this paper, some of these methods are applied to the elastic piston rod. The validity
of the reduced models is checked by applying Modal Correlation Criteria (MCC), since
only the eigenfrequency comparison is not sufficient. Diagonal Perturbation is proposed
as an efficient method for iteratively solving ill-conditioned large sparse linear systems
(Ax = b, A: ill-conditioned) when direct methods fail due to memory capacity problems.
This is the case of FE-discretized systems, when tolerance failure occurs during the dis-
cretization procedure.

Keywords Model reduction · Elastic piston rod · Modal Correlation Criteria · Large sparse
linear systems · Diagonal perturbation · ANSYS · MATLAB

1 Introduction

Model reduction is a key issue for the efficient analysis and simulation of increasingly large
models obtained from different research areas, e.g., dynamical systems, circuit simulation,

P. Koutsovasilis () · M. Beitelschmidt

Faculty of Transportation and Traffic Sciences “Friedrich List”, Institute of Railway Vehicles and
Railway Technology, Chair of Vehicle Modelling and Simulation, TU Dresden, Dresden, Germany
e-mail: panagiotis.koutsovasilis@tu-dresden.de
M. Beitelschmidt
e-mail: michael.beitelschmidt@tu-dresden.de
112 P. Koutsovasilis, M. Beitelschmidt

structural mechanics, etc. The application resp. adaptation of the right reduction technique
is an important factor, since the aim is to reduce the storage and computation time require-
ments, but without damaging the dynamical properties of the model.
In the case of mechanical Multi Body Systems (MBS), a common spatial discretization
method used is the Finite Element Method (FEM). The Partial Differential Equation (PDE),
which describes the dynamics of the elastic body is transformed into a linear second order
Ordinary Differential Equation (ODE) of the form

Mẍ(t) + Dẋ(t) + Kx(t) = Bu(t), (1)

where M, D, K ∈ Rn×n are the system matrices (mass-, damping, and stiffness matrix, re-
spectively), Bu(t) ∈ Rn×1 the load vector and x ∈ Rn×1 the unknown state vector with n
Degrees of Freedom (DoF). In many cases 105 ≤ n ≤ 106 , which leads to large dimension
system matrices, and thus to vast storage and computation time needs.
The general concept of model reduction is to find a low dimension subspace T ∈ Rn×m
with m  n in order to approximate the state vector x, i.e: x = TxR + . There is always the
possibility to reduce (1) to a 1st-order ODE and then apply the reduction methods originally
designed for such systems (MIMO, SISO, etc.). The only problem is that the information
about the structure properties of (1) (mass, damping, stiffness) might be altered. Thus, struc-
ture preserving model reduction is considered.
By projecting (1) on that subspace, a lower dimension linear 2nd-order ODE is obtained

MR ẍR (t) + DR ẋR (t) + KR xR (t) = bR , (2)

with MR = TT MT, DR = TT DT, KR = TT KT being the reduced system matrices and bR =

TT B the reduced load vector of dimensions Rm×m and Rm×1 , respectively. The reduction’s
effectiveness and reliability depends on the size of .
Based on the choice of T, various techniques have been developed through the last
decades, most of which share some common characteristics and, therefore, can be divided
into three categories [1, 2]:
1. Static condensation, substructuring and modal truncation.
– Guyan reduction [3, 4]
– Dynamic reduction [5]
– Component Mode Synthesis (CMS) [4, 6]
– Improved Reduction System Method (IRS) [4, 7, 8]
– System Equivalent Expansion Reduction Process (SEREP) [9]
2. Padé and Padé-type approximations.
– Krylov Subspace Method [2, 10–15]
3. Balancing-related truncation techniques [2, 10, 16–18].
– This approach uses the idea of computing a special realization of a linear-time-
invariant system, called balanced realization.
For the purpose of this paper, the method belonging to this category was not used for the
reduction of the elastic piston rod, thus the reader is prompted to the previous denoted
citations for further information.
Comparison of model reduction techniques for large mechanical systems 113

2 Modal truncation, sub-structuring and static condensation

For the methods belonging to this category, the system matrices of the FEM discretized
linear-time invariant 2nd-order ODE (1) are partitioned into subblocks, originated by the
theory of master (m) or external and slave (s) or internal DoFs:

M ¨ +
 x̃(t) ˙ + Kx̃(t)
Dx̃(t)  = f̃,
 
[ ]mm [ ]ms
[] := , [ ] = {M, D, K} ,
[ ]sm [ ]ss
    (3)
xm f
x̃ := , f̃ := m ,
xs fs
m ∪ s = n, n = DOFtotal , m ∩ s = ∅.

In the following and without loss of generality, the undamped variant of (3) is considered,
i.e., 
D = 0.

2.1 Guyan reduction

It is assumed that there is no force applied on the internal DoFs, i.e., fs = 0 and then (3)
is solved for xs . By omitting the equivalent inertia terms (5) (“static”), the transformation
matrix for the static reduction (6) is obtained. The low-dimension subspace is in this case
represented by Tstatic :

xs = −K−1
ss (Msm ẍm + Mss ẍs + Ksm xm ), (4)
Msm ẍm + Mss ẍs = 0, (5)
   
xm I
= · xm = Tstatic · xm . (6)
xs −K−1 ss Ksm

Guyan reduction is a good approximation for the lower eigenvalues (eigenfrequencies,

eigenvectors). For high frequency motion, the effect of inertia terms is significant, thus the
method becomes inaccurate.

2.2 Dynamic reduction

Applying the Laplace transformation on the undamped system (3), the following system (7)
is obtained:
 
 2+K
−Mω   X(ω) = B(ω)
X(ω) = F(ω). (7)
The transformation matrix (8) for the dynamic reduction is then:
 
I
=⇒ Tdynamic = ,
−B−1
ss (ω)Bsm (ω)
(8)
Bi,j (ω) := −Mi,j ω2 + Ki,j , i, j ∈ {s, m} .

Guyan reduction is a special case of the dynamic reduction, i.e., when ω = 0, but the
dynamic reduction is a better candidate for approximating high-frequency motion. Still,
Tdynamic depends on the choice of an appropriate initial frequency ω, which is not a trivial
task.
114 P. Koutsovasilis, M. Beitelschmidt

2.3 Improved Reduction System Method (IRS)

IRS perturbs the static transformation by taking in account the inertia terms as pseudo-static
forces [7]. The free vibration of the undamped reduced system (2) gives:

MR ẍm + KR xm = 0 =⇒ ẍm = −M−1

R KR xm . (9)

By differentiating (6):

ẍs = −K−1
ss Msm ẍm , (10)
−1
(9), (10) =⇒ ẍs = K−1
ss Msm MR KR xm . (11)

Substituting (9), (11) in (4), the transformation matrix for IRS is obtained:
 −1

xs = −K−1 −1
ss Ksm + Kss SMR KR xm , (12)
S= Msm − Mss K−1
ss Ksm , (13)
   
xm 0 0
= TIRS xm , P= ,
xs 0 K−1
ss

TIRS = Tstatic + PMTstatic M−1

R KR . (14)

TIRS depends on the reduced mass and stiffness matrices obtained by the static reduction. In
order to minimize the error produced by this scheme, IRS could be extended to the iterated
IRS method [8], where the improved estimates MR , KR are used in the definition of TIRS
according to the subsequent iterations:

TIRS,i+1 = Tstatic + PMTIRS,i M−1

R,i KR,i . (15)

The subscript i denotes the i-th iteration. In (15), TIRS,i is the current IRS transformation
and MR,i , KR,i are the associated reduced system matrices. A new transformation TIRS,i+1
is obtained, which then becomes the current IRS transformation for the next step.
The algorithm converges to yield the eigenvalues and eigenvectors of the full system.
However, the reduced IRS stiffness matrix is stiffer than the analogous Guyan or dynamic
reduced matrix producing small deviations by orthogonality checks.

2.4 Component Mode Synthesis (CMS): Craig–Bampton

For the internal DoFs (s), the Craig–Bampton set [6, 19, 20] is introduced. It consists of
some lower eigenmodes Φ of the internal or slave structure, which are calculated having the
external or master DoFs (m) blocked, i.e.:
 

xs = Φ sin(ωt) =⇒ Kss − ω2 Mss Φ = 0. (16)

The displacement of the slave-coordinates is then given by a superposition of the master

DoFs and the Craig–Bampton modes:


l
xs = −K−1
ss Ksm xm + φ k yk = Γ xm + Φy, l  n − m. (17)
k=1
Comparison of model reduction techniques for large mechanical systems 115

Thus, the transformation matrix for the CMS method is obtained:

    
I 0 xm x
x= = Tcms m (18)
Γ Φ y y

CMS as IRS delivers good approximation results of the reduced structure.

The major drawback of CMS is that the effectiveness of the Craig–Bampton’s set highly
depends on the definition of the master’s DoF (m): by choosing two different m-sets of the
same dimension by the same structure, the Craig–Bampton modes are not the same, since
the internal structure is not the same, thus implying the m-DoF dependence.

2.5 System equivalent expansion reduction process (SEREP)

In SEREP [5, 9], the eigenmodes and eigenfrequencies of the original full model are calcu-
lated. Thus, x = Ψ q, where Ψ is the modal matrix and q the vector of modal coordinates. By
partitioning the displacement x and the modal matrix into the active (master) and omissive
(slave):
   
xm Ψm
x= =x= q (19)
xs Ψs
   
xm Ψm
=⇒ = Ψ+m xm = TSEREP xm , (20)
xs Ψs

where Ψ + −1 T
m := (Ψ m Ψ m ) Ψ m is the pseudo-inversion of Ψ m .
T

SEREP approximates high-frequency motion (eigenfrequencies and eigenvectors), up to

the predefined limit, perfectly.
The main functionality of the so far mentioned methods, belonging to the category of
“static condensation, substructuring and modal truncation”, is to catch the dominant modes
of the system in the reduced order model. This is done either
– by trying to accumulate specific eigenvalues, which dominate the long-term dynamics of
the ODE-solution (modal truncation)
– or by selecting a small set of the lowest eigenvalues
– or by demanding the total structure to be split up into smaller structures (substructuring)
– or finally by neglecting the effect of inertia terms (static reduction)
From the above mentioned techniques, only the Guyan and the CMS reduction are imple-
mented in commercial FEM software packages, with CMS being the best available one.

3 Padé and Padé-type approximations

This category corresponds to techniques developed for obtaining a reduced order model
based on projection methods [13] and the theory of moment matching and partial realization
[2, 12].

3.1 Krylov subspace method

Suppose the constant matrix  A ∈ Rn×n , a start vector b̃ ∈ Rn×1 and q ∈ N∗ . The Krylov
subspace is defined as the subspace spanned by the q column-vectors b̃,  Ab̃, . . . , 
Aq−1 b̃,
i.e.:
  
Kq  A, b̃ := span b̃, 
Ab̃, . . . , 
Aq−1 b̃ . (21)
116 P. Koutsovasilis, M. Beitelschmidt

In the case of the undamped system (22) (a proof also exists for the case of proportionally
damped second order systems [21]), it is proven that 
A := K−1 M and b̃ := K−1 f, i.e.,
 
Mẍ(t) + Kx(t) = f =⇒ TKrylov ∈ Kq K−1 M, K−1 f , (22)
 −1  −1  −1  −1 q−1 −1
TKrylov = span K f, K M K f, . . . , K M K f . (23)

Equation (1) can always be written as an input–output system if we write the outputs as
linear combination of states
y = CT x. (24)
An evidence that proves Krylov’s reduction efficiency of (1), (24) is the equality of some
input–output behavior parameters for both the full and reduced systems, under the assump-
tion of 
AR regularity; the so-called moment matching [22], where
 

AR = TTKrylov K−1 M TKrylov .

Moments mi are defined as the Taylor-coefficients of the transfer function G for the
Laplace transformation of (22), (24):
 −1
G(s) = CT s2 M + K f. (25)

It is proven that the first q moments for the full and reduced {(1), (24)} system agree. For the
first moment mR0 (reduced model) and m0 (full model), the proof is given according to [12]:
 T −1 T
mR0 = CTR K−1
R fR = CR TKrylov KTKrylov
T
TKrylov f
  −1
= CTR TTKrylov KTKrylov TTKrylov KTKrylov r0 = CTR r0

= CT TKrylov r0 = CT K−1 f = m0 .

Krylov subspace method is implemented using the classical Arnoldi algorithm and a modi-
fied Gram–Schmidt orthogonalization in order to obtain the q orthonormal basis vectors.
This method ends up with well approximated eigenmodes and eigenfrequencies. The
nonnecessity to define the partitioned master/slave DoF space is an advantage that minimizes
user intervention and moreover contributes to better numerical properties, since the band
diagonal structure of the system matrices (obtained by FEM software packages (Fig. 2) is
not damaged.

4 Elastic piston rod

The reduction methods were implemented as an independent Model Order Reduction

(MOR) interface (Fig. 12) in MATLAB. By activating this interface all necessary informa-
tion (system matrices, eigenvectors, etc.) are automatically obtained by FE software pack-
ages (in this case ANSYS), rewritten in standard matrix format and the user is free to apply
the wished condensation approach.
In order to test the efficiency and computational applicability of each method, an ade-
quately large and realistic model should be chosen. Its dimension should indicate that itera-
tive approaches for the solution of the final linear algebraic system Ax = b are necessary to
apply. Thus, the elastic piston rod was chosen.
Comparison of model reduction techniques for large mechanical systems 117

The original model was discretized with FE in ANSYS. The number of elements (tetra-
hedron SOLID95) produced is nelem = 13,868 and the number of nodes nnode = 23,835.
Each node was appointed with 3 DoFs (UX, UY, UZ—unconstrained model). The produced
system matrices (M, K) have a dimension of

dim(M) = dim(K) = (3 · nnode , 3 · nnode ) = (7,1505, 7,1505).

For the first five previously introduced methods, the master nodes respectively DoFs set is
selected as shown in Fig. 1 (black points). The m-set is selected according to standard criteria
[23], restricted though to a relative low number (m = 10); this is done in order to prove that a
possible inappropriate master node selection (number and position) vastly affects the result’s
quality of certain reduction methods.
The Krylov subspace method being mnode -free (only the number of Krylov modes q need
to be defined) appears to be advantageous concerning this restriction.
The reordering of the system matrices into block matrices according to (3) produces
a different sparsity pattern (Fig. 2) in comparison to the one produced by FEM software
packages and in this case by ANSYS (due to the wavefront solver). This specific sparsity
pattern occurs by any kind of 3D structure discretized by the FEM [1]. In Fig. 2, the mass
matrix is depicted; the stiffness matrix has qualitatively the same sparsity pattern, but with
notedly more nonzero entries.

Fig. 1 Master nodes

selection—piston rod

Fig. 2 Mass matrix produced by ANSYS (right). Reordered mass matrix (left)
118 P. Koutsovasilis, M. Beitelschmidt

It should be mentioned that the reordered stiffness matrix is a symmetric indefinite and
ill-conditioned matrix. The indefiniteness aggravates the right application of direct [24, 25]
or iterative schemes [26, 27], since most of them are meant for positive definite matrices.

5 Modal Correlation Criteria (MCC)

In order to check the validity of a reduction method, a comparison between the original (full)
and the reduced model has to be done. A useful approach is to compare the eigenvalues
(eigenfrequencies and eigenvectors) of both models by a modal analysis.
Since the FEM software packages give only the eigenfrequency information, almost all
comparisons rely on this feature. A good eigenfrequency-correlation does not always imply
the analogous eigenvector-correlation. Thus, criteria for comparing the eigenvectors are nec-
essary. This is feasible only if the reduction’s transformation matrix is known (not directly
available by commercial FEM packages). With it, eigenvectors of the same dimension can
be compared: either by expanding the reduced model eigenvector to the dimension of the
full model’s eigenvector or the opposite.
In this paper, the following criteria are applied to the elastic piston rod for the comparison
of the reduction methods [28, 29]:
– Normalized Relative Eigenfrequency Difference (NRFD)
– Modified Modal Assurance Criterion (modMAC)
– Mass Normalized Vector Difference (MNVD)
– Stiffness Normalized Vector Difference (SNVD)
– Normalized Modal Difference (NMD)

5.1 Normalized Relative Eigenfrequency Difference (NRFD)

Theory implies the eigenfrequencies of the reduced model to be higher than the eigenfre-
quencies of the original (full) model. Thus, in Fig. 3, the normalized relative difference for
the first 14 of the nonrigid body eigenvectors are depicted

reigdif = |eigsub − eigfull |2 /|eigsub |2 . (26)

The lower the value of this criterion is, the better the reduction method used is.

5.2 Modified Modal Assurance Criterion (modMAC)

This criterion gives the information about the compared eigenvectors’ angle. The eigenvec-
tors are mass-normalized and expanded or reduced to the same dimension

(Φ Tk MΨ l )2
modMACk,l = ,
(Φ Tk MΦ k )(Ψ Tl MΨ l )
Φ k : k-th eigenvector of the full model,
Ψ l : l-th expanded eigenvector. (27)

A value “modMAC = 100%” means absolute correlation; the less this value becomes,
the worst the eigenvector correlation is as shown in Fig. 4. A thumb-rule assures that a mod-
MAC correlation of a magnitude larger or equal than 80% implies a qualitative successful
reduction.
Comparison of model reduction techniques for large mechanical systems 119

Fig. 3 NRFD—piston rod

Fig. 4 modMAC—piston rod

5.3 Mass Normalized Vector Difference (MNVD)

MNVD gives the relative vector difference of mass-normalized eigenvectors

 sub − M
|M  full |2
MNVDk,l = ,
 sub |2
|M
 full : modal mass of the original model,
M
 sub : modal mass of the reduced model.
M (28)

By this criterion, all methods are identical (should be), except for CMS, which gives minor
deviations due to numeric implementation, in comparison to the other (Fig. 5).
120 P. Koutsovasilis, M. Beitelschmidt

Fig. 5 MNVD—piston rod

5.4 Stiffness Normalized Vector Difference (SNVD)

Analogously to the MNVD criterion, SNVD gives information about the relative vector
difference of stiffness-normalized eigenvectors

sub − K
|K full |2
SNVDk,l =
sub |2
|K
full : modal stiffness of the original model,
K
sub : modal stiffness of the reduced model.
K (29)

In this case, results are sensitive, concerning small deviations of the modal stiffness for
both the original and reduced model as shown in Fig. 6. This criterion resembles the
normalized relative difference comparison, since what is compared are the modal stiff-
nesses matrices for both the reduced and full model, which are eigenfrequency depen-
dent:

sub or full = Φ T Ksub or full Φ = diag ω2 .
K (30)
The smallest SNVD value depicts the best result.

5.5 Normalized Modal Difference (NMD)

NMD is a criterion that delivers important information concerning the deviation of sin-
gle coordinates (DoFs) of eigenvector pairs. The actual purpose of this criterion is
to isolate the vector coordinates that result with the worst correlation. The user is
able to define at this or near this specific node position a master DoF and reapply
the reduction procedure. Thus, the bad coordinate correlation of both vectors is mini-
mized.
The fact that this criterion is normalized makes it an important tool for modal correlation,
but it should be mentioned that its objective is clearer for smaller dimension models, where
the DoFs are easier isolated and better visualized.
Comparison of model reduction techniques for large mechanical systems 121

Fig. 6 SNVD—piston rod

The calculation is based on Modal Scale Factor (MSF), which is a scale factor according
to the principle of least-square error.

|Ψ k (r) − MSF · Φ k (r)|2

NMDk,r = , (31)
Ψ k (r)
Ψ Ti Φ j
MSFi,j = ,
Ψ Ti Ψ i
Φ k (r): r-th coordinate of the k-th full eigenvector,
Ψ l (r): r-th coordinate of the l-th expanded eigenvector. (32)

Eigenvectors Nr. 7, 8, 18 have been randomly chosen in order to illustrate the results of this
criterion shown in Figs. 7, 8, 9.
SEREP delivers the best results concerning NMD followed by Krylov and an interchange
between IRS, CMS. All these results are discussed in the last chapter.

6 Solution methods—diagonal perturbation

All the above mentioned reduction methods (except SEREP) require the matrix inverses
K−1
ss or K
−1
in order to calculate the equivalent transformation matrices. Due to dim(s) ≈
dim(n) ∈ (104 , 106 ) a direct calculation of the inverses using a matrix factorization method
(LU, Choleksy, QR, Dulmage–Mendelsohn, etc.) could lead to memory capacity problems.
Certain software packages, e.g., CSparse [24] or TAUCS [30], can efficiently handle factor-
ization methods, and thus the direct solving of large sparse linear systems. The computation
time is comparable to the time needed for a static-analysis solution by commercial FEM
software packages. Still, the applicability of a direct calculation depends on the hardware
architecture and for constantly increasing large FEM models it could not always be feasi-
ble.
122 P. Koutsovasilis, M. Beitelschmidt

Fig. 7 NMD eigenvector 7—piston rod

Fig. 8 NMD eigenvector 8—piston rod

Comparison of model reduction techniques for large mechanical systems 123

Fig. 9 NMD eigenvector 18—piston rod

For that reason, iterative methods [26, 27] are preferred. By taking, for instance, the static
method, (6) can be rewritten as
   
xm I
= x = Tstatic xm , (33)
xs T m
T = −K−1
ss Ksm (34)
=⇒ Kss T = −Ksm (35)
⇐⇒ Ax = b, A := Kss , x := T, b := −Ksm , (36)

where (36) can be iteratively solved for as many steps as defined by m

(10, 500)  dim(m)  dim(n).

Then T is known, i.e., Tstatic is known. This procedure is also applied for the other reduction
methods.
By choosing a right preconditioner depending on the structure of the matrix (incom-
plete LU-, Cholesky factorization, block Jacobi, incomplete band-diagonal in case of band-
diagonal matrices) faster convergence is achieved.
There is always the case though, by which the matrix A is ill-conditioned (large condition
number): FEM discretized linear systems, which occur due to tolerance failure by the dis-
cretization procedure. The selection of a preconditioner does not sufficiently accelerate the
convergence resulting in a large number of iteration steps, i.e. increase of the computation
124 P. Koutsovasilis, M. Beitelschmidt

time as shown below

C(A) = λmax /λmin , C: condition number, λ: eigenvector, (37)

√
NCG ≈ C, N : number of iteration steps-CG. (38)

Different kinds of techniques have been developed (e.g. deflation) [25, 31] in order to
reduce the condition number of the ill-posed matrix. Here, the diagonal perturbation method
is proposed.

6.1 Diagonal perturbation

Instead of solving the original linearized system Ax = b, we solve the perturbed system
shown below

(A + αAd )x = b, (39)
α := 10−(n+k) , (40)
  
n = max f (j ), ∀i ∈ 1, dim(A) , k ∈ Z, (41)
j ∈N

f (j ) := 10±j · aii ∈ A, floating point number form, (42)

k ≥ min(j ) − max(j ), (43)
 
Ad := diag diag(A) , diagonal matrix A. (44)

By this small perturbation of the diagonal elements of A, the eigenvalues are vastly af-
fected reducing the condition number and consequently the convergence rate. This method
produces the following numeric error:
 
(αAd )x = 10−(n+k) |Ad x|2 ,
2

which is easy to validate by using the MCC.

Diagonal perturbation is reduction and preconditioning independent. The results de-
picted in the following figures concern the Krylov subspace method.
By trying to iteratively compute the Krylov subspace, without applying the diagonal per-
turbation method, the iterative algorithm (in this case the preconditioned conjugate gradient)
fails to converge according to the applied tolerance (10−3 ), since the condition number of
the stiffness matrix is high.
By applying the diagonal perturbation algorithm, the user can choose to vary the values
of the k-parameter as long as (43) is satisfied. For this example, the following values are
chosen:
k = {−2, −1, 0, 1, 2}.
The value k = 2 represents a small perturbation for the diagonal entries of the elastic piston
rod’s stiffness matrix, i.e., the condition number is almost unaffected, thus the simulation
time is vast (but the algorithm does not fail as in the previous case). By reducing the values
of k, a faster convergence rate is achieved. A thumb rule of 0 ≤ k ≤ 2 assures a satisfactory
convergence rate keeping intact the dynamic properties of the model, as shown in Figs. 10
and 11.
Comparison of model reduction techniques for large mechanical systems 125

Fig. 10 Computation time and convergence rate

Fig. 11 Frequency and MAC comparison

126 P. Koutsovasilis, M. Beitelschmidt

The case k = 0 reduces the computation time by 33.5%, without affecting the result’s
quality, whereas the choice k = −1 delivers a computation time reduced by 48.35%, but
damaging the eigenfrequency and eigenvector correlation Figs. 10, 11.
Since diagonal perturbation solves a perturbed and not the original system it is impera-
tive that the obtained solution to be validated by MCC, such as in this case Figs. 10, 11.

7 Coupling FE reduced models into MBS codes

Commercial MBS software packages contain implemented interfaces for the FEM-MBS
coupling [32]; FEMBS [33] is such an interface implemented for SIMPACK. It offers two
options for the FEM-MBS data transfer, namely either using the Guyan or the CMS reduc-
tion. The FEM data are then written in the so-called “Standard Input Data” (SID) format
[34], which is then used for importing the MBS properties in SIMPACK (Fig. 12).
In this paper, the quality of each reduction method is based on the correlated accuracy
by a modal analysis for the reduced and the full model. The reason is that the data required
for the SID construction are actually modal analysis results-data of the reduced model [35,
36], e.g., system matrices (free, or free and fixed for fixed structures), modal displacement
matrix, eigenfrequencies, master nodes, and their position with respect to a certain reference
frame, etc. Thus, if the correlation of these data is good (this step is done prior the SID gen-
eration), then the imported into the MBS-code model represents accurately the original FE
model independent of applying or not later-on constraints or forces on some of the model’s
DoF.
Here, the case of reducing and correlating free-structures is considered. In [35, 36], con-
strained models are examined and it is shown again that the Krylov Subspace Method (KSM)
is an approach that offers qualitatively better results than the already standardized CMS
method. There, an independent SID interface has been built (Fig. 12), which allows the
utilization of reduced data generated by any kind of condensation procedure. In case of
applying a method, which does not use the notion of physical coordinates (e.g., KSM or
balanced truncation) an extra transformation is needed in order to (re-)allocate the Krylov
or balanced truncation DoFs into the wished physical coordinates set (equivalent to the
m-set) without damaging the reduction’s quality; not possible by any of the other meth-
ods, since the m-set should be defined before the beginning of the reduction’s process. After
this allocation procedure the model is imported into a MBS code and at these reallocated
DoF forces or constraints can be normally applied.

8 Results and discussion

MCC lead to the conclusion that SEREP and Krylov deliver the best eigenfrequency/eigen-
vector results (Table 1).

Fig. 12 FEM-MBS interface workflow

Comparison of model reduction techniques for large mechanical systems 127

Table 1 Relative advantages and

disadvantages for each MOR MOR Criteria
Low freq. Middle freq. High freq. Master DoF
motion motion motion dependence
√
Guyan × × YES
√
Dynamic × × YES
√ √
CMS × YES
√ √
IRS × YES
√ √ √
SEREP YES
√ √ √
KSM NO

Guyan reduction is by far the least reliable method for approximating high-frequency
motion due to its static nature.
IRS, as a perturbed Guyan method, ends up with good correlation results for the lower
as well as a great number of higher-frequency motions. The IRS transformation matrix can
always be optimized by additional iteration steps [8] reducing the modeling error, and thus
giving good approximation results.
The algorithms of both CMS and dynamic reduction promise good correlation for high-
frequency motion, but in many cases this is not feasible due to the following reasons:
– The error of dynamic reduction depends on the right choice of the initial frequency (8),
the finding of which is not a trivial task.
– CMS depends on the definition of a sufficient number of eigenmodes (Craig–Bampton
modes) for the internal structure, the effectiveness of which is a consequence of the m-set
selection.
All the above results could have been improved, if a different set of master DoFs were
chosen. This fact enables Krylov as a promising reduction method for mechanical MBS,
since there is no such dependence (Table 1). The user has only to define the maximum di-
mension of the reduced system (Krylov modes q), without having to select dominant eigen-
modes or master DoFs. Thus, user intervention is minimized [35, 36].
The diagonal perturbation method is proposed for the efficient solution of large sparse
linear systems obtained by FEM discretization in case of failure, when applying a direct
factorization approach. By this method the iteration procedure is speed up, keeping intact—
up to a point (definition of the k parameter)—the dynamical properties of the model.

References

1. Koutsovasilis, P., Beitelschmidt, M.: Model reduction comparison for the elastic crankshaft mechanism.
In: Proc. 2. International Operational Modal Analysis Conference (IOMAC), vol. 1, pp. 95 –106, Copen-
hagen (2007)
2. Benner, P.: Numerical linear algebra for model reduction in control and simulation. Mitteilungen Ges.
Angew. Math. Mech. 29(2), 275–296 (2006)
3. Guyan, J.: Reduction of stiffness and mass matrices. AIAA J. 3, 380 (1965)
4. Myklebust, L.I., Skallerud, B.: Model reduction methods for flexible structures. In: 15th Nordic Seminar
on Computational Mechanics, Aalborg, Denmark, October 2002
5. Gloth, G.: Vergleich zwischen gemessenen und berechneten modalen Parametern. In: Carl-Cranz
Gesellschaft e.V, Oberpfaffenhofen (2001)
6. Craig, R., Bampton, M.: Coupling of substructures in dynamic analysis. AIAA J. 6, 1313–1319 (1968)
128 P. Koutsovasilis, M. Beitelschmidt

7. O’Callahan, J.C.: A procedure for an improved reduced system (IRS) model. In: Proc. 7. International
Modal Analysis Conference, Las Vegas (1989)
8. Friswell, M.I., Garvey, S.D., Penny, J.E.T.: Model reduction using dynamic and iterated IRS techniques.
J. Sound Vibr. 186, 311–323 (1995)
9. O’Callahan, J.C., Avitabile, P., Riemer, R.: System equivalent reduction expansion process (SEREP). In:
Proc. 7. International Modal Analysis Conference, Las Vegas (1989)
10. Antoulas, A.C.: Approximation of Large-Scale Dynamical Systems. Advances in Design and Control,
1st edn. SIAM, Philadelphia (2005)
11. Krylov, A.N.: On the numerical solution of the equation by which in technical questions frequencies of
small oscillations of material systems are determined. Izv. Acad. Sci. USSR, Otd. Mat. Estestv. Nauk.
2(4), 491–539 (1931) (original text in Russian)
12. Lohmann, B., Salimbahrami, B.: Ordnungsreduktion mittels Krylov-Unterraummethoden. Automa-
tisierungstechnik 52, 1 (2004)
13. Gildin, E.: Model and controler reduction of large-scale structures based on projection methods. Ph.D.
thesis, Faculty of the Graduate School of the University of Texas at Austin, December 2006
14. Beitelschmidt, M., Koutsovasilis, P., Quarz, V.: Zur modellierung und simulation der kolbenmaschinen-
dynamik unter berücksichtigung von strukturelastizitäten. In: Proc. ANSYS Conference and 24. CAD-
FEM Users Meeting, Stuttgart (2006)
15. Lehner, M., Eberhard, P.: A two-step approach for model reduction in flexible multibody dynamics.
Multibody Syst. Dyn. 17, 157–176 (2007)
16. Benner, P., Mehrmann, V.: Dimension Reduction of Large-Scale Systems. Lecture Notes in Computa-
tional Science and Engineering, vol. 45. Springer, Berlin (2005)
17. Chahlaoui, Y., Lemonnier, D., Vandendorpe, A., Van Dooren, P.: Second-order balanced truncation. Lin-
ear Algebra Appl. 415, 373–384 (2006)
18. Stykel, T.: Balanced truncation model reduction of second order systems. In: Proc. 5. MATHMOD,
Vienna (2006)
19. Craig, R.R. Jr.: Coupling of substructures for dynamic analyses: an overview. AIAA-2000-1573 (2000)
20. Benninghof, J.K., Lehoucq, R.B.: An automated multilevel substructuring method for eigenspace com-
putation in linear elastodynamics. SIAM. J. Sci. Comput. 25(6), 2084–2106 (2004)
21. Eid, R., Salimbahrami, B., Lohmann, B., Rudnyi, E.B., Korvink, J.G.: Parametric order reduction of pro-
portionally damped second order systems. Technical Reports on Automatic Control, TRAC-1, October
2006
22. Lehner, M., Eberhard, P.: Modellreduktion in elastischen Mehrkörpersystemen. Automatisierungstech-
nik 54, 4 (2006)
23. Waltz, M.: Dynamisches Verhalten von gummigefederten Eisenbahnrädern. Ph.D. thesis, Technische
Hochschule Aachen, Fakultät für Maschinenwesen (2005)
24. Davis, T.A.: Direct Methods for Sparse Linear Systems. Fundamentals of Algorithms. SIAM, Philadel-
phia (2006)
25. Manteuffel, T.A.: An incomplete factorization technique for positive definite linear systems. Math. Com-
put. 34(150), 473–497 (1980)
26. Sami, A., Seid, F., Sameh, A.: Efficient iterative solvers for structural dynamic problems. Comput. Struct.
82, 2363–2375 (2004)
27. Saad, Y.: Iterative Methods for Sparse Linear Systems, 2nd edn. SIAM, Philadelphia (2003)
28. Reichelt, M.: Anwendung neuer Methoden zum Vergleich der Ergebnissen aus rechnerischen und exper-
imentellen Modalanalyseuntersuchungen. VDI Berichte 1550, 481–495 (2000)
29. McGowan, P.E., Angelucci, A.F., Javeed, M.: Dynamic test/analysis correlation using reduced analytical
models. Technical report, NASA Technical Memorandum 107671, September 1992
30. Toledo, S.: TAUCS a library of sparse linear solvers. Technical report, School of Computer Science,
Tel-Aviv University (2003)
31. Burrage, K., Erhel, J., Pohl, B.: A deflation technique for linear systems of equations. Technical Report
94-02, Eidgenössische Technische Hochschule Zürich (1994)
32. Schwertassek, R., Wallrap, O.: Dynamik flexibler Mehrkörpersysteme. Grundlagen und Fortschritte der
Ingenieurwissenschaften. Vieweg, Wiesbaden (2005)
33. SIMPACK. FEMBS. Technical report, SIMPACK Release 8.6, November 2004
34. Wallrap, O.: Standardization of flexible modeling in multibody system codes, part I: definition of stan-
dard input data. Mech. Struct. Mach. 22(3), 283–304 (1994)
35. Koutsovasilis, P., Quarz, V., Beitelschmidt, M.: FEM-MKS coupling: transfering elastic structures into
the SID file-format by the use of Krylov subspace method. In: Proc. ANSYS Conference and 25. CAD-
FEM Users Meeting, Dresden (2007)
36. Koutsovasilis, P., Quarz, V., Beitelschmidt, M.: Standard input data for FEM-MBS coupling: importing
alternative model reduction methods into SIMPACK. Math. Comput. Model. Dyn. Syst. (2008, submit-
ted)

View publication stats