Model Order Reduction Techniques for Damped Vibrating Systems -

the Moment-Matching Approach

Martin Buchschmid / Raul Rodriguez Sanchez / Gerhard Muller

Technische Universitat Munchen, Chair of Structural Mechanics
Arcissstrae 21
80333 Munich, Germany
In Computational Mechanics one comes across with models based on finite element formulations, which
often need to be repeatedly solved, e.g. in the scope of frequency response analyses in structural dynamics
and vibroacoustics. Usually, such problems represent both a data storage and time expensive task, when
the whole system is considered and classical solvers are used. In the last decades several Model Order
Reduction (MOR) techniques have been developed for decreasing the computational cost and the memory
store of data, though maintaining the main features of the original model.
This work presents moment-matching MOR techniques, namely Pade approximation, Krylov-based Galerkin
Projection method and Derivative-based Galerkin Projection. A numerical example of a vibroacoustical
problem is presented. The results show the suitability of the moment-matching methods for non symmetri-
cal and non proportional damped systems and highlights the time efficiency and the suitability for parallel
implementation of those methods.

Keywords: Model Reduction, Krylov Subspace Method, Structural Acoustics, Frequency

Response, Moment-Matching, Pade approximation.

1 Introduction
This work deals with a vibroacoustical problem; i.e. the frequency response of a fluid contained in
a five fixed-faces box subjected to a pressure load applied at the free face of the box. The equation
of this fluid-structure interaction problem is given by
          
MS 0 uS (t) 0 0 uS (t) KS RS uS (t) f S (t)
+ 0 D
RTS MF p(t) F p(t) + 0 KF p(t) = f F (t) (1)

where MS , MF , DF , KS and KF are the structure mass matrix, fluid mass matrix, fluid damping
matrix, structure stiffness matrix and fluid stiffness matrix respectively. RS is the coupling matrix,
and u(t), p(t) f S (t) and f F (t) are the displacement vector of the structure, the discretized pressure
field, the load vector acting in the structure and the load vector  acting
 in the fluid respectively.

Applying a Fourier transformation to eq. (2), with uS () = F uS (t) , p() = F p(t) , fS () =
   
FS f S (t) and fF () = FF f F (t) leads to
     !    
MS 0 0 0 KS RS uS ( j ) fS ( j )
2j + i j 0 D + 0 p( j ) = fF ( j ) (2)
RTS MF F KF

or
( j 2 M + i j D + K)u( j ) = f( j ) j = 1, 2, 3...n (3)
where u = [uS p]T and f = [fS fF ]T and M, D, K RNN are the total mass, total damping and
total stiffness matrix respectively. Those matrices are non symmetric. f RN1 is the load vector
(force vector) and u RN1 is the displacement vector. n is the number of times that u() is
computed in eq. (3) in the scope of frequency response analysis. If n and N are large numbers
solving eq.(3) for u( j ) becomes computationally expensive, in these cases one may search for
a reduced-order model which would lead to a lower computational time. Such a reduced-order

1
model is achieved using a suitable MOR technique. The main idea of MOR techniques is to find
a vector space spanned by the columns of V CNnr , with nr  N, which maps a reduced set of
degrees of freedom (dofs) ur Cnr 1 into the original set of degrees of freedom u, such that
u u = Vur (4)
where u is the approximation of u. In the following the approximation will not be labeled any-
more. By substituting the approximation of u given by eq. (4) in eq. (3), there exists an error er
defined by
er = (2 M + iD + K)Vur f (5)
This error is not contained in the Ran(V), therefore it is said to be in the null space of V, i.e. the
transpose of V is orthogonal to er
 
VH (2 M + iD + K)Vur f = 0 (6)

where H is used for the conjugate transpose of a matrix. Eq. (6) is equivalent to performed
Galerkin projection [7] or applied the least square method to a overestimated system of equations
resulting from substituting eq. (4) in eq. (3). From eq. (6) the following reduced-order model is
achieved:
( j 2 Mr + i j Dr + Kr )ur ( j ) = fr ( j ) (7)
where Mr , Dr , Kr Cnr nr and fr Cnr 1 are defined by
Mr = VH MV Dr = VH DV Kr = VH KV
(8)
fr = VH f
For the construction of that matrix V the modal analysis is not applicable due to the unsymmetrical
nature of matrices in eq. (3). Another way before the application of a Modal analysis with real
modes would be the symmetrization of the matrices in eq. (3), but it increase the dimensions of the
problem. One is looking for the opposite and also the damping is not proportional, thus a modal
analysis is not an appropriate option for this kind of problem. Condensations methods, e.g. Guyan
Reduction, Dynamic Reduction etc. would not lead to an improvement in the computational time
of eq. (3) due to the loss of sparsity which is possessed by the original system. In this work
the matrix V is constructed using moment matching methods. Those methods have been used in
system and control field and now they are applied for second order dynamical systems, e.g. the
Krylov Subspace method was reformulated in 1989 by Craig and Su [1] to solve second order
dynamical systems, thus the reduced-order model preserves the second order form of the original
model, without transforming the second order dynamical model into the first order formulation
(state space form), which destroys the meaning of the original model. It turns out that the Krylov
subspace method results in a very efficient method in computational cost and its parallelization
could be easily achieved. In this contribution four MOR techniques are reviewed and classified as
follows
1. Modal Decomposition Method
Complex Modal Analysis (CMA)
2. Interpolatory Methods
Multi-point Pade Approximation (mP)
Krylov Subspace Method (KSM)
Derivative-based Galerkin Projection (DGP)
The general procedure of a frequency response analysis using MOR techniques is summarized as
follows:
1. Define a frequency interval = [L R ]
2. Define a set of reduced dofs ur and construct the basis for the matrix V, see secs. 2 - 3, such
that Vur approximates the physical coordinates u.
3. Compute the reduced force vector and the reduced matrices of eq. (8)

2
 4. Compute ur () in eq. (7) and use eq.(4) to get u in the interval .
(a) Define n frequency points in , those frequency points are denominated j .
(b) Compute ur ( j ) n times, i.e. for each j
 1
ur ( j ) = 2j Mr + i j Dr + Kr fr j = 1, 2, 3, ..., n (9)

(c) Project the vector ur into the Ran(V) to get the solution u( j ) of the original model
using (4):
u( j ) = Vur ( j ) (10)

2 MOR: Modal Decomposition Methods

This section presents an overview of the classical modal decomposition methods
Complex Modal Analysis

2.1 Complex Modal Analysis

In case of non-proportional damping a reduced decoupled system is obtained using a complex
modal analysis, i.e. using complex eigenvectors and eigenvalues. Therefore the second order
ordinary differential equations given by eq. (2) is expressed as a 2N-dofs first order differential
equation, i. e. the state-space formulation
z(t) = Az(t) + B f (t) (11)

where the state vector z R2N1 , the state matrix A R2N2N , the input matrix B R2NN and
the input vector f R2N1 are defined by
     
u 0 I 0
z = u ; A = ; B= (12)
M1 K M1 D M1

Considering the free vibration problem, i.e. the external force vector f equal to zero and using a
solution of the form z = nk=1
r
ek k leads to

k k = Ak k = 1, 2, ..., nr (13)
which is the standard eigenvalue problem. A is a real and nonsymmetric matrix, therefore the
2N eigenvalues must either be real or they must occur in complex conjugate pairs and the modal
matrix is composed by N eigenvectors and their N complex conjugates. The right eigenvectors
(eigenvectors of A) j and k are not orthogonal with respect to the matrix A. Nevertheless the
eigenvalues of A and the eigenvalues of AT are the same. Furthermore the right eigenvectors k
, with k = 1, 2, ..., nr , are biorthogonal with respect to the matrix A to the left eigenvectors j
(eigenvectors of AT ), with j = 1, 2, ..., nr , [9]. For convenience the normalization of the eigenvec-
tors is done such that
Tj k = k j , j, k = 1, 2, ..., nr (14)
holds true. Then nr right eigenvectors of A are used as a basis for a vector space which contains z
approximately, i.e. the expansion theorem permits to represent the state vector z as a combination
of the right eigenvectors of A
nr
z= k k (t) (15)
k=1

or in the frequency domain

nr
z= k k (i) (16)
k=1

3
  
where k (i) = F k (t) . Substituting eq. (16) in the Fourier transform of eq. (11) and multi-
plying it by the conjugate transpose of the left eigenvectors H
k leads to

i k (i) = k k () + H
k Bf(i) (17)
or
H Bf(i)
k (i) = k (18)
i k
where the set of modal coordinates = [ 1 , 2 , .., n ]T Cnr 1 , and the set of right eigenvectors
r
= [1 , 2 , ..., nr ] C2Nnr are the reduced set of dofs ur and the matrix V in eq. (4) respectively
H
and the set of left eigenvectors = [1 , 2 , ..., nr ]H Cnr 2N is VH in eq. (6) and (8). Because
of the increase in the dimension of the problem, CMA is a computationally expensive procedure
and will not be consider in the numerical example of sec. 4.

3 MOR: Moment Matching methods

The standard formulation of the second order dynamical system, (3), in the frequency domain is
given by
(s2 M + sD + K)u(s) = fp(s)
(19)
y(s) = CT u(s)
where s = i, CT will be considered here equal to the identity matrix I and the function p(s) equal
to one in case of a response having the same form of the excitation force, i.e. the input vector is
constant. The transfer function of eq. (19) is defined by
H = C(s2 M + sD + K)1 f (20)
The main approach of an interpolatory MOR method is to match the first p terms of the Taylor
expansion of eq. (20) about p with p = 0, 1, 2, ..., P, i.e. about P frequency points in a frequency
interval = [L , R ], which is the same as to find an approximated vector u which matches
the first terms of the Taylor expansion of the vector u about the same p frequency points in the
context of structural mechanics problems, i.e.
!
P
u(s) = u (s) + O (s p ) p
(21)
p=1

where s p = s s p and 
p is the
 order of the Taylor series expansion at the interpolation frequency
(k) d k
p . Thus u (s p ) = ds k u(s) . Therefore the main task for the Moment-Matching MOR methods
is to find a vector space that contains the first moments (see, sub-sec. 3.2) of H . In the case of
the multi-point Pade approximation such an approach is simplified to find the coefficients of the
Pade approximation [L/M] of a single dof, which implies to match the first L + M + 1 moments
of the transfer function, eq. (20). Due to the analogy to the Pade approximant those methods are
also known as Pade type methods [2] [3]. Here a similar terminology as in [7] was adopted where
those methods a labeled as Interpolatory MOR methods.

3.1 Multi-Pade approximation

Every dof u j (s), j = 1, 2, 3, ..., N, of the displacement vector can be approximated by a Pade
approximant about a frequency point 0 . The Pade approximant of the jth dof has the form
p0 + p1 0 + + pL L0
u j () = [L/M] = + O (L+M+1
0 )
1 + q1 0 + + qM M0
Pj ()
= + O (L+M+1
0 ) (22)
Q j ()

4
where = 0 + , Pj () and Q j () are polynomials of degree L and M respectively. By Cross-
multiplying eq. (22) and disregarding the error of the right hand side leads to
u j ()Q j () = Pj () (23)
Successively differentiating eq. (23) with respect to the following equation is obtained [6]
!
k M L
k! (kr) m! mr l!
r(k r)! u j (m r)! qm pl lk = 0 (24)
r=0 m=r l=k (l k)!
k = 0, 1, . . . , L + M
with
qm = 0 i f r > M
(25)
pl = 0 i f k > L
  
dk
where ukj = dk
u(0 ) and = 0 . Then the coefficients of the polynomials Pj and Q j
j
can be computed if u j and its first L + M derivatives are known, i.e. the Pade approximant given
by eq.(22) matches the first L + M + 1 terms of the corresponding Taylor series about 0 , therefore
the Pade approximant matches the first L + M + 1 terms (moments) of a power expansion of the
transfer function given by eq. (20) [2].
It is said to be a multi-point Pade Approximation [6] if P frequency points are defined and
interpolated in the frequency interval of interest, , where each frequency is denoted by p with
p = 1, 2, 3, ..., P, using
!
k M L
k! (kr) m! l!
r!(k r)! u j,p (m r)! qm p (l k)! pl lk
mr
p =0 (26)
r=0 m=r l=k
 
L+M+1
k = 0, 1, . . . , 1
P
(k)
where d e is the ceiling function and u j, p is the k derivative of the jth dof of the displacement
vector u evaluated at p and is defined by

K1 f k=0
1 


(1)
(k) (k)
u j, p = (u p ) j = K (2 p M iD)u p + f k=1 (27)
 
K kM(2 p u p + (k 1)u(k2)
1 (k1) (k1)

ikDu k2

p p

where p = p 0 and K = (2p M + iD p + K). Eq. (27) implies to solve P times a gen-
l m
eralized impedance problem with L+M+1 P different right hand sides.
From eq. (27) it is possible to obtain the coefficients of the Pade polynomials for each component
u j of u, those coefficients are collected in the vector x = [p0 p1 p2 ... pL q1 q2 ... qM ]T and are
found by solving Ax = b which is formed as follows:
 
A1 [n n ]
P b1
..
.
. ..


A  b
x[nP 1] = p (28)

p [n nP ]
.. .
..

. 


AP [n n ] bP [n 1]
P
P [nP nP ]

d (L+M+1) e(L+M+1) d (L+M+1) 1

where A p C P , bp C , p = 1, 2, . . . , P, n = d L+M+1
P
P e and nP = L + M +
1. A is a symmetric matrix but ill-conditioned, thus this approach can be used only when few
interpolation points are considered [2].

5
 l m
L+M+1
A p and b p are computed by expanding eq. (26) for k = 0, 1, . . . , P 1, i.e the row k + 1 of
A p is an equation given by

(k)! (k+1)! 1 (k+2)!

0! pk 1! ( p ) pk+1 2! (
L!
)2 pk+2 (Lk)! ( p )Lk pL
p 
+ kr=0 r!(kr)!
k! 1!
(1r)! ()1r u(kr) q1
  (29)
+ kr=0 r!(kr)!
k! 2!
(2r)! () 2r (kr)
u q2
 
(k)
+ kr=0 r!(kr)!
k! M!
(Mr)! ()
Mr (kr)
u qM = u p

Then the coefficients of the left hand side of eq. (29) are placed in the row k + 1 of A p , and the
right hand side is placed at the k + 1 position of the vector b p . It permits to express eq. (26) in a
matrix notation of the form [A p ][n nP ] [x][nP 1] = [b p ][nP 1] as follows

(0)
up
p0 (1)

p1 u(2)
p
up
p2
1 ( p ) ( p )2 ( p )L 2 M .
 ( p )u p   ( p ) u p   ( p ) u p  ... ..

(0) (1) (0) (1) (0) (1)
0 1 2( p ) L( p )L1 u p +( p )u p 2( p )u p +( p )2 u p M( p )M1Up +( p )M u p pL = u(L)

p
.. .. .. .. .. .. .. .. .. q1 (L+1)

. . . . . . . . . q u p
2 (L+2)
. u p
.. .
.
qM .
(M)
up
(30)
Thus P small systems of equations, which share the same unknown vector x, of the form given
by eq.(30) are assembled, i.e they are gathered in the assembled-system of equations given by
eq.(28), which permits to compute the coefficients of the multi-point Pade approximation with a
very low computational cost.

3.2 Krylov-based Galerkin projection Method

First Order approach
The second order dynamical system in eq. (19) is reformulated in the standard first order form,
see sub-sec. 2.1, and given by
s z(s) = Az(s) + Bp(s)
(31)
y(s) = CT z(s)
Where z R2N1 , A R2N2N , B R2NN and C R2NN are the state vector, the state matrix,
the output matrix and the output matrix respectively and are defined by
       
u 0 I 0 C
z = u A= B= C = 0 (32)
M1 K M1 D M1 f

The transfer matrix of eq. (31) is given by

h(s) = CT (sI A)1 B (33)

If A is non singular, the transfer function h(s) can be approximated about s0 = 0 using the Taylor
series as follows
 
T 1 1 1 1 i 1 i
h(s) = C (A B) + (A )(A B)s + + (A ) (A B)s + (34)

For convenience one may define

   
1
H = A = KI D
1 K1 M b=A B= K1
1f (35)
0 0

and
b0 = K1 f H 11 = K1 D H 12 = K1 M (36)

6
The resulting non-negative coefficients of the series (34) are said to be the systems moments

mi = CT (A1 )i (A1 B) = CT (H )i b, i = 0, 1, 2, . . . (37)

where mi C2N1 is the ith moment of eq. (33). It is clear from eq. (37) that z(s) of (31)
is contained in a vector space spanned by the columns of the Krylov subspace K (H , b) with
starting vector b and matrix H . A simplification of the dimension of the vectors in the Krylov
subspace can be done if one notes that
       
1 b0 H + H b
m0 = b = K f = 0
2
m2 = H b = 11 b 1 12 b0 = b2
0 b1 1
       
m1 = H b = H 11 b0 = b1 m3 = H b = H 11 b2 + H 12 b1 = b3
b 3 b
b0 0 b1 2

therefore the displacement part u(s) of the vector z(s) is contained in the column space spanned
by {b0 , b1 , b2 , b3 , ...}, i,e. by the Krylov subspace K (H 11 , H 12 ; b0 ),

u() u K (K1 D, K1 M; K1 f) (38)

which holds true for low frequency range, because the interpolation point s0 was chosen equal to
zero.
It remains to show how to get the Krylov subspace for s 6= 0, such that the matrix V contains
information of many points, i.e. a interpolatory approach over the frequency interval . It is
achieved by noting that the moments of h(s) about s0 6= 0 are equal to the moments of h(s + s0 )
about zero [5]. One also can proof that the transfer function given by eq. (20) is equal to that
given by eq. (33), thus

h(s + s0 ) = C((s + s0 )2 M + (s + s0 )D + K)1 f

= C(s2 M + s(2s0 M + D) + (s20 M + (s0 D + K))1 f (39)
For convenience one may define

Ds = 2s0 M + D and Ks = s20 M + s0 D + K (40)

Ds and Ks are the generalized damping and the dynamic stiffness respectively. The substitution
of D and K in the Krylov subspace in eq. (38) by the definitions given by eq. (40) leads to

K (K1 1 1
s Ds , Ks M; Ks f) (41)
which approximates the solution u for any s0 . Using s = i and defining P frequencies as inter-
polation points in a vector space V may be defined by

SP
V[N3P] = p=1 K p (2p M + i p D + K)1 (2i p M + D),
 (42)
2 1 2 1
( p M + i p D + K) M; ( p M + i p D + K) f

where p is the interpolation frequency, with p = 1, 2, ..., P and p is the number of matched terms
about p of the Taylor expansion of the transfer function given by eq.(33). If the generalized
damping is neglected eq. (42) simplifies to
P  
K (2p M + i p D + K)1 M, (2p M + i p D + K)1 f
[
V[N3P] = (43)
p=1

which is the suitable form for proportional (Rayleigh damping and modal damping) damped sys-
tems.

7
 Second order approach
The transfer matrix of the second order dynamical system of eq. (19) is given by eq. (20) and
rewritten here for convenience
H (s) = CT (s2 M + sD + K)1 f = CT Ks f (44)
The Taylor expansion of eq. (44) is given by

1 d(k)  
k
H (s0 6= 0) = H (s0 (s)
) (45)
k k! ds(k)

therefore the moments of eq. (45) are

1 d(k)  
mk = H (s0 ) k = 0, 1, 2, 3, ..., (46)
k! ds(k)
By solving the first moments, eq. (46), of the Taylor series given by eq. (45) and considering C
equal to the identity matrix it is possible to group together their terms such that a Krylov subspace
of the form given by (41) is obtained. It is illustrated by solving the first four moments, using eq.
(46), of the Taylor series about s0 , i.e.

m0 = K1
s f = b0 (47)
m1 = K2 1
s (2s0 M + D)f = (Ks Ds )b0 (48)
m2 = K3 2 2
s (2s0 M + D) f Ks (M)f
= (K1 1 1
s Ds )(Ks Ds )b0 + (Ks M)b0 (49)
8 3
m3 = K4 3
s (2s0 M + D) f + Ks (2s0 M + D)(M)f
6
4 3
+ Ks (2s0 M + D)(M)f
6
= (K1 1 1 1 1
s Ds )(Ks Ds )(Ks Ds )b0 + (Ks Ds )(Ks M)b0
+ (K1 1
s M)(Ks Ds )b0 (50)
These show that the moments m0 , m1 , m2 , ..., m of the second order dynamical system given by
eq. (19) span the Krylov subspace defined by eq. (41) and vice versa.

3.3 Derivative-based Galerkin projection Method

The transfer function of the original second order dynamical system of eq. (19) can be redefined
as
H (s) = CT (s2 M + sD + K)1 = CT K1 (51)
fp(s) of eq. (19) is interpreted as a constant input with p(s) = 1. Analogously the transfer function
of a reduced second order system is defined by

H (s) = CT V(s2 Mr + sDr + Kr )1 VH = CT VK1

s, V
H
(52)

One searches for a H which acting in a vector f approximates the original transfer function acting
on the same vector, i.e.

k H (s p + s)f H (s p + s)f k c|s|h (53)

where s is sufficient small and h > 0. An equivalent expression to eq. (53) is given by
H (s)f = H (s)f + O ((s) ) (54)
which implies
dk dk
(H ) = (H ) f or k = 0, 1, 2, 3, ..., 1 (55)
dsk dsk

8
 dk
Thus if dsk
((s2p M + 2s p D + K)1 f) V for k = 0, 1, 2, ..., 1 then one proofs eq. (53) as follows
[4]
kH (s p + s)f H (s p + s)fk = kCK1 T 1 H
s (s p + s)f C VKs, (s p + s)V fk

Using the skew projection P = VK1 H

s, (s + s)V Ks (s + s)

kCT K1 T 1 H
s (s p + s)f C VKs, (s p + s)V fk
= kCT K1 T 1
s (s p + s)f C P Ks (s p + s)fk

1 d (k)  1 
= kCT (I P ) (k)
K s (s p + s) sk fk
k=0 k! ds
(k) 
1 d 1

= kCT (k)
K s (s p + s) sk fk
k= k! ds

1 d (k+)  1 
= |s| kCT (k+)
K s (s p + s) sk fk
k=0 (k + )! ds
c|s| (56)
where c is chosen uniformly for all s sufficient small. For the second order dynamical system
given by eq. (19) the matrix C is considered equal to the identity matrix and one is interested in
the state vector u, therefore taking the vector f of eq. (54) equal to the force vector, and the fact
that s = i, eq. (54) holds true if
( )
P
[ d   d 1  
span u( p ), u( p ) , , 1 u( p ) Ran(V) (57)
p=1
d d

u and its derivatives are found by solving eq. (27) for k = 0, 1, 2, ..., 1.

4 Numerical Example
A box-shaped room with Lx = 4m Ly = 3m Lz = 2m is filled with a fluid (air) with a density of 12
kg 6
m3 and bulk modulus of 1.42 10 . Five sides of that room are rigid walls and the elastic wall
has the following dimensions: lz = 3m, ly = 2m and thickness = 0.01m. It posses the following
properties: elasticity modulus Ex = Ey = Ez = 21Pa, Poisson ratio x = y = z = 0.3 and density
of 7580 mkg3 . The system is excited by a load Fx = 10 at 0.3Ly and 0.5Lz . The computations were
performed on a Intel(R) core(TM) i5-4670 CPU@ 3.10GHz 3.10 GHz.
The first analysis in done using a 2546 dofs model in a frequency band = [1, 100], n = 100,
P = 16, and p ={ 2, 10, 30, 40, 50, 65, 70, 85, 95}. mP uses P = 9 and p ={10, 20, 30, 40, 50,
65, 70, 80, 90}.
The computation time required for this analysis is shown in tab. 1 and the amplitude and phase
angle of the frequency response for the pressure at the point of excitation are depicted in fig. 1
and 2 respectively.

Method c.t.[s] c.t.parallel[s]

DM 10.192
KGP 2.656 1.497
DGP 2.778 1.529
mP 3.660 1.856

Table 1: Computation time using DM, KGP, DGP and mP

The comparison in computation time (c.t.) shows a faster achievement of the solution for the par-
allel implementation. the KGP and DGP are 4 times faster than the DM runs in the conventional

9
way. For the mP approximation, it is recommended to used a mesh with few interpolation points,
e.g. for this example P < 10, due to the fact that the matrix A in eq. (28) is ill-conditioned for large
P [2], which yields to wrong coefficients and therefore wrong approximation of the response.

Frequency Response Plot

3
10
DM
KGP
2 DGP
10 mP
Amplitude [Pa]

1
10

0
10

1
10

2
10
0 10 20 30 40 50 60 70 80 90 100
Frequency [Hz]

Figure 1: Amplitude of the response using DM, KGP, DGP and mP.

Frequency Response Plot

200

150

100
Phase angle [rad]

50

-50

-100 DM
KGP
-150 DGP
mP

-200
0 10 20 30 40 50 60 70 80 90 100
Frequency [Hz]

Figure 2: Phase angle of the response using DM, KGP, DGP and mP.

If one chooses to use the mP interpolation points for DGP and KGP one gets the response depicted
in figs. 3 and 4.

10
 Frequency Response Plot
10 3
DM
KGP
10 2 DGP
mP

Amplitude [Pa]
10 1

10 0

10 -1

10 -2
0 10 20 30 40 50 60 70 80 90 100
Frequency [Hz]

Figure 3: Phase angle of the response using DM, KGP, DGP and mP.

Frequency Response Plot

200

150

100
Phase angle [rad]

50

-50

-100
DM
KGP
-150 DGP
mP
-200
0 10 20 30 40 50 60 70 80 90 100
Frequency [Hz]

Figure 4: Phase angle of the response using DM, KGP, DGP and mP.

In a second investigation = [1, 200], n = 200, P = 18, p ={2, 10, 30, 40, 50, 65, 70, 85, 95,
105, 115,125,135,145,155,170,180,190}.
The computation time required for this analysis is written in tab. 2 showing that the KGP and
DGP (run in parallel) are approximately 7.5 times faster than the DM.

Method c.t.[s] c.t.parallel[s]

DM 21.403
KGP 5.488 2.855
DGP 5.506 2.795

Table 2: Computation time using DM, KGP and DGP.

The amplitude and the phase angle of the frequency response for the pressure at the point of
excitation are depicted in figs. 5, 6.

11
 Frequency Response Plot
10 3
DM
KGP
DGP
10 2

Amplitude [Pa]
10 1

10 0

10 -1

10 -2
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Figure 5: Amplitude of the response using DM, KGP and DGP.

Frequency Response Plot

200

150

100
Phase angle [rad]

50

-50

-100
DM
KGP
-150
DGP

-200
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Figure 6: Phase angle of the response using DM, KGP and DGP.

The last investigation is done in a 15 041 dofs model. The frequency response of the pressure at
the excitation point is depicted in figs. 7 and 8

Method c.t.[s] c.t.parallel[s]

DM 736.631
KGP 196.438 131.337
DGP 202.956 130.859

Table 3: Computation time using DM, KGP and DGP.

12
 Frequency Response Plot
10 2
DM
KGP
DGP

10 1

Amplitude [Pa]

10 0

10 -1

10 -2
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Figure 7: Amplitude of the response using DM, KGP and DGP.

Frequency Response Plot

200

150

100
Phase angle [rad]

50

-50

-100
DM
KGP
-150
DGP

-200
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Figure 8: Phase angle of the response using DM, KGP and DGP.

The solutions given by the KGP and DGP overlap the exact solution almost everywhere in the .
At the end of , the Moment-Matching MOR methods do not capture the exact solution due to
the fact that the last interpolation point considered in the computation was 190Hz.

5 CONCLUSIONS
As shown in this work, the moment matching approach can be used to obtain a reduced-order
model for a vibroacoustical problem of the form given by eq. (3). For non proportional-damped
systems and unsymmetrical matrices the CMA and the Moment-Matching MOR methods are
appropriate choices, but the CMA increases the dimension of the matrices. The mP is applicable
in a narrow or when few interpolation points are needed. The KGP and DGP can be applied
to any kind of system defined by eq. (3) over a narrow or large . Thus they are very promising
methods to obtain a reduced-order model with less computational cost and good approximation
to the exact solution for non-proportional damped vibroacoustical systems.

References

[1] Su, T. J., & Craig, R. R. (1991). Model reduction and control of flexible structures using Krylov
vectors. Journal of guidance, control, and dynamics, 14(2), 260-267.

13
 [2] Bai, Z. (2002). Krylov subspace techniques for reduced-order modeling of large-scale dynamical sys-
tems. Applied Numerical Mathematics, 43(1), 9-44.

[3] Freund, R. W. (2003). Model reduction methods based on Krylov subspaces. Acta Numerica , 12,
267-319.
[4] Beattie, C., & Gugercin, S. (2009). Interpolatory projection methods for structure-preserving model
reduction. Systems & Control Letters, 58(3), 225-232.

[5] Salimbahrami, S. B. (2005). Structure preserving order reduction of large scale second order models
(Doctoral dissertation, Universitat Munchen).
[6] Avery, P., Farhat, C., & Reese, G. (2007). Fast frequency sweep computations using a multipoint
Padbased reconstruction method and an efficient iterative solver. International Journal for Numerical
Methods in Engineering, 69(13), 2848-2875.

[7] Hetmaniuk, U., Tezaur, R., & Farhat, C. (2012). Review and assessment of interpolatory model order
reduction methods for frequency response structural dynamics and acoustics problems. International
Journal for Numerical Methods in Engineering, 90(13), 1636-1662.
[8] Craig, R. R., & Kurdila, A. J. (2006). Fundamentals of structural dynamics. John Wiley & Sons.

[9] Meirovitch, L. (2010). Fundamentals of vibrations. Waveland Press.

[10] OCallahan, J. C. (1989, February). A procedure for an improved reduced system (IRS) model. In
Proceedings of the 7th international modal analysis conference (Vol. 1, pp. 17-21). Union College
Press, Schenectady, NY.

[11] Gordis, J. H. (1992). An analysis of the improved reduced system (IRS) model reduction procedure.
In 10th International Modal Analysis Conference (Vol. 1, pp. 471-479).
[12] Blair, M. A., Camino, T. S., & Dickens, J. M. (1991). An iterative approach to a reduced mass matrix.
In 9th Conference International Modal Analysis Conference (IMAC) (Vol. 1, pp. 621-626).

14