Proceedings of ESDA04

7th Biennial Conference on Engineering Systems Design and Analysis

Proceedings of ESDA04
7TH Biennial ConferenceJuly
on Engineering Systems Design
19-22, 2004, Manchester, Unitedand Analysis
Kingdom
July 19-22, 2004, Manchester, United Kingdom

ESDA2004-58542
ESDA2004-58542

COMBINED DETERMINISTIC-STOCHASTIC FREQUENCY-DOMAIN SUBSPACE

IDENTIFICATION FOR EXPERIMENTAL AND OPERATIONAL MODAL ANALYSIS.

Bart Cauberghe∗, Patrick Guillaume, Peter Verboven, Eli Parloo, Steve Vanlanduit
Department of Mechanical Engineering
Vrije Universiteit Brussel
Brussels, B-1050
Belgium
Email: [email protected]

engineers to get more physical insight from the identified mod-

els. Continuously expanding its application base, modal analysis
Keywords: experimental and operational modal analysis, system
is today successfully applied in automotive engineering (engine,
identification,
suspension, body-in-white, fully trimmed cars, ...), aircraft en-
gineering (ground vibration test, landing gear, control surfaces,
ABSTRACT
in-flight tests), spacecraft engineering (launchers, antennas, solid
Until recently frequency-domain subspace algorithms were
panels, satellites,...), industrial machinery (pumps, compressors,
limited to identify deterministic models from input/ouput mea-
turbines, ...) and civil engineering (bridges, off-shore platforms,
surements. In this paper, a combined deterministic-stochastic
dams, ...).
frequency-domain subspace algorithm is presented to estimate
models from input/output spectra, frequency response functions Experimental modal analysis (EMA) identifies a modal
or power spectra for application as experimental and operational model from the measured forces applied to the test structure and
modal analysis. The relation with time-domain subspace identifi- the measured vibration responses. The modal model expresses
cation is elaborated. It is shown by both simulations and real-life the dynamical behavior of the structure as a linear combination
test examples that the presented method outperforms traditional of different resonant modes. Each resonance mode is defined by
frequency-domain subspace methods. a resonance frequency, damping ratio, mode shape and partici-
pation vector. These modal parameters depend on the geometry,
material properties and boundary conditions of the structure. Vi-
INTRODUCTION brations of the structure originate from its resonance modes that
During the last decade modal analysis has become a key are inherent properties of the structure. Forces exciting the struc-
technology in structural dynamics analysis. Starting from simple ture at one or more of these resonance frequencies cause large
techniques for trouble shooting, it has evolved to a ”standard” vibration responses resulting in possible damage, discomfort and
approach in mechanical product development. Beginning from malfunctioning.
the modal model, design improvements can be predicted and the More recently, system identification techniques were devel-
structure can be optimized. Based on the academic principles oped to identify the modal model from the structure under its op-
of system identification, experimental modal analysis helps the erational conditions from vibration responses only. These tech-
niques, referred as operational modal analysis (OMA) or output-
only modal analysis, take advantage of the ambient excitation as
∗ Address all correspondence to this author. e.g. wind, traffic and turbulence. The latter experimental situ-

1 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 ation can differ significantly from the real-life operating condi- cussed on time-domain subspace methods, while linear systems
tions. An important advantage of OMA is that the structure can often are characterized in the frequency domain. Frequency do-
remain in its normal operating condition resulting in more real- main system identification has several advantages especially for
istic models for in-operational structures. modal testing such as simple frequency-band selection, simple
Nevertheless, the current evolution in mechanical engineer- pre-filtering for elimination of disturbing components, physical
ing towards the use of Computer Aided Design (CAD) like Fi- interpretation, ... . Therefore, it is quite natural to consider sub-
nite Element Models (FEM) results in a changing role for test- space identification algorithms in the frequency domain to iden-
ing [1], [2]. Today the optimization process in product develop- tify models directly from input/output spectra, FRFs or power
ment is under strong pressure because of the competitive market, spectra. In [7] a basic projection frequency-domain subspace al-
increasing customers’ demands and by consequence the design gorithm is proposed in a deterministic framework, which does
cycle becomes shorter in time. This results in an increasing use not require uniform frequency grid and which allows to use a
of simulations based on numerical models to reduce the num- frequency weighting. By considering the covariances on the pri-
ber of prototypes and expensive experiments. Still, testing plays mary data as a frequency weighting, it is shown that this projec-
an important and evermore critical role, in every step of the de- tion algorithm is strongly consistent [8], [9]. An other approach
velopment process for target setting, bench-marking and model consists by transforming the frequency-domain data to the time-
updating. All this, together with the decreasing expertise of the domain by the IDFT and combining this with classical time do-
users, since EMA/OMA have been transferred from the realm of main techniques [8].
the research experts to the product development workfloor [3], In general three different types of subspace identification al-
makes that the demands for modal parameter estimation (MPE) gorithms exist in the time-domain [10]:
algorithms still increase in terms of accuracy, speed, automation
and physical interpretation. Identification of deterministic models given by
In general the proposed algorithm tries to fulfill the require-
ments for new MPE algorithms [1], i.e.
xn+1 = Axn + Bun
1. Reducing the time of testing. yn = Cxn + Dun
2. Allow maximal test data exploitation
3. Increase the accuracy of the estimates
4. Decrease the complexity of the analysis, allowing less- Deterministic algorithms only estimate the dynamics be-
experienced staff to process the data tween the measured outputs yn and the measured inputs un .
5. Extend the limits and ranges (large number of sensors, high Their frequency-domain equivalents, discussed in [8], can
damping, noisy data, short data records, ...) be made consistent in the case of colored output noise by

introducing the covariances of output noise Ck = E Nk NkH
Furthermore, in [4], [5], [6] an extension of the EMA and OMA
as a frequency weighting in the projection algorithm
concepts, to a so-called combined EMA-OMA framework is pro-
posed. In this combined framework, the vibration response is
considered as a result of both measured artificial applied forces zk Xk = AXk + BUk
and unmeasurable ambient excitation, i.e. an Operational Modal Yk = CXk + DUk + Nk (1)
Analysis with eXogenous inputs (OMAX), resulting in a maxi-
mum data exploitation. In the OMAX framework, the stochastic
contribution, i.e. the part of the response which can not be re- with Nk zero mean complex circular independent noise. Re-
lated to the measurable input forces, is considered as valuable mind that, although this algorithm is consistent, dynam-
information. Under the assumption that the stochastic contribu- ics excited by ambient forces are not identified, since no
tion in the response is related to unmeasurable ambient forces, information is extracted from Nk . Therefore, this algo-
extra information about the system can be extracted from this rithm has no combined deterministic-stochastic meaning
contribution i.e. both modes excited by the measurable and/or by in the OMAX framework. Extensions of the determinis-
the unmeasurable forces can be identified from the data. tic frequency-domain state-space formulation to take into
Frequency-domain subspace algorithms estimate state-space account intial/final conditions and a validation procedure
models by means of geometrical projections and can be con- based on this extension are proposed in [11], [12].
sidered as the frequency-domain counterparts of time-domain Identification of stochastic models [13] given by
subspace algorithms. Compared to the traditional identification
algorithms, subspace algorithms are non-iterative and as a re-
sult they always yield a solution without the risk of convergence xn+1 = Axn + wn
problems. Until recently, research efforts were and still are fo- yn = Cxn + vn

2 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Stochastic models estimate the dynamics of a system with- Assumption 1: Wk and Vk are zero mean circular complex
out the use of artificially applied forces un , but instead con- independent and identically distributed (over k) noise sources
siders all unmeasured forces to be white zero mean (random) with covariance matrix
noise wn and vn . These type of models have their applica-
tions in the OMA framework and their frequency-domain
    
Wk Q S
WkH VkH


equivalent is presented in [14], [4]. E = (4)
Vk SH R
Identification of combined deterministic-stochastic models
given by [15]
where E is the expected value.
The goal of this chapter is to estimate the unknown matrices
xn+1 = Axn + Bun + wn A, B,C, D, Q, R, S (up to a similarity transformation) from the
yn = Cxn + Dun + vn given spectra Yk and Uk and assuming the data is generated by
an unknown combined deterministic-stochastic system.
white wn and vn white zero mean noise sources. Under
the assumption that the measured input forces un are uncor- 1.2 Main theorem
related with the unmeasurable inputs wn and vn combined Consider the input and output Vandermonde matrices U ∈
stochastic-deterministic subspace algorithms are consistent. C 2rNi ×N and Y ∈ C 2rNo ×N respectively given by
The combined algorithms identify the dynamics excited by
the measured inputs, the dynamics excited by unmeasurable z−r z−r . . . z−r
1 U1 2 U2 N UN
 
forces as well as the coupled dynamics i.e. modes excited −r+1 −r+1 −r+1
 z1 U1 z2 U2 . . . zN UN 
by both the measurable and unmeasurable forces. This com- 

.. .. .. ..


bined interpretation of this subset of subspace algorithms  −1.
 . . . 
 def 
z−1 −1

fits in the framework of Input/Output Spectra-data driven  z U1
1 2 U2 . . . z N UN 
 U−
U= = (5)
OMAX identification. In this paper a frequency-domain  U1
 U2 . . . UN   U+
counterpart for these combined stochastic-deterministic al-  z1U1
 z2U2 . . . zN UN  
gorithms is developed and its applicability is shown by sev-  .. .. .. .. 
eral examples and simulations. This combined algorithm
 . . . . 
can also be used to obtain consistent estimates from noisy zr−1
1 U1
r−1
z2 U2 . . . zN UN r−1

FRFs and power spectra. Profound studies about frequency-

domain system identification for modal analysis can be and similarly for
found in [4], [16], [17]
def  
Y−
Y= = (6)
1 Theoretical Aspects Y+
1.1 System description
Consider a proper, stable nth order multiple-input/multiple- with r > n/No . Both the input and output Vandermonde matri-
output combined deterministic-stochastic system. The term com- ces are divided in a ’positive’ U+ , Y+ and a ’negative’ U− , Y−
bined means in the context of this paper that the deterministic, the Vandermonde matrices as defined above. The main theorem of
stochastic and the coupled deterministic-stochastic dynamics are this paper is then given by
considered simultaneously. The frequency-domain state-space Theorem 1 Under Assumption 1
equations of a combined discrete-time system is given by
 
U−
zk Xk = AXk + BUk +Wk (2) Y+ /U+ → Or X for N → ∞ (7)
Y−
Yk = CXk + DUk +Vk (3)
with Or the extended observability matrix given by
with Yk ∈ C No ×1 and Uk ∈ C Ni ×1 respectively the vectors of
the output and input spectra at spectral line k, Xk ∈ C n×1 the  
C
state vector at frequency spectral lines k and zk = ei2πk/N (k =
CA
1, 2, . . . , N) covering the full unit circle. The vectors Wk ∈ C n×1
 
Or =  (8)
 
..
and Vk ∈ C No ×1 contain respectively the process and the output

 . 
measurement noise at spectral line k. CAr−1

3 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 and X = [ X1 X2 . . . XN ] the spectra of the forward Kalman filter with x = [x0 . . . xN ]T the forward Kalman filter state sequence
state estimates. The oblique projection A/BC is defined by [10] estimate. Consider the Discrete Fourier Transformation matrix

z01 z02 ... z0N

 
 CCH CBH †
 
CH BH z−1 z−1 z−1

A/BC = A C (9) 1 
BCH BBH [:,1:r] 1 2 N

F= √  .. .. (16)
 
.. 
N . . . 
−(N−1) −(N−1) −(N−1)
and the geometrical interpretation of this oblique projection is z1 z2 . . . zN
given in [10].
From the definition of the DFT and IDFT it follows that
Proof. Consider the Inverse Discrete Fourier Transform (IDFT)
of both the output and input spectra given by
Y− = yp F
Y+ = yf F
1 N
y(n) = IDFT (Yk ) = √ ∑ Yk znk (10) U− = up F
N k=1
U+ = uf F (17)
1 N
u(n) = IDFT (Uk ) = √ ∑ Uk znk (11)
N k=1 and F is orthogonal, i.e. FFH = I. Applying the property
A/BC = D ⇔ [AF]/[BF] [CF] = DF if FF H = I, which can be
From the definition of the IDFT it follows that if y(n) = DFT (Y ) proven by the definition of the oblique projection, for the time
then y(n + m) = DFT (zmY ). Taking the IDFT from the basic domain theorem results in
model equations (2) and (3) results in the following time-domain    
state space model U− [up F]
Y+ /U+ = [yf F]/[uf F] (18)
Y− [yp F]
 
u
x(n + 1) = Ax(n) + Bu(n) + w(n) (12) = yf /uf p F (19)
yp
y(n) = Cx(n) + Du(n) + v(n) (13)
= Ôr x̂F (20)
and w(n) = √1N ∑Nk=1 Wk znk , v(n) = √1N ∑Nk=1 Vk znk zero mean, = Ôr X̂ (21)
white noise vector sequences. The input and output block Hankel
matrices are then given by which ends the proof.
The main theorem Eq. 15 in time-domain subspace is proven to
be consistent, when at least one of the following conditions is
 
u−r u−r+1 ... uN−r
 .. .. .. ..  satisfied [15]:
 . . . . 
The system is purely deterministic, i.e. wn = 0 and vn = 0
 
 u−2 u−1 ... uN−2  def  
The input signal is white noise.

 u−1 u0 ... uN−2  = up
 (14) r → ∞, since for r → ∞ the non-steady state Kalman filter
 u0 u1 ... uN−1  uf
converges to a steady state Kalman filter. This is intuitively
 
 u1 u2 ... uN 
clear, since by the time the Kalman filter is in steady state,
 
 . .. .. .. 
 .. . . .  the effect of the initial conditions has died out. In practice
ur−1 ur . . . uN+r−2 r ≥ 8 already results in good estimates for mechanical sys-
tems [15].
 
yp
and similarly for .
yf 1.3 From states to system matrices
The main theorem for combined deterministic-stochastic sub- Once the states Xk are estimated by the oblique projection
space identification in the time-domain, presented in [10], states according the main theorem Eq. 7 the system matrices A, B, C
that and D can be estimated in a LS sense

ρ
        
up zk Xk AB Xk
yf /uf → Or x for N → ∞ (15) = + Wk (22)
yp Yk C D Uk ρVk

4 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Since it is assumed that no input noise is present and the state 1. Form the Vandermonde matrices Y+ , Y− , U+ and U− from
spectra Xk are consistent estimates, the least squares estimate of the measured spectra Yk and Uk . The inputs can be extended
the system matrices is consistent. The covariance matrix of the by zk to model the transients in the measurements. In case
stochastic sources Wk and Vk is then obtained from the residuals one wants to estimate real system matrices A, B, C and D, the
ρWk and ρWk as Vandermonde matrices V are replaced by [ Re(V) Im(V) ]
(V = Y+ , Y− , U+ and U− ).
2. Calculate the RQ decomposition as
1 N ρWk
    
Q S
∑ ρVk ρ H ρH


= (23)
SH R N k=1 Wk Vk     
Y+ R11 0 0 0 Q1
 U+   R21 R22 0 0   Q2 
 U−  =  R31 (28)
    
1.4 Taking into account effects of transients and leak- R32 R33 0   Q3 
age Y− R41 R42 R43 R44 Q4
Until now it was assumed that the given spectra Yk and
Uk obey the frequency-domain state-space model. Since the then
spectra of the outputs and inputs are obtained from the DFT
of the measured time signals Yk = √1N ∑n=0
N−1
y(n)z−n
k and Uk =

U−
 
R31 R32 R33 0

u(n)z−n = Ra Q with Ra = (29)
√1
N
∑n=0
N−1
k the proposed model is only true if x(N) = x(0) Y− R41 R42 R43 R44
or N = ∞. In all other cases an extra term T must be taken 
U+ = Rb Q with Rb = R21 R22 0 0

(30)
into account in the state-space model to model the intial/final  
Y+ = Rc Q with Rc = R11 0 0 0 (31)
conditions. The extended combined deterministic-stochastic
frequency-domain state-space model is then given by (32)

zk Xk = AXk + BUk + T zk +Wk (24) with QT = [ QT1 QT2 QT3 QT4 ]. The oblique projection
U−
Yk = CXk + DUk +Vk (25) Y+ /U+ is then given by
Y−
x(N)−x(0)
where T = √ models the influence of the initial and final 
U−
 h
H †
ih h
H †
ii†
N = Rc I − RH Rb Ra I − RH
   
Y+ /U+ b Rb Rb b Rb Rb Rb Ra Q (33)
Y−
conditions. These extra parameters T (T ∈ C n×1 ) model the non-
steady state response of the system. Taking into account these ad-
ditional parameters makes the frequency model robust for leak- A profound discussion of this expression for the oblique
age and transients effects in case a rectangular window (uniform projection is given in [10].
window) is used for the calculation of the spectra. This ob- 3. In the third step the SVD of the projection is calculated as
servation generalizes the stochastic frequency-domain subspace  
identification method proposed in chapter [14], since the influ- U−
Y+ /U+ = USV H (34)
ence of leakage and transients can be modelled by a combined Y−
deterministic-stochastic model with input Uk = √1N zk
4. Determine the extended observability matrix Or and the
zk Xk = AXk + BUk +Wk (26) state estimate X̂. For a chosen order n (in theory the order n
is equal to the rank of S)
Yk = CXk + DUk +Vk (27)
1/2
Or = U1 S1 (35)
and B represents the transient term T . As a conclusion the
1/2
presented algorithm in this chapter allows to identify combined X̂ = S1 V1H (36)
deterministic-stochastic models from IO data and purely stochas-
tic models from output-only spectra, without suffering from tran- with U1 = U[:,1:n] , S1 = S[1:n,1:n] and V1 = V[1:n,:] . In the case
sient and leakage errors. 1/2
that real system matrices are estimated X̂ ′ = S1 V1H and X̂ =
′
X̂[:,1:N] ′
+ iX̂[:,N+1:2N] .
2 Practical Implementation 5. Knowing Yk ,Uk and X̂k the system matrices are obtained by
In this section the implementation of the different algorithm solving the LS problem 22 and the covariance matrix of the
steps is discussed: process and noise is obtained by 23

5 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 3 Remarks 4 Producing clear stabilization charts for modal anal-
1. Equation (34) can be replaced by ysis experiments
Processing MIMO measurements is common practice to es-
timate the modal parameters i.e. natural frequencies, damping
ratios, mode shapes and participation factors from force and vi-
  bration response measurements. A typical problem is the re-
U− quired user interaction to distinguish physical poles from the
W1 Y+ /U+ W2 = USV H (37)
Y− mathematical poles. It is known from stochastic frequency-
domain subspace system identification that always stable poles
are identified in a consistent way [14]. By changing the ba-
sis function zk into z−1k in the Vandermonde matrices the poles
of the stochastic system are identified as unstable. For a com-
with W1 ∈ C No r×No r and W2 ∈ C N×N such that W1 is of bined deterministic-stochastic approach this results in stable
full rank and rank([ U− Y− ]T ) = rank([ U− Y− ]T W2 ). For poles modelling the deterministic contribution and unstable poles
time-domain combined deterministic stochastic subspace it modelling the stochastic contribution. This allows to distinguish
is shown in [10] that special choices of W1 and W2 corre- the physical poles from the stochastic poles (which model the
spond to well-known algorithms like the N4SID (Numerical noise i.e. the stochastic contribution on the data) by inspecting
algorithms for Subspace State Space System Identification), the sign of the damping value. This idea can be applied for sev-
MOESP (Multivariable Output-Error State Space) and the eral system identification techniques to result in clear stabiliza-
Canonical variate algorithm (CVA). Analogously the fre- tion charts [18], [4], [19]. A stabilization diagram is a typical tool
quency counterpart of this weighing matrices correspond to used in modal analysis which illustrates the estimated poles for
the frequency counterparts of the N4SID, MOESP and CVA different model orders. However, it should be mentioned that in
algorithms. the case of coupled deterministic-stochastic dynamics (as in the
2. The presented algorithm can also be used as a com- OMAX case) the identified poles will be biased i.e. the damping
bined deterministic-stochastic frequency-domain subspace ratios are underestimated, since the stochastic contribution tends
method processing Frequency Response Functions (FRFs) to estimate unstable poles, while the deterministic contribution
or positive power spectra. By simple substitution of Yk = Hk tends to stable poles.
or Yk = G+ yy,k and Uk = Ik with Hk ∈ C
No ×Ni
and G+ yy,k ∈
No ×Nre f
C respectively the FRF matrix and power spectra ma-
trix at spectral line k and Ik a unity matrix with dimensions 5 Simulations and Real-life measurement examples
Ni . Since the inputs Uk are considered white in this case, the 5.1 Monte Carlo Simulation
projection results in a consistent estimate of the state spectra Lets consider a discrete 6th order system excited by 2 Gaus-
and the observability matrix. Compared to the consistent de- sian random sequences. The A, B, C and D matrices are given
terministic projection algorithm [8], the advantage is that no in appendix 6. The first input is a standard gaussian distributed
a-priori known covariance matrix of the noise has be taken sequence with standard deviation 1, while the second input has a
into account. The need for the covariance matrix of the noise standard deviation of 0.5. All simulations are performed in the
on the primary data to be consistent is a serious drawback, time-domain to include the non-steady state responses. During
since this matrix must be convertible and thus at least No av- the identification process the second input U2,k is considered as
erages must be taken, resulting in a long measurement time unknown and by consequence the identification process can be
for a practical test setup (e.g. No > 50). considered as a combined deterministic-stochastic problem. The
3. The same motivations to perform the identification in the fre- response in the frequency-domain can be considered as a sum of
quency domain as demonstrated for the stochastic subspace three contributions i.e. the deterministic part, the transient part
identification are valid for the combined identification: i.e. and the stochastic part
easy pre-filtering of the data and simple frequency band se-
lection. The chosen frequency band is re-scaled to cover
Yk = C(Izk − A)−1 B[:,1] + D[:,1] U1,k +C(Izk − A)−1 T zk
 
uniformly the full unit circle (half unit circle in case of real
system matrices) to ensure consistency and the numerical + C(Izk − A)−1Wk +Vk (38)
stability is improved, since the basis functions zr are orthog-
onal. In practice, it is advised to cover only 90% to make with in this simulation Wk = B[:,2]U2,k and Vk = D[:,2]U2,k . To
a trade-off between consistency and model errors (since the examine the consistency properties (asymptotic unbiasedness),
frequency band of the underlying continuous-time domain Monte Carlo simulations are performed for 4 different cases of
model is approximated by a discrete-time domain model). the total number of time samples Nt (N = Nt /2) in order to check

6 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 −20 −20

−40
−40
Table 1. MSRE for Monte Carlo simulations comparing the presented
−60
−60
−80 combined algorithm including transient effects (A1), combined algorithm
Ampltitude

Ampltitude
−80
−100 not including transient effect (A2) and the classical frequency domain sub-
−100
−120
−120
space projection algorithm (A3). From the simulation results it is clear that
−140

−140
the MSRE of A1 and A2 decreases if the number of frequency lines N in-
−160

−180 −160 creases, while A3 remains biased

0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
frequency (Hz) frequency (Hz)

(a) (b) N=150 N=250 N=400 N=600

A1 0.0097 0.0054 0.0042 0.0027
Figure 1. Mean and variance of the estimated system for N=250 (◦ exact A2 0.0373 0.0074 0.0066 0.0041
¯
system, full line Ĝ, dotted line σ2 : a) Combined deterministic-stochastic A3 0.2810 0.2132 0.2760 0.3404
Ĝ
frequency domain subspace algorithm (A1) b) Classical frequency-
domain projection subspace algorithm (A3). One can clearly notice the
bias of algorithm A3. output can be considered as the sum of transient effects and a
stochastic contribution as shown in Eq. 40.

Yk = C(Izk − A)−1 T zk +C(Izk − A)−1Wk +Vk (40)

if the expectance value of the estimated model is equal to the
true model for the number of frequency lines N going to infinity.
As an error measure the mean square relative error (MSRE) over The presented algorithm can still estimate the system matrices C
the different frequencies between the mean estimated model (for and A. This is impossible by the classical frequency-domain pro-
the 100 runs and 100 estimated models) and the true model is jection subspace algorithm, since it models only a deterministic
calculated model between the input and output spectra. From the 100 simu-
lations the mean value and the standard deviation of the estimates
of the magnitude R and angle θ of the second pole p2 = Reiθ with
1 N Ĝ¯ k − Gk 2 R = 0.9858 and θ = 0.5938 is shown in Table 2. (similar for the
|E|2 = ∑ | Gk |
N k=1
(39)
other poles). The simulations show clearly that the differences

with Gk the ’true’ system between the output and the first input
and Ĝ¯ k the mean of the 100 estimated models. This MSRE is Table 2. Monte Carlo simulations for a stochastic identification for differ-
calculated for N = 150, 250, 400 and 600 given in table 1. The ent numbers of spectral lines N
presented subspace model is compared with the classical pro-
jection frequency domain projection algorithm. The presented
N=300 N=600 N=900
algorithm was applied including the estimation of the transient
R̄ 0.9826 0.9840 0.9851
T by considering an conditional input zk and without estimating
σR 0.0061 0.0044 0.0034
the transient effects. The identification starts from the spectra of
θ̄ 1.5950 1.5944 1.5939
the response Yk and the first input U1,k . All spectra are calcu-
σθ 0.0070 0.0043 0.0039
lated using a rectangular window. From table 1 it is clear that
the proposed combined algorithm (A1,A2) outperforms the clas-
sical projection algorithm (A3). Furthermore, the error is further
reduced by taking into account the transients effect T to com- between the mean values of the 100 times estimated magnitude
pensate for leakage and the non-steady state behavior (A1 versus and phase of the pole p2 and the exact value decreases for an
A2). This is more important for small Nt , since for large Nt both increasing number of spectral lines (asymptotic unbiased). The
leakage and the transient effect have less influence. In figure 1, variances on the estimated R and θ decrease for an increasing
number of spectral lines (consistency). Figure 2 shows the out-
the mean estimated system Ĝ¯ and the variance σ2Ĝ over the 100
put spectra and the estimated stochastic model from this spectra
simulations are shown for the different identification cases for
for 1 simulation.
N = 250. Assume now that, both inputs are unknown and thus
only the spectra of the outputs are known. In this case the iden-
tification procedure becomes a stochastic identification problem. 5.2 Flight flutter measurements
The influence of the final and initial conditions can still be mod- To show the applicability of the combined deterministic-
elled by considering the parameters T and the input zk . The stochastic frequency domain the same real-life in-flight vibra-

7 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 40
(deterministic) projection algorithm and the combined determin-
20
istic stochastic approaches. Based on the relative differences
0
between the eigenfrequencies and damping ratios estimated for
amplitude (dB)

−20
subsequent model orders n, the poles are labelled in the stabiliza-
−40
tion diagram. The symbols s, f , d and o respectively mean
−60

−80 s: both relative damping ratio difference < 15% and relative
−100 eigenfrequency difference < 3%
−120
0 0.1 0.2 0.3 0.4 0.5 f : relative damping ratio difference ≥ 15% and relative
frequency (Hz)
eigenfrequency difference < 3%
Figure 2. Output spectra (dots), estimated stochastic model (full line) d: relative damping ratio difference > 15% and relative
eigenfrequency difference ≥ 3%
o: relative damping ratio difference ≥ 15% and relative
14 14 eigenfrequency difference ≥ 3%
12 12

10 10 After applying these criteria it is clear that the stabilization di-

model order

model order

8 8 agram of the combined algorithm results in a easier interpreta-

6 6
tion of the chart than the deterministic algorithm. Table 3 clearly
4 4
shows that some of the natural frequencies and damping ratios
2 2
are not identified by the deterministic approach. Furthermore,
4 5 6 7 8 9 10 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz) large deviations in the estimated damping ratios can be observed
(a) (b) between the deterministic and combined identified parameters.
From the aircraft manufacturer it was confirmed that all the 9
f o f f f o f o f o f f f f s s ss s f sd fs s s
14 s
s
o s
f
sf
fo f
f
f
o o
o s
sf s
s f
f f
f s
o 14 s
s
f
s
sf s
ss
s
s o
o
d
s
s
fs
s
s
s
s
s
poles identified with the combined are the only 9 poles in the
12 s
f s
f f f
s f
o
so
o f
s
f
s
s
s
of
f f
12 o s
s
s
s
ss
ss
s
s
fs
s
s
s
s
s
s
s
frequency band and the damping ratios were in very close agree-
10 s f ff f f f f f 10 s f ss s f f s d
ment with the manufacturers specification. From figure 4, which
model order
model order

f f ff f s s f o s s ss o s s s o
8 f o sf o s f s 8 f o fs f f s

6
s
f
f fs
ss o
f s
f
s
f 6
s o
f
fs
sf o
o s
f
s
f
compares the synthesized FRFs for both the deterministic and
4
o sf
f f o
o
f
f
4
d sf
f f
o
s
s
combined algorithm with the FRFs obtained from the H1 esti-
f f s ff s
2 2 mator, it is clear that the deterministic algorithm results in much
4 5 6 7 8 9 10 4 5 6 7 8 9 10 larger model errors than the combined algorithm. By reason of
Freq. (Hz) Freq. (Hz)
confidential agreements the correct sample period is not used and
(c) (d)
thus all damping and frequency values do not agree with the ac-
tual true values.
Figure 3. Comparison between stabilization charts obtained by the de-
terministic and combined subspace algorithm: (a) deterministic algorithm
(b) combined algorithm (c) deterministic algorithm after applying relative 5.3 Averaged Based Spectral function driven identifi-
criteria (d) combined algorithm after applying relative criteria. cation
The combined deterministic-stochastic algorithm can also
start from FRFs (EMA), power spectra (OMA) or both simulta-
neously (OMAX) as primary data to estimate the system matri-
ces and modal parameters. The next two examples show that the
tion measurements are now analyzed. The measured signal of
algorithm is capable to start from FRF data from modal testing
the angle perturbation of the flaps is used as the input force
in the laboratory ,which illustrates its capability to handle with a
signal, since the system is excited by injection a noise signal
large number of outputs, several inputs, frequency band selection
in the flight-by-wire control system. The stochastic contribu-
and large model orders.
tion in the responses is caused by the turbulent forces acting
on the airplane during flight. In fact this measurement setup is
an OMAX situation, since both a measurable artificial applied 5.3.1 Subframe of an engine The accelerations were
force and unmeasurable turbulent forces are acting on the air- measured at 23 response locations and two random inputs were
plane. Starting from the input and output spectra in a frequency applied by electrodynamic shakers. Before the actual identifi-
band from 3Hz to 11Hz, a state-space model with real system cation task started, the frequency band of interest between 210
matrices is estimated by both the classical projection and com- Hz and 410 Hz is re-scaled to cover 90% of the full unit cir-
bined deterministic-stochastic frequency-domain subspace algo- cle (complex model). A model order of 30 was used. Both the
rithms. Figure 3 compares the stabilization charts for both the models identified by the classical projection algorithm and the

8 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 25
25
20
20

15
15 Table 3. Natural frequencies and damping ratios identified by the deter-
10
10 ministic and combined deterministic-stochastic frequency-domain algo-
Ampl. (dB)

Ampl. (dB)
5
5

0
0
rithm.
−5
−5 −10

−10 −15

−15 −20
−25
fdet (Hz) ddet (%) fcom (Hz) dcom (%)
−20
4 6 8 10 4 5 6 7 8 9 10
Freq. (Hz) Freq. (Hz)
4.08 1.95 4.09 3.06
−5
−10
−5 4.60 4.07 4.69 3.39
−10
−15 −15
−20 −20 5.27 3.46 5.16 5.59
Ampl. (dB)

Ampl. (dB)

−25 −25
−30
−35
−30
−35
/ / 5.35 2.39
−40
−40
−45
−50
−45 6.06 0.17 6.01 5.19
−50
−55
−55
4 6
Freq. (Hz)
8 10 4 5 6 7
Freq. (Hz)
8 9 10 11 8.07 2.95 8.07 3.27

15 15
8.80 2.20 8.71 3.25
10 10
5 5 9.07 2.91 9.17 3.60
0 0
Ampl. (dB)

Ampl. (dB)

−5

−10
−5 / / 9.87 3.81
−10
−15
−15
−20
−20
−25
−25
−30
4 5 6 7 8 9 10 11
−30
4 5 6 7 8 9 10 11 tion approaches. The presented combined approach resulted in
Freq. (Hz) Freq. (Hz)
an average MSRE over all FRFs of 0.2, while the classical pro-
jection based subspace method resulted in a mean MSRE of 5.4.
Figure 4. Comparison between the FRFs (×) and the synthesized FRFs Figure 6 compares some identified FRFs for both identification
obtained by the classic projection algorithm (left) and the combined approaches with a high quality validation data set. Figure 7 com-
deterministic-stochastic frequency domain subspace algorithm (right). pares the stabilization diagram for identification for both basis
The classical projection approach results in large bias errors. (times: functions zk or z−1
k . It is clear that starting from the basis func-
measured FRFs, full line: synthesized FRFs)
tion z−1
k facilitates the pole selection.

5.3.2 Laser vibrometer measurements on a car

combined algorithm, are validated with a high quality validation door To illustrate that this combined frequency-domain sub-
data set by means of the MSRE over all frequencies per FRF. space algorithm is capable to process FRF measurements with
Figure 5 shows the MSRE for the 46 FRFs for both identifica- high modal and high spatial density measurements were per-
formed by a scanning laser vibrometer on the door of a car. The
FRF matrix has 79 responses, 1 input and a frequency resolu-
10
2
tion of 0.31Hz. A frequency band from 46Hz up to 120Hz is
1
modelled by a real state-space model containing 50 modes, with
10
the frequencies covering 90% of the half unit circle. Figure 8
10
0 shows the high quality of the synthesized FRFs for 4 different
MSRE

FRF measurements.
−1
10

−2
10

6 Conclusion
10 20
FRF
30 40
This paper presented a combined deterministic-stochastic
frequency domain subspace algorithm resulting in consistent es-
Figure 5. Comparison between the mean square relative errors for the timates of the system matrices of a state-space model. The pro-
46 FRFs. (◦ classical projection frequency-domain subspace method, posed algorithm can be used to process output-only measure-
∗ combined deterministic-stochastic frequency domain subspace algo- ments with a transient term (OMA), input-output measurements
rithm) (OMAX), frequency response functions (EMA) and power spec-

9 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 20 20 able and unmeasurable forces. The applicability is illustrated
10 10
by means of both simulations and several real-life measure-
0 0
ments. The method identifies simultaneously the deterministic
Ampl. (dB)

Ampl. (dB)
−10 −10

−20 −20 dynamics, the stochastic dynamics and the coupled stochastic-
−30
−30
−40
deterministic dynamics an therefore outperforms classical deter-
−40

−50
−50 ministic subspace algorithms. Furthermore, the algorithm can
200 250 300
Freq. (Hz)
350 400 200 250 300
Freq. (Hz)
350 400 take advantage of the frequency-domain by a simple frequency
band selection. The proposed combined stochastic-deterministic
30 30 frequency-domain subspace algorithm is very promising to pro-
20 20
cess all types of frequency-domain data (Input/Output spectra,
10 10
FRF data, power densities). Compared to other frequency-
Ampl. (dB)

Ampl. (dB)

0 0
domain algorithms, the proposed algorithm needs no a priori
−10 −10

−20 −20
noise information and no optimization procedure is required to
−30 −30
be consistent.
200 250 300 350 400 200 250 300 350 400
Freq. (Hz) Freq. (Hz)

20 20 ACKNOWLEDGMENT
10 10

0 0
The financial support of the Fund for Scientific Research
−10 −10 (FWO Vlaanderen); the Concerted Research Action ”OPTI-
Ampl. (dB)

Ampl. (dB)

−20 −20
Mech” of the Flemish Community; the Research Council (OZR)
−30 −30

−40 −40 of the Vrije Universiteit Brussel (VUB); IUAP V/22;the Con-
−50 −50
certed Research Action ”ILiNOS” of the Flemish Community
−60 −60

−70
are gratefully acknowledged. The author holds an Research As-
200 250 300 350 400 200 250 300 350 400
Freq. (Hz) Freq. (Hz)
sistant Grant of the Fund for Scientific Research - Flanders (Bel-
gium)(F.W.O. - Vlaanderen).
Figure 6. Comparison between validation FRFs (×) and the synthe-
sized FRFs by the classical projection algorithm (left) and the combined
deterministic-stochastic frequency domain subspace algorithm (right). REFERENCES
The classical projection approach results in large bias errors. (×: mea- [1] Van der Auweraer, H., 2002. “Testing in the age of virtual
sured FRFs, full line: synthesized FRFs)

−75
−85
−80
−90
−85
30 30 −95
−90
−100
Ampl. (dB)

Ampl. (dB)

−95
25 25 −105
−100
−110
−105
number of modes

number of modes

20 20 −115
−110
−120
−115
15 15 −125 −120
−130 −125
10 10
50 60 70 80 90 100 110 50 60 70 80 90 100 110
Freq. (Hz) Freq. (Hz)
5 5
−70
−80
240 260 280 300 320 340 360 380 250 300 350 −80
frequency (Hz) frequency (Hz)
−90
−90
(a) (b) −100
Ampl. (dB)

Ampl. (dB)

−100
−110
−110

−120
−120
Figure 7. Comparison between stabilization diagrams for both basis
−130
−1 −1 −130
functions: (a) basis function zk ; (b) basis function zk (∗ stable poles −140 −140

i.e. positive damping). 50 60 70 80 90 100 110 50 60 70 80 90 100 110

Freq. (Hz) Freq. (Hz)

Figure 8. Comparison between FRFs (×) and the synthesized FRFs by

the combined deterministic-stochastic frequency domain subspace algo-
tra (OMA). In the case of input-output data, the algorithm can
rithm. (×: measured FRFs, full line: synthesized FRFs)
be considered in the OMAX framework, where the vibration re-
sponse is considered as the sum of a contribution of measur-

10 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 protoyping.”. In Proceedings of the International Confer- quency domain subspace algorithms for the stochastic iden-
ence on Structural Dynamics Modelling. tification problem”. Under review in, Automatica .
[2] Van der Auweraer, H., 2003. “System identification for [15] Van Overschee, P., and De Moor, B., 1994. “N4sid:
structural dynamics and vibra-acoustics design engineer- Subspace algorithms for the identification of combined
ing.”. In Keynote, Proceedings of the 13th IFAC Sympo- deterministic-stochastic systems”. Automatica, 30 (1) ,
sium on System Identification. pp. 75–93.
[3] Peeters, B., Guillaume, P., Cauberghe, B., Verboven, [16] Verboven, P., 2002. Frequency-domain System Identifica-
P., and der Auweraer, H. V., 2004. “Automotive and tion for Modal Analysis. PhD thesis, Dept. of Mechanical
aerospace applications of a new fast-stabilizing polyrefer- Engineering, Vrije Universiteit Brussel, Belgium.
ence frequency-domain parameter estimation method”. In [17] Parloo, E., 2003. Application of Frequency-domain System
Proceedings of the 21th International Modal Analysis Con- Identification in the Field of Operational Modal Analysis.
ference. Ph.D. Thesis, Vrije Universiteit Brussel, Departement of
[4] Cauberghe, B., 2004. Applied Frequency-Doamin System Mechanical Engineering.
Identification in the Field of Experimental and Operational [18] Cauberghe, B., Guillaume, P., Verboven, P., Vanlanduit, S.,
Modal Analysis. Ph.D. Thesis, Vrije Universiteit Brussel, and Parloo, E. “The secret behind clear stabilization di-
Departement of Mechanical Engineering, PDF download- agrams: The influence of the parameter constraint on the
able www.avrg.ac.be/member info cauberghe.htm. stability of poles”. Under review, Mechanical Systems and
[5] Cauberghe, B., Guillaume, P., Verboven, P., and Parloo, E., Signal Processing (MSSP).
2003. “Modal analysis in presence of unmeasured forces [19] Cauberghe, B., Verboven, P., S.Vanlanduit, Guillaume, P.,
and transient effects”. Journal of Sound and Vibration and Parloo, E., 2004. “The secret behind clear stabiliza-
(JSV), 265 , pp. p.609–625. tion charts in experimental modal analysis”. In Proceedings
[6] Cauberghe, B., Guillaume, P., Verboven, P., Parloo, E., of the 10th SEM International Conference on Experimental
and Vanlanduit, S., 2003. “A combined experimental- and Apllied Mechanics.
operational modal analysis approach in the frequency do-
main”. Proceedings of the 21th International Modal Anal-
ysis Conference (IMAC21) . Appendix
[7] Liu, K., Jacques, R., and D.W.Miller, 1994. “Frequency do- The 6th order model used for the simulation is given by
mains structural systm identification by observability range
space extraction”. Proceeding of the American Control  
−0.0103 0.9928 0.0377 −0.1746 −0.0367 0.0365
Conference, Baltimore, 1 , pp. 107–111.  −0.9409 −0.0070 0.2141 0.2065 0.1868 −0.0467 
[8] McKelvey, T., Akcay, H., and Ljung, M., 1996. “Subspace- 
 0.0076 −0.1240 0.6504 −0.8057 −0.1547 0.0538 

based multivariable system identification from frequency A= 
 0.1601 0.0689 0.6680 0.6289 −0.1739 −0.0240 

response data”. IEEE Transactions on Automatic Control,  −0.0136 −0.0056 −0.0103 0.0168 −0.2689 0.7985 
41 (7) , pp. 960–978.
0.0003 −0.0024 −0.0315 −0.0193 −0.9110 −0.6742
[9] Pintelon, R., 2002. “Frequency-domain subspace system  
identification using non-parametric noise models”. Auto- C = −0.3971 −0.0433 −0.3410 −0.0192 −0.3572 −0.0234
 
matica, 38 , pp. 1295–1311. −1.2561 −1.0395
[10] Van Overschee, P., and De Moor, B., 1996. Subspace  0.1651 0.0782 
 
Identification for Linear Systems: Theory-Implemantation-  −0.7939 −0.6635  −3
Applications. Kluwer Academic Publishers. B=  10
 0.7051 0.7799 

[11] Cauberghe, B., Guillaume, P., Pintelon, R., and Verboven,  −0.1138 0.5930 
P., 2003. “Mimo state space frequency domain system iden- −0.6482 −1.4204
tification and model validation from arbitrary signals”. Un-
D = −0.3937 −0.5708 10−3
 
der review in, IEEE Transaction on Automatic Control .
[12] Cauberghe, B., Guillaume, P., Pintelon, R., and Verboven, (41)
P., 2003. “Frequency domain subspace identification using
frf data from arbitrary signals”. Under review in, Journal
of Sound and Vibration .
[13] Van Overschee, P., and De Moor, B., 1993. “Subspace al-
gorithms for the stochastic identification problem”. Auto-
matica, 29 (3) , pp. 649–660.
[14] Cauberghe, B., Guillaume, P., and Pintelon, R., 2003. “Fre-

11 Copyright
c 2004 by ASME

Downloaded From: https://proceedings.asmedigitalcollection.asme.org on 06/19/2019 Terms of Use: http://www.asme.org/about-asme/terms-of-use