Journal of Sound and Vibration 476 (2020) 115314

Contents lists available at ScienceDirect

Journal of Sound and Vibration

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

Improving the double-exponential windowing method to

identify modal frequencies and damping ratios of dynamically
large structures
S. Zhang a, X. Sheng a, b, *, S. Yang a, Z. Chen a
a
State Key Laboratory of Traction Power, Southwest Jiaotong University, Chengdu, Sichuan, 610031, China
b
School of Urban Railway Transportation, Shanghai University of Engineering Science, Shanghai, 201620, China

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

Article history: Dynamically large structures can have many closely spaced modes. A particular example is
Received 5 September 2019 an extruded aluminium panel used for building high-speed trains. Such a structure may be
Received in revised form 19 February 2020 modelled deterministically with a large number of degrees of freedom (dofs), or statisti-
Accepted 9 March 2020
cally represented as a statistical energy analysis (SEA) system having particular SEA pa-
Available online 13 March 2020
Handling editor: I. Trendafilova
rameters. In both cases, experimental modal parameter identification may be required.
Although there are several commercial packages which can be used for modal parameter
identification, alternative methods with different strengths and/or convenience are still
Keywords:
Statistical energy analysis
desirable. In this paper, the double-exponential windowing method developed by the
Experimental modal analysis second author is revisited and improved. Improvements are achieved by taking into ac-
Modal damping ratio count the effect of the finite time duration of Fourier transforms and making use of the real
Double-exponential windowing parts of driving point mobilities. The method is easy to implement and the improvements
make the choice of the decay rate in the exponential window much flexible: for a very
lightly damped structure, it allows to use a large positive decay rate to depress mea-
surement noise, and for a highly damped structure, it allows to use a negative decay rate to
make modes more evident or to separate overlapped modes. The usefulness of the
improved method are demonstrated for a 30-bladed wheel model with known modal
parameters, and for an extruded aluminium panel.
© 2020 Elsevier Ltd. All rights reserved.

1. Introduction

Nowadays the statistical energy analysis (SEA) method is more and more often used to analyse vibro-acoustic behaviours
of complex built-up structures such as those found on ships [1], aeroplanes [2] and high-speed trains [3]. In order to build a
SEA model, SEA parameters of subsystems, including modal densities, internal loss factors (or modal damping loss factors)
and coupling loss factors etc., have to be determined. Modal densities and coupling loss factors may be determined
analytically, numerically or experimentally, but for internal loss factors, the last means is normally an only choice. It is also a
fact that, when determining the coupling loss factor between two subsystems analytically or numerically, the internal loss
factors of the subsystems must be known. There are a number of measurement-based techniques for SEA parameter

* Corresponding author. State Key Laboratory of Traction Power, Southwest Jiaotong University, Chengdu, Sichuan, China.
E-mail addresses: [email protected] (S. Zhang), [email protected] (X. Sheng), [email protected] (S. Yang), [email protected] (Z.
Chen).

https://doi.org/10.1016/j.jsv.2020.115314
0022-460X/© 2020 Elsevier Ltd. All rights reserved.
2 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

determination [4], e.g., the power input method (PIM) [5e8], the decay rate (DR) method [9,10], and of course, the experi-
mental modal analysis (EMA) method [11,12].
The PIM is simply based on the principle that power input to a structure is equal to power dissipated by the structure [5].
By definition, the internal damping loss factor of the structure is the ratio of the power dissipated in one cycle to the product
of the reversible mechanical energy and 2p. This method is applied in Ref. [6] to a typical high-speed train-used extruded
aluminium panel to identify loss factor for each bay (i.e. vibration energy of only one bay is computed, without considering
vibrational energy of other parts of the panel). In addition, with PIM, internal and coupling loss factors can also be determined
by inversing the SEA power balance equations [7].
There are some difficulties in PIM. The main difficulty is the estimation of the vibrational energy, since in theory, it requires
vibrations at many points of a structure to be measured. In practice however, vibrations at only a limited number of points are
measured and those at other locations have to be estimated using certain assumptions and/or methods.
According to the DR method, the internal loss factor of a structure is determined from the decay rate of a free vibration
response which remains after excitations stop suddenly. The decay rate may change with time. In Ref. [8], the authors point
out that the overall average energy in a frequency band is dominated by the energy in lightly damped modes. If loss factors in
a frequency band vary largely, it is difficult to deduce an appropriate average value. It is pointed out in Ref. [9] that, an increase
in the number of modes in a frequency band and/or in damping can reduce the spatial variation of the estimated initial decay
rate. Ref. [10] presents an automatic algorithm for identifying loss factors, but for multi-slope cases, the error of the algorithm
is significant and remains to be explored. When there are multiple decay rates in a frequency band, it may be difficult to use
the DR method to accurately identify the frequency band averaged damping loss factor.
A number of studies can be found on comparison between the above two methods of determining internal loss factors for a
structure. Authors in Ref. [13] demonstrate that both methods can give accurate loss factor estimations as long as the damping
values remain realistic for a linear system and at least one modal resonance is present in each frequency band. However,
Ref. [8] indicates that, when one or two very lightly damped modes are accompanied by several more heavily damped modes
in a frequency band, results of these two methods may be quite different. It can be seen that there are some difficulties and
uncertainties in using these two methods to identify loss factors.
In many cases, SEA parameters can still be determined by traditional experimental modal analysis (EMA). EMA has evolved
to become a ‘standard’ approach in mechanical product development. In the technical service market, there are several
commercial EMA packages, for example, LMS PolyMAX [14], DASP-Modal [15] and Me’Scope [16] etc. EMA methods can be
grouped as time domain methods (e.g. the least-square complex exponential method) and frequency domain methods (e.g.
the least-squares complex frequency-domain method and PolyMAX). A time domain method is normally based on driving-
and/or cross-point impulse response functions (IRF) and a frequency domain method is normally based on driving- and/or
cross-point frequency response functions (FRF). In general, a frequency response function is calculated to be the Fourier
transform of the response divided by that of the force generating the response, and the associated impulse response function
is then estimated by inversely Fourier transforming the frequency response function. Only the force and response are actually
measured and recorded as a discrete time history.
Although several EMA packages are commercially available, alternative methods or tools with different strengths and/or
conveniences are still emerging [17e20]. This is driven either by specific engineering applications or by budget consideration.
Alternatives can be in excitation means or modal parameter identification method. In Ref. [17], several identification methods
are implemented and compared for a funnel-shaped inlet of magnetic resonance tomography. Ref. [18] introduces a new
experimental modal analysis (EMA) method developed for synchronous machines in which the traditional instrumented
hammer impulse is substituted with a magnetic pulse generated by currents applied at two poles.
For the purpose of identifying modal damping ratios for very lightly damped turbocharger turbine wheels, the second
author has developed the so-called double-exponential windowing method in Refs. [19,20] in three versions. The first version
is presented in Ref. [19]. In this version, two sets of exponentially windowed FRFs are first produced for the structure in
question using two slightly different decay rates. Modal frequencies of the structure are determined by simply locating peaks
in any one, or a combination of several, exponentially windowed FRFs. The modal damping ratios of the structure and the
participation factors are determined based on the condition that they satisfy the two corresponding windowed FRFs at all
modal frequencies simultaneously. The main drawback of this version is that it cannot separate two modes having the same
modal frequency.
Repeated modal frequencies always exist to a bladed wheel since the wheel is always designed to be a cyclically periodic
structure. Although practically unavoidable wheel mistuning may destroy the wheel periodicity and a repeated modal fre-
quency may become two different frequencies, these two frequencies are normally still close to each other. The second
version developed in Ref. [19] is mainly for accounting for closely spaced, or repeated, modal frequencies of a bladed wheel.
According to this version, two windowed FRF matrices associated with all the necessary degrees of freedom are first produced
using two slightly different decaying rates. Then singular value decomposition (SVD) is performed on these two FRF matrices
to generate two sets of singular values as function of frequency (called complex modal indication function, or CMIF). Modal
frequencies are determined by simply locating the peaks in the first CMIF (i.e. the CMIF formed with the maximum singular
value) of one of the windowed FRF matrices. The CMIFs of the windowed FRF matrices at the modal frequencies provide
sufficient conditions to determine the modal damping ratios and modal masses. The eigen vector obtained in SVD at a modal
frequency gives the corresponding modal shape.
 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 3

The second version requires measurement of FRFs between all the degrees of freedom which are necessary to catch up the
modal shapes. For example, if the number of dofs is 12, the number of FRFs which are required to measure is 78 if symmetry of
the FRF matrix is utilised. Therefore, for a structure of many dofs (e.g. the extruded aluminium panel studied in this paper),
this method is too time-consuming.
The second version of the double-exponential windowing method, like other classic methods, require driving-point FRFs
to be measured. This is sometimes difficult or impossible. For example, it is not easy to measure driving point FRFs for a
turbine wheel used on a down-sized petrol engine, since the blades are too small. Thus, a third version of the exponential
windowing method is proposed in Ref. [20] which does not require any driving point FRF to be measured. The main idea is to
mathematically construct the unmeasured driving point FRFs from the measured cross-point FRFs to generate a full FRF
matrix. By applying the second version of the double-exponential windowing method to this full FRF matrix, modal pa-
rameters are determined.
As pointed out above, the second and third versions of the double-exponential windowing method requires many FRFs to
be measured, and the number of measurements may be too large for a complex structure. In contrast, requirement on the
number of FRFs of the first version of the method is limited, although with a limited number of FRFs it may not be able to
produce modal shapes. This, however, is not an issue for SEA, since SEA does not require modal shape information. It is also a
fact that, although the first version is presented in Ref. [19], the usefulness of the method has never been investigated. For
EMA and SEA workers to have options of different strengths and/or conveniences, the usefulness of, and possible improve-
ment to, this method are required to be investigated.
In this paper, the first version of the double-exponential windowing method is revisited and improved. Details are pre-
sented in Section 2. In Section 3, the usefulness of the improved method compared with PolyMAX are demonstrated for a 30-
bladed wheel model with known modal parameters. In Section 4, these two methods are applied to identify modal fre-
quencies and damping ratios for a train-used extruded aluminium panel, generating further comparison between them. The
paper is concluded in Section 5.

2. Improvement of the double-exponential windowing method in Ref. [19].

In this section the double-exponential windowing method developed in Ref. [19] is revisited and improved. The method is
based on the concept of exponentially windowed mobility derived from force and velocity time histories measured in
hammer testing. Improvement is made by taking into account the effect on the exponentially windowed mobility of the fact
that Fourier transforms are based on two sets of truncated data.
The theoretical exponentially windowed mobility is described in Section 2.1. The expression for this exponentially
windowed mobility changes if the associated time histories are truncated, as discussed in Section 2.2. The double-exponential
windowing method is formulated in Section 2.3. The method is further commented in Section 2.4. Since use is made of
mobility in the method while often it is acceleration, rather than velocity, that is measured, the exponentially windowed
mobility has to be estimated from measured acceleration data. This is discussed in Section 2.5.

2.1. Theoretical exponentially windowed mobility

A mobility or velocity frequency response function (FRF) of a structure can be expressed as [21].

X
R
as
HðuÞ ¼ ; (1)
s¼1
i uus  i uus þ 22s0

pffiffiffiffiffiffiffi
if the structure is assumed to be linearly and proportionally damped, where, i ¼ 1, R is the number of modes which
contribute to the FRF, as is a real constant, termed the sth modal participation factor, us and 2s0 are, respectively, the sth modal
frequency and damping ratio. It is known that as > 0 if the mobility is a driving point mobility. From Eq. (1) it can be seen that,
in case of a single degree-of-freedom (dof), the real part absolute of the FRF peaks just at the modal frequency. This is the
reason for using mobility to formulate the double-exponential windowing method. For R > 1 and relatively light damping, it
can still be assumed that peaks of the real part absolute of the FRF occur at each modal frequency (more comments are given
in Section 2.4).
A mobility can be calculated from the Fourier transform of the vibrational velocity divided by that of the force exciting the
vibration. An exponential window is often applied to both the force and response signals before the Fourier-transforms are
performed. An exponential window is defined by

wðtÞ ¼ eb1 t ; ðt  0Þ; (2)

where b1 > 0 is termed the decay rate of the window function (as can be seen in this paper, the method in this paper allows a
negative value for b1. In this case, it is still termed the decay rate of the window function). The decay of a signal, x(t), multiplied
by the window function (in other words, x(t) is windowed), at the end of a time duration T is given by
4 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

   
 xð0Þe0   
20log10   ¼ 20log xð0Þ þ 20b1 Tlog e ¼ D0 þ D1 ; (3)
xðTÞe b 1 T  10 
xðTÞ 10

where, D0 in dB is the decay of the original signal during the time duration, and

D1 ¼ 20b1 Tlog10 e ¼ 8:686b1 T; (4)

in dB is an extra decay caused by the exponential window. From Eq. (4),

b1 ¼ D1 =ð8:686TÞ: (5)

with the window (Eq. (2)) applied to both the force and response, the mobility expressed in Eq. (1) (called a theoretical
exponentially windowed mobility) becomes [22,23].

X
R
as
H 1 ð uÞ ¼   ; (6)
s¼1
i uus  i 1 þ 22s0 2s1 þ 22s1 uus þ 2ð2s0 þ 2s1 Þ

where,

2s1 ¼ b1 =us ; (7)

is the sth artificial damping ratio introduced by the exponential window. Small damping is assumed so that

22s0 2s1 þ 22s1 ≪1; (8)

and Eq. (6) is approximated by

X
R
as
H 1 ð uÞ ¼ : (9)
s¼1
i uus  i uus þ 2ð2s0 þ 2s1 Þ

It can be show that 22s0 2s1 þ 22s1 < 0:01 if j2s1 j 2s0 and 2s0 < 0:06. In other words, Eq. (8) holds if damping ratio is less than
0.06. This can be met by most bare metal structures.
In both numerical simulation and measurement, the force and response (velocity) (they are termed measured force and
response) can only be recorded as a time series for a finite time duration T. After that they are assumed to be zero. The mobility
calculated based on those time series may be termed measured exponentially windowed mobility. By equating this measured
mobility to that expressed in Eq. (9), an equation is established from which modal parameters (i.e. as , us and 2s0 ) can be
determined (identified). Current methods are all based on this equality. It can be seen that, unless the measured force and
response decay sufficiently (this is one of the purposes to apply the exponential window Eq. (2)) at the end of the time
duration, significant difference may exist between the measured mobility and Eq. (9) (see the next section). To reduce the
difference, Eq. (9) should be modified as shown in the next section. The modification constructs the main improvement to the
double-exponentially windowing method developed in Ref. [19].

2.2. Truncated exponentially windowed mobility

Now the force is denoted by f(t), the velocity by v(t), and the associated displacement impulse response function is denoted
by h(t). Since the force is assumed to be an impact force, it vanishes when t  t0 , where t0 ≪T is the time within which the
hammer is in contact with the structure. According to the Duhamel Integral,

Z∞ Z∞ Zt0
d _  tÞf ðtÞdt ¼ _  tÞf ðtÞdt:
vðtÞ ¼ hðt  tÞf ðtÞdt ¼ hðt hðt (10)
dt
0 0 0

Multiplied by the exponential window defined in Eq. (2), Eq. (10) becomes

Zt0
eb1 t vðtÞ ¼ _  tÞf ðtÞdt:
eb1 t hðt (11)
0

Performing Fourier transform for a period T on Eq. (11) gives

 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 5

ZT
  
~vwd u ¼ eb1 t v t eiut dt
0 0 t 1
ZT Z0
_    
¼ @ e 1 h t  t f t dtAeiut dt
b t

0 0 0t 1
ZT Z0
b1 ðttÞ b1 t _
   
¼ @ e e h t  t f t dtAeiuðttÞ eiut dt
0 0 0 1
Zt0 ZT
b1 ðttÞ _   iuðttÞ  
¼ @ e h tt e dt Aeb1 t f t eiut dt
00 0 1
Zt0 ZTt
   
¼ @ eb1 t h_ t eiut dt Aeb1 t f t eiut dt: (12)
0 t

Since 0  t  t0 , t0 ≪ T and dh(t)/dt ¼ 0 for t < 0, Eq. (12) reduces to

0 1
Zt0 ZT
~vwd ðuÞ ¼ @ _
eb1 t hðtÞeiut A b1 t
dt e f ðtÞeiut dt
0 0
0 10 t 1
ZT Z0
¼@ eb1 t _
hðtÞeiut A@
dt e b1 t iut
f ðtÞe dtA
0 0
~ ðuÞF~ ðuÞ;
¼H (13)
1 wd

where, F~wd ðuÞ is the windowed Fourier transform of the force, given by

Zt0 ZT
F~wd ðuÞ ¼ eb1 t
f ðtÞe iut
dt ¼ eb1 t f ðtÞeiut dt; (14)
0 0

and

ZT
~ ðuÞ ¼ _ iut ~vwd ðuÞ
H eb1 t hðtÞe dt ¼ : (15)
F~wd ðuÞ
1
0

is an approximation of the theoretical exponentially windowed mobility (Eq. (9)), which, according to Eq. (13), can be
calculated to be the Fourier transform of the windowed velocity divided by that of the windowed force for the time period T.
Eq. (15) is called ‘truncated exponentially windowed mobility’.
To identify modal parameters for a structure from measured data, a relationship must be established between the
truncated exponentially windowed mobility, H ~ ðuÞ, and modal parameters of the structure. It is well known that, the
1
displacement time response function in units m/N, h(t), can be calculated to be the inverse Fourier transform of the
displacement frequency response function, i.e.

Z∞
1 XR
as =ðiuÞ
hðtÞ ¼ eiut du; (16)
2p s¼1 i uus  i uus þ 22s0
∞

which, for small damping, can be worked out to be

8
> 0; t < 0;
>
>
<
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi  X 
hðtÞ ¼ X R
as 2s0 us t
R
as (17)
>
>
> q ffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi e sin 1  22s0 us t z e2s0 us t eius t  eius t ; t  0:
: 2i
s¼1 1  22s0 s¼1

Insertion of Eq. (17) into Eq. (15) gives (note: 2s1 ¼ b1 =us , h (0) ¼ 0)
6 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

ZT ZT
~ ðuÞ ¼
H _
eb1 t hðtÞe iut
dt ¼ hðTÞeb1 T eiuT þ ðiu þ b1 Þ eb1 t hðtÞeiut dt
1
0 0
X
R
¼ as eð2s0 þ2s1 Þus T eiuT sinðus TÞ
s¼1
8 T 9
Z ZT
XR
as ðiu þ 2s1 us Þ < i½1þið2s0 þ2s1 Þus t iut i½1ið2s0 þ2s1 Þus t iut
=
þ e e dt  e e dt
2i : ;
s¼1
0 0
X
R
¼ as eð2s0 þ2s1 Þus T eiuT sinðus TÞ
s¼1
X
R
as ð1  i2s1 us =uÞ h i
þ 1  eð2s0 þ2s1 Þus T eiðuus ÞT
s¼1
2½ið1  us =uÞ þ ð2s0 þ 2s1 Þus =u
X
R
as ð1  i2s1 us =uÞ h i
 1  eð2s0 þ2s1 Þus T eiðuþus ÞT ;
s¼1
2½ið1 þ us =uÞ þ ð2s0 þ 2s1 Þus =u

or,

X
R
~ ðuÞ ¼
H as eð2s0 þ2s1 Þus T eiuT sinðus TÞ
1
s¼1

X
R
as ð1  i2s1 us =uÞ h i
þ 1  eð2s0 þ2s1 Þus T eiðuus ÞT
s¼1
2½ið1  us =uÞ þ ð2s0 þ 2s1 Þus =u

X
R
as ð1  i2s1 us =uÞ h i
 1  eð2s0 þ2s1 Þus T eiðuþus ÞT : (18)
s¼1
2½ið1 þ us =uÞ þ ð2s0 þ 2s1 Þus =u

Eq. (18) gives the truncated exponentially windowed mobility expressed in terms of modal parameters. This is much more
complicated than the theoretical exponentially windowed mobility, Eq. (9). However, it can be shown that, as T/ ∞, the
truncated exponentially windowed mobility becomes the theoretical exponentially windowed mobility.
The real part of Eq. (18) is

~ ðuÞÞ ¼
ReðH 1

X
R    
as eð2s0 þ2s1 Þus T sin us T cos uT
s¼1
n         o  
XR as 1  eð2s0 þ2s1 Þus T cos u  us T þ 2s1 us u eð2s0 þ2s1 Þus T sin u  us T 2s0 þ 2s1 us u
þ h i
s¼1 2 ð1  us =uÞ2 þ ð2s0 þ 2s1 Þ2 ðus =uÞ2
n     h   io 
XR as eð2s0 þ2s1 Þus T sin u  us T  2s1 us u 1  eð2s0 þ2s1 Þus T cos u  us T 1  us u
þ h i
s¼1 2 ð1  us =uÞ2 þ ð2s0 þ 2s1 Þ2 ðus =uÞ2
n         o  
XR as 1  eð2s0 þ2s1 Þus T cos u þ us T þ 2s1 us u eð2s0 þ2s1 Þus T sin u þ us T 2s0 þ 2s1 us u
 h i
s¼1 2 ð1 þ us =uÞ2 þ ð2s0 þ 2s1 Þ2 ðus =uÞ2
n     h   io .
XR as eð2s0 þ2s1 Þus T sin u þ us T  2s1 us u 1  eð2s0 þ2s1 Þus T cos u þ us T 1 þ us u
 h i : (19)
s¼1 2 ð1 þ us =uÞ2 þ ð2s0 þ 2s1 Þ2 ðus =uÞ2

Eq. (19) may be written in a more compact form

 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 7

X
R
~ ðuÞÞ;
Cðu; 2s0 ; 2s1 ; TÞas ¼ ReðH (20)
1
s¼1

where, Cðu; 2s0 ; 2s1 ; TÞ is a function of u, 2s0 , 2s1 and T, and its expressions can be easily deduced from Eq. (19).
To quantify the difference between the theoretical exponentially windowed mobility and the truncated one, a single-
degree-of-freedom system is considered. For a single-degree-of-freedom system, Eqs. (19) and (9) becomes

~ ðus ÞÞ ¼ as eð2s0 þ2s1 Þus T sinðus TÞcosðus TÞ þ as 1  eð2s0 þ2s1 Þus T

ReðH 1
2ð2s0 þ 2s1 Þ

as 1  eð2s0 þ2s1 Þus T cosð2u s TÞ þ 2s1 e s0 þ2s1 Þus T sinð2u
ð2
s TÞð2s0 þ 2s1 Þ

8 þ 2ð2s0 þ 2s1 Þ2

as eð2s0 þ2s1 Þus T sinð2us TÞ  2s1 1  eð2s0 þ2s1 Þus T cosð2us TÞ
 ; (21)
4 þ ð2s0 þ 2s1 Þ2

as
ReðH1 ðus ÞÞ ¼ ; (22)
2ð2s0 þ 2s1 Þ

and the relative error between them is plotted in Fig. 1 for us =ð2pÞ ¼ 500 Hz, 2s ¼ 0:005 and for different decay rates and time
lengths. This figure demonstrates that difference between the truncated exponentially windowed mobility and the theo-
retical one can be significant. The difference can be reduced by increasing either the sampling duration T or the decay rate of
the exponential window, with the former being much more effective. However, if the measurement environment is not good
enough, a too long sampling duration will reduce the signal-to-noise (SeN) ratio of the signals.
In summary, when identifying modal parameters from measured force and response, the truncated exponentially
windowed mobility, Eq. (18), should be made to equate the measured exponentially windowed mobility, as discussed below
in Section 2.3.

2.3. Determination of damping ratios from a set of measured driving point mobilities

Now Eq. (20) is written for the driving point mobility of the kth dof,

X
R
~ ðuÞÞ;
C1 ðu; 2s0 ; 2s1 ; TÞaks ¼ ReðH (23)
1k
s¼1

~ ðuÞ represents the measured

where, aks > 0 is the sth modal participation factor of the kth driving point mobility, and now H 1k
exponentially windowed mobility. Any one driving point mobility may not contain all the modes interested, therefore, a
number of, or all, driving point mobilities are measured. To minimise measurement noise, these driving point mobilities are
averaged to give

Fig. 1. Difference between the truncated exponentially windowed mobility and the theoretical one. dd, T ¼ 0.1s,   , T ¼ 0.2s;  ∙ , T ¼ 0.3s.
8 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

X
R
1 X
M
1 X
M
C1 ðu; 2s0 ; 2s1 ; TÞ aks ¼ ~ ðuÞÞ;
ReðH 1k
s¼1
M M
k¼1 k¼1

or

X
R
~ ðuÞÞ;
C1 ðu; 2s0 ; 2s1 ; TÞas ¼ ReðH (24)
1
s¼1

where, M is the number of mobilities,

1 X
M
as ¼ aks ; (25)
M
k¼1

is a positive real number, and

X
M
~ ðuÞ ¼ 1
H ~ ðuÞ;
H (26)
1 1k
M
k¼1

is the average of all the measured exponentially windowed driving point mobilities.
Based on the absolute of Eq. (26), a curve can be produced for the frequency range considered. Modal frequencies can be
determined by looking at the peaks in the curve. At the rth modal frequency, Eq. (24) becomes

X
R
~ ður ÞÞ; ðr ¼ 1; 2; /RÞ
C1 ður ; 2s0 ; 2s1 ; TÞas ¼ ReðH (27)
1
s¼1

In Eq. (27), the right-hand terms are known, but as and 2s0 (s ¼ 1; 2; /R) are yet to be determined. Since the number of
equations in Eq. (27) is R while the number of unknown is 2R, another set of at least R equations has to be established. It can be
established in different ways, for example, by making use of the imaginary parts of the measured exponentially windowed
driving point mobilities. However, the imaginary part has a local minimum at a modal frequency, and is vulnerable to noise
contamination.
Following the idea of the double-exponential windowing method proposed in Ref. [19], a new set of equations like Eq. (27)
is established with an exponential window of a slightly different decay rate b2 (e.g. b2 ¼ 1:005b1 ),

X
R
~ ður ÞÞ; ðr ¼ 1; 2; /RÞ
C2 ður ; 2s0 ; 2s2 ; TÞas ¼ ReðH (28)
2
s¼1

where, 2s2 ¼ b2 =us .

Now Eqs. (27) and (28) provide a sufficient number of equations to determine unknowns as and 2s0 (s ¼ 1; 2;/R). However,
they are non-linear in terms of the unknown modal damping ratios, and a nonlinear equation solver should be employed. A
nonlinear equation solver normally requires initial values for the unknowns. These initial values may be determined by
treating each mode to be an isolated degree of freedom. Under this approximation, Eqs. (27) and (28) reduce to

ð2s0 þ2s1 Þus T

~ ðus ÞÞ ¼ 1as eð2s0 þ2s1 Þus T sinð2us TÞ þ as 1  e
ReðH 1
2 2ð2s0 þ 2s1 Þ
as 1  eð2s0 þ2s1 Þus T cosð2us TÞ þ 2s1 eð2s0 þ2s1 Þus T sinð2us TÞ ð2s0 þ 2s1 Þ

8 þ 2ð2s0 þ 2s1 Þ2

as eð2s0 þ2s1 Þus T sinð2us TÞ  2s1 1  eð2s0 þ2s1 Þus T cosð2us TÞ

4 þ ð2s0 þ 2s1 Þ2
¼ C1 ðus ; 2s0 ; 2s1 ; TÞas ; (29)
ð2s0 þ2s2 Þus T
~ ðus ÞÞ ¼ 1as eð2s0 þ2s2 Þus T sinð2us TÞ þ as 1  e
ReðH 2
2 2ð2s0 þ 2s2 Þ
as 1  eð2s0 þ2s2 Þus T cosð2us TÞ þ 2s2 eð2s0 þ2s2 Þus T sinð2us TÞ ð2s0 þ 2s2 Þ

8 þ 2ð2s0 þ 2s2 Þ2
 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 9


as eð2s0 þ2s2 Þus T sinð2us TÞ  2s2 1  eð2s0 þ2s2 Þus T cosð2us TÞ

4 þ ð2s0 þ 2s2 Þ2
¼ C2 ðus ; 2s0 ; 2s2 ; TÞas ; (30)

Thus, a single nonlinear equation for 2s0 is generated,

~ ðus ÞÞ ¼ C ðus ; 2 ; 2 ; TÞReðH

C2 ðus ; 2s0 ; 2s2 ; TÞReðH ~ ðus ÞÞ; (31)
1 1 s0 s1 2

from which initial values for the modal damping ratios can be worked out. With these initial modal damping ratios, initial
values for as (s ¼ 1; 2; /; R) can be easily worked out from Eq. (27) since now Eq. (27) is a set of linear algebraic equations.

2.4. Some comments

In Section 2.3, the double-exponential windowing method is formulated in terms of the real part of the average of a
number of driving point mobilities, and only values at modal frequencies are required. These treatments have the following
advantages: 1) a driving point mobility normally is less noise contaminated than a cross point mobility; 2) measurement
noise can be further reduced by the averaging process; 3) a unique estimate is produced for modal frequencies and damping
ratios. These advantages are demonstrated by the first application in Section 3.
However, a mode of a structure may be exhibited at slightly different modal frequencies in different driving point mo-
bilities due to various reasons (e.g. measurement noise, local mass effect of transducer, damping effect, frequency resolution
etc.), multiple closely spaced peaks may occur to the real part of the average of these driving point mobilities. On the other
hand, the average process may make two genuine, but closely spaced, modes become a single one (this, however, may be
overcome by using a negative decay rate in the exponential window). These issues are more problematic for dynamically large
structures. In this case, the double-exponential windowing method may be applied to each and every driving point mobility,
possibly generating different estimates for modal frequencies and damping ratios. This, however, is not an issue for the
structure treated as a SEA subsystem, since SEA only requires band-averaged loss factors. When the double-exponential
windowing method is applied to a driving point mobility of a dynamically large structure with relatively high damping,
negative decay rate may be used to separate possibly overlapped modes.

2.5. Mobility estimated from measured acceleration

_ for a finite duration of time T,

In many situations an accelerometer is used to pick up vibrational acceleration aðtÞ ¼ vðtÞ
~wd ðuÞ, i.e.
where for t  0, aðtÞ ¼ 0 and vðtÞ ¼ 0. Fourier transform of a(t) windowed with Eq. (2), is denoted by a

ZT
iut
~wd ðuÞ ¼
a eb1 t vðtÞe
_ dt
0

h iT ZT

¼ eb1 t vðtÞeiut  þ ðiu þ b1 Þ eb1 t vðtÞeiut dt
0
0
b1 T iuT
¼ ðiu þ b1 Þ~vwd ðuÞ þ vðTÞe e : (32)

According to Eq. (15),

b1 T eiuT
~ ðuÞ ¼ ~vwd ðuÞ ¼ awd ðuÞ  vðTÞe
~
H ; (33)
F~ wd ðuÞ ðiu þ b1 ÞF~wd ðuÞ
1

in which v(T) is unknown. Fortunately, the term vðTÞeb1 T eiuT can be dropped due to the fact that at the end of the time
~wd ðuÞ (the double-exponential window
period, the vibrational velocity becomes negligible compared to the peak values of a
method only requires those peak values), thus

~wd ðuÞ
a
~ ðuÞ ¼
H ; (34)
ðiu þ b1 ÞF~ wd ðuÞ
1

In other words, the Fourier transform of the windowed velocity is given by that of the windowed acceleration divided by
ðiu þb1 Þ. In what follows, responses are always taken to be vibrational velocity.
10 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

3. Application to a 30-bladed wheel model

In this section, application is made to a bladed wheel with 30 blades. A bladed wheel is normally designed to be tuned, i.e.
cyclically periodic, having repeated modal frequencies. However in practice, blade mistuning always exists due to casting
process, machining tolerance and so on, and the mistuning may split a repeated modal frequency to two closely spaced modal
frequencies (it is likely that a repeated modal frequency remains unchanged). To identify mistuning and predict order
response for a mistuned turbine wheel, a lumped mass-spring model and its differential equations of motion are established
in Ref. [24]. The 30-bladed wheel model considered in this section is the same as that in Ref. [24] but with the number of
blades being 30 and blade mistuning being 10%. Blade vibrational velocity time-histories lasting T ¼ 1 s at a time spacing
Dt ¼ 1  105 s due to a given impact force at the same or other blade tips are generated through numerically solving the
differential equations. The impact force is shown in Fig. 2. There are 30 modes of which the actual modal frequencies and
damping ratios are known and shown in Fig. 3. The modal frequencies and shapes are obtained by directly solving the eigen-
value problem formed by the mass and stiffness matrices of the bladed wheel model. The damping ratios are defined when
forming the damping matrix for the bladed wheel model and they are between about 0.040% and 0.054%.
Mode shapes are shown in Fig. 3. They are formed by the tangential displacement of the blade tips. It can be seen that the
bladed wheel model has a high modal density in frequency range between 4700 Hz and 5600 Hz, but there exists a fairly wide
non-resonance frequency range between 5500 Hz and 6500 Hz where a mode exists at 6502 Hz. This is just the dynamic
characteristics of the bladed wheel. Modes between 4700 Hz and 5600 Hz are featured by that only blades vibrate while the
wheel hub nearly remains stationary. From the last mode shape in Fig. 3 it can be seen that, in the mode at 6502 Hz, all the
blades vibrate in phase and at almost the same amplitude. To balance, the wheel hub has to vibrate out of phase of the blades.
This mode is effectively a mode of zero-diameter number. Compared to other modes in which the hub is nearly stationary, this
mode has a similar modal mass, but much higher modal stiffness, giving rise to a much higher modal frequency.
To demonstrate the robustness of the proposed method, noise with a 3 dB signal-to-noise (S/N) ratio is added to each
velocity time-history independently. The velocity time-history of the first blade tip when the impact force shown in Fig. 2 is
applied at the same blade tip is shown in Fig. 4. It can be seen that the signal is heavily contaminated by the added noise,
showing no decay after a short period of time.
For this bladed-wheel model, the average of the 30 driving point mobilities is used with the double-exponential win-
dowing method. When using the proposed method to identify modal parameters, the exponential window with a high decay
rate has the effect of supressing measurement noise, thereby improving the accuracy of the identified modal damping ratios.
However, the decay rate cannot be too large, otherwise it will make closed spaced modes too overlapped or even indistin-
guishable. From Fig. 3 it can be seen that all but two modes are well separated (compared with the corresponding half power
bandwidth). The modes at 5192 Hz and 5197 Hz are only around 5 Hz apart while the half power bandwidths of these two
modes are about 4.5 Hz. For the response data here produced for the bladed wheel model, it is found that the best choice of
decay is 40 dB, introducing artificial damping ratios ranging from 0.071% to 0.096%. According to the requirement on fre-
quency resolution discussed in Ref. [19], FFT frequency resolution is set to be 0.1 Hz. This fine resolution is realised by zero
padding to the data.
The real and imaginary part of the average of the measured and reconstructed (using the identified modal parameters) 30
driving point mobilities are shown in Fig. 5 and Fig. 6, respectively. It is seen that the measured and reconstructed almost

Fig. 2. The impact force.

 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 11

Fig. 3. Actual modal frequencies, damping ratios and modal shapes of the 30-bladed wheel model. The horizontal axis is for blade number.

Fig. 4. The velocity time-history of the first blade when the impact force is applied at the same blade.

overlap each other. It can also be seen that, for the given S/N ratio level, the averaged mobility is smooth around the modal
frequencies. The identified modal damping ratios are presented and discussed later.
Modal analysis for the bladed-wheel model is also performed using PolyMAX based on the force and velocity time-
histories. Since responses of all the blades are obtained due to the impact force applied at any one blade, a 30  30
mobility matrix can be generated. The mobilities are estimated using FFT with the same exponential window as in the double-
12 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

Fig. 5. Real part of the average of the 30-driving point mobilities. Blue, measured; red, reconstructed.

Fig. 6. Imaginary part of the average of the 30-driving point mobilities. Blue, measured; red, reconstructed.

exponential windowing method applied to both the force and responses. FFT frequency resolution is 1 Hz. The 1st, 5th, 16th
and 23rd column in the mobility matrix is selected in turn for performing the PolyMAX analysis, generating four sets of modal
parameters. These four columns have been chosen for no special reason apart from that the column numbers are roughly
uniformly distributed between 1 and 30. The stabilization diagram between 4800 Hz and 5600 Hz based on the 5th column of
mobilities is shown in Fig. 7, where, ‘s’ represents the stable poles and the modal model size is set to be 200.
Damping ratios obtained with those four columns are shown in Fig. 8. It can be seen that with a single column of mo-
bilities, PolyMAX may fail to identify some modes.
PolyMAX can work on multiple columns of mobilities. So next all the 4 chosen columns of data are input to PolyMAX to
identify modal parameters for the bladed wheel model. Modal damping ratios identified are shown in Fig. 9, in comparison
with those identified using the method in this paper. It can be seen that, now all the 30 modal damping ratios are identified by
PolyMAX, but there two modes (the 1st and the 20th) for which the relative error in the identified damping ratio is greater
than 20%. In contrast, damping ratios identified by the method proposed in this paper are much closer to the actual values. In
fact, the average relative error of the method in this paper is 1.96%, while with PolyMAX, it is 4.05%. The better performance of
the method proposed in this paper may be explained by the facts that the method works on the average of the driving point
mobilities and only requires information at the modal frequencies. These two aspects may have largely minimised the
negative effect of measurement noise.
Except for the last mode, all other modes are closely spaced. However, due to low damping, the modal overlaps are still
greater than 1, although in some cases only just. Therefore, the exponential window used neither damages the visibility of the
 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 13

Fig. 7. The stabilization diagram obtained using the 5th column of mobilities.

Fig. 8. Actual and PolyMAX-estimated damping ratios.

peaks nor invalidates the assumption that the peaks occur at the modal frequencies. However, this is not the case for an
extruded aluminium panel from a train floor, as discussed below.

4. Application to a train-used extruded aluminium panel

In this section, application is made to an extruded aluminium panel from a train floor. The vibro-acoustic behaviour of the
panel has great effect on train interior noise and has been a topic of much research [25,26]. To clarify modal characteristics of
the panel, FE analysis is firstly carried out for the panel in Section 4.1. In Section 4.2, FRF measurements of an extruded
aluminium panel by a force hammer and an accelerometer are described. In Sections 4.3 and 4.4, modal frequencies and
damping ratios are identified by PolyMAX and the method proposed in Section 2, and comparisons are then made between
these two methods.
14 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

Fig. 9. Damping ratios and relative errors obtained by using the method in this paper and PolyMAX.

4.1. FE analysis of the extruded aluminium panel

The cross-section and a photograph of the extruded aluminium panel are shown in Fig. 10. The cross-section has di-
mensions of 0.970  0.985  0.050 m. The structure consists of two bare face plates of 4.2 mm thickness and 22 rib plates. The
top and bottom face plates are divided into 10 and 9 bays, respectively, by the stiffening rib plates, as shown in Fig. 10. Material
density is 2752 kg/m3, Young's modulus is 7.1  1010 Pa, and Poisson's ratio is 0.33.
In order to investigate modal characteristics of the structure, an FE model is established in ANSYS under free boundary
conditions, as shown in Fig. 11. The structure is discretised using Solid185 elements, and node spacing is set to be about
5e7 mm to ensure that there are sufficient number of elements per structural wavelength, giving a total of 382,194 elements.
Furthermore, weld-lines are also discretised in a sufficiently detailed manner.
According to the FE analysis, there are 124 (the 6 rigid body modes are not included) modal frequencies in the frequency
range up to 2500 Hz. These modal frequencies are plotted in Fig. 12 from which modal densities can be derived. The average
spacing between two neighbouring modal frequencies is only about 9 Hz. A histogram plot of the modal frequencies is shown
in Fig. 13 with a frequency spacing of 200 Hz. It can be seen that the number of modes in the band centred at 2100 Hz is about
22.
Mode shapes at two modal frequencies, 194 Hz and 975 Hz, are shown in Fig. 14. It is observed that, when modal frequency
is high, modes local to a bay appear, as indicated in Fig. 14(c) and (d). For these complex modes, modal shapes can hardly be
obtained by conventional measuring techniques.

Fig. 10. The cross-section and a photograph of the extruded aluminium panel.
S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 15

Fig. 11. FE model of the extruded panel.

Fig. 12. Modal frequencies of the panel.

Fig. 13. Mode count of the extruded panel in 200 Hz bandwidth.

16 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

Fig. 14. Mode shapes seen from both sides of the panel. (a) At 194 Hz seen from the top; (b) At 194 Hz seen from the bottom; (c) At 975 Hz seen from the top; (d)
At 975 Hz seen from the bottom.

4.2. FRF measurement of the extruded aluminium panel and determination of modal frequencies

A photograph of the experiment is shown in Fig. 15. The extruded aluminium panel was hung with flexible strings to
simulate free boundary conditions. The instruments used were a B&K force hammer Type 8206e002, a B&K accelerometer
Type 4514B, and a B&K LAN-XI frame Type 3053-B-120.
A large structure, like the panel considered here, has too many degrees of freedom (some are inaccessible) to measure
force- and response-time histories at all the dofs. Therefore, the first step to do measurement is to determine appropriate
loading and response points. Modal analysis may be used to help determine these points by making sure that the chosen
points can catch as many local modes (or substructures) as possible. In this case, the extruded aluminium panel just has 19
bays, thus one point is chosen for each bay on the two faces of the panel. These points have been chosen in such a way that
their locations in the bays are random satisfying a uniform distribution. There are 19 points in total, the locations of which are
shown in Fig. 16. Accelerations of the panel were measured for each bay by gluing an accelerometer to the point in that bay

Fig. 15. A photograph of the experiment.

 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 17

Fig. 16. Locations of the measurement points. (a) On the top face plate and (b) on the bottom face plate.

and varying the excitation point among the 19 chosen points. By making use of reciprocity, 190 FRFs are obtained which can
be used to form a 19  19 acceleration FRF matrix. The signals are sampled for 2 s at a time step of 1.2207  104 s, ensuring a
frequency resolution of 0.5 Hz.
Of course, to pick up the response, a laser vibrometer can also be used. Using a laser vibrometer can overcome some
shortcomings of the accelerometer, e.g. the mass load effect. If a 3D scanning laser vibrometer is used, then the number of
measurement points can be much more. This is beneficial for SEA parameter estimation.
To illustrate, the acceleration of the panel at Location No. 7 is shown in Fig. 17 in both time and frequency domain. It can be
seen that the response time-history has been sufficiently attenuated at the end of sampling time and measurement noise is
insignificant. This suggests that in the double-exponential windowing method, a small decay rate, or even a negative one,
may be used for the exponential window.
It has been pointed out that one special strength of the method is that the choice of the decay rate in the exponential
window is much more flexible, even a negative value is allowed. A negative decay rate can make modes more evident, and
two closed spaced modes less overlapped and more distinguishable. This is demonstrated in Fig. 18 for the 11th driving point
mobility of the panel, where the solid line is produced with an exponential window of 0.1 dB while the dotted line is produced
with an exponential window of 20dB. Care must be exercised in choosing a negative decay rate, since now the damping
ratios introduced by the exponential window are negative, possibly making the denominator 2ð2s0 þ2s1 Þin Eq. (29) (especially
for low frequency modes) so small that the equation becomes irregular, or negative.
The modal frequencies can be identified by simply looking for the peaks of the real part of a driving point mobility, as
shown in Fig. 19 for the 14th driving point mobility. The number of modes identified is 71. Since the total damping is small, the
modal frequencies obtained this way are very close to the true values.

Fig. 17. The acceleration of the panel at Location No. 7. (a) Time-history; (b) frequency spectrum.
18 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

Fig. 18. Real part of the 11th driving point mobility. dd, with an exponential window of 0.1 dB; …, with an exponential window of 20dB.

Fig. 19. Modes identified on the real part of the 14th driving point mobility with decay rate of 20dB.

It is understandable if, for a given frequency range, modal frequencies predicted from the FE model are more than seen in
the measurement. However, the number of identified modal frequencies based on the 19 driving point FRFs is close to that of
predicted ones. This means that, to identify modal frequencies and damping ratios of the panel, it is sufficient to measure one
driving point FRF for each bay, as here. Predicted modal frequencies are compared with measured ones in Table 1 for a
frequency range of 100 e 1000 Hz. It can be seen that measurement and prediction are close to each other, and the absolute
relative errors are less than 2%.

Table 1
Predicted and measured modal frequencies of the extruded panel.

Mode FE predicted modal frequency (Hz) Measured modal frequency (Hz) Relative error (%)
(relative to prediction)
1 194.11 196 0.9737
2 304.51 306 0.4893
3 380.96 384 0.7980
4 471.53 463 1.8090
5 493.29 501 1.5630
6 708.51 718 1.3394
7 762.19 758 0.5497
8 788.45 782 0.8181
9 865.78 874 0.9494
11 975.14 977 0.1907
 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 19

Fig. 20. The stabilization diagram obtained using all the 19 columns of the acceleration FRF matrix.

Although modal frequencies can be accurately determined both numerically and experimentally, the determination of
modal damping ratios has to rely on measurement. Modal damping ratios have been estimated in Ref. [6] using different
methods for panels similar to that shown in Fig. 10, and a high level of uncertainty is shown. This uncertainty is related to the
method and procedure used, the assumptions underpinning them, and a wide range of potential testing and processing
errors. It should be pointed out that, the second version of the double-exponential windowing method cannot be applied to
the FRF matrix described above for determining modal parameters for the panel, since the dofs involved in this FRF matrix is a

Fig. 21. The stabilization diagram obtained using the 1st column of the acceleration FRF matrix.
20 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

small portion of the dofs required for the frequency range considered. However, based on the measured FRFs, damping ratios
of the panel shown in Fig. 10 can be identified, first using PolyMAX (Section 4.3), and then using the method developed in this
paper (Section 4.4).

4.3. Damping ratios of the extruded aluminium panel identified using PolyMAX

With the force and response time-histories, modal frequencies and damping ratios may be determined using PolyMAX.
PolyMAX is normally executed for a single, or several, columns in the FRF matrix. In order to obtain enough stable poles, the
modal size is set to be 300 (note that the number of non-rigid body modes predicted by FE is 124). The stabilization diagram
between 1000 Hz and 2500 Hz is shown in Fig. 20 which is produced using all the 19 columns of the acceleration FRF matrix.
In this figure, ‘s’ (meaning stable) indicates that both frequency, damping and pole vector are stable within tolerances, ‘v’
means that only the pole vector does not change within the tolerances, ‘d’ implies that both the damping and frequency of the
pole do not change within the tolerances, ‘f’ shows that only the frequency of the pole does not change within the tolerances,
and finally, ‘o’ simply tells that the pole is not stable. It can be seen that this diagram only shows few stable poles as the model
order increases. This may be caused by inconsistency in modal information between the columns.
Instead of using all the columns of the FRF matrix, now only a single column is employed to generate a stabilization di-
agram, as shown in Fig. 21 for the first column. Now much more stable poles are clearly displayed.
Stable poles obtained in Fig. 21 may contain spurious numerical modes, although some modes indicated in Fig. 19 do not
appear in Fig. 21 as a stable pole. In order to make sure that a stable pole shown in Fig. 21 is not a false mode, only poles at the
modal frequencies identified in Section 4.2 are selected for the estimation of damping ratios. The results are shown in Fig. 22.

Fig. 22. PolyMAX-estimated damping ratios based on the 1st, 7th, 14th and 18th column of FRFs. (a) For mode number 1e60; (b) for mode number 60e120.
 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 21

There are four sets of estimated damping ratios each with a single column of acceleration FRFs. Two observations can be
made: 1) some modes are missed by PolyMAX with a single column of FRFs; 2) for a given modal frequency, modal damping
ratios estimated from different columns of the FRF matrix can be significantly different.
Since there are 19 columns of FRFs, 19 estimates can be produced for each modal damping ratio. The minimum, maximum
and mean of the 19 estimates are shown in Fig. 23. It can be seen that, for the data measured for this extruded aluminium
panel, the estimated maximum damping ratios can be as large as 4 times the minimums, once again demonstrating a high
level of uncertainty in damping identification for such a complex structure.

4.4. Damping ratios of the extruded aluminium panel identified using the method developed in this paper

Modal damping ratios are now re-identified using the method developed in this paper. A decay rate of 20 dB is chosen for
the exponential window. The method is applied for each of the 19 driving point mobilities. It turns out that, the reconstructed
driving point mobility is quite close to that measured. For example, the real and imaginary part of the measured and
reconstructed (using the identified modal parameters) 14th driving point mobility are compared in Fig. 24 and Fig. 25,
respectively. It is seen that the reconstructed matches the measured quite well.
As commented in Section 2.4, for a given mode, modal frequencies identified from different driving point mobilities may
be different, although slightly. The average of the associated modal damping ratios is taken to be the modal damping ratio of
that mode. The results are shown in Fig. 26 in comparison with the mean of the 19 estimates from PolyMAX. Damping ratios
averaged in every 200 Hz are shown in Fig. 27. It can be seen from these two figures that two methods agree relatively well,
and as expected, they agree better at low frequencies.

Fig. 23. The minimum, mean and maximum modal damping ratios estimated by PolyMAX.

Fig. 24. Real part of the 14th driving point mobility. Blue, measured; red, reconstructed.
22 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314

Fig. 25. Imaginary part of the 14th driving point mobility. Blue, measured; red, reconstructed.

Fig. 26. Damping ratios identified using the method in this paper and the mean by PolyMAX.

Fig. 27. Bandwidth (200 Hz)-averaged damping ratios identified using the method in this paper and PolyMAX.
 S. Zhang et al. / Journal of Sound and Vibration 476 (2020) 115314 23

5. Conclusions

In this paper, the double-exponential windowing method is revisited and improved. Improvements are achieved by
exactly expressing a measured driving point mobility in terms of modal parameters, and by making use of the real part of the
measured driving point mobility at modal frequencies. This method may be regarded as the fourth version of the double-
exponential windowing method.
The strength of the method improved in this paper is that it only uses driving point motilities, which normally have high S/
N ratios, of a structure, is easy to implement and flexible to choose the decay rate in the exponential window. For a lightly
damped structure, it allows to use a large positive decay rate to depress measurement noise, and for a highly damped
structure, it allows to use a negative decay rate to make modes more evident or to separate two overlapped modes.
The usefulness of the improved method are demonstrated for a 30-bladed wheel model with a set of 30 known modal
parameters, and for a train-used extruded aluminium panel, both are dynamically large structures. For the 30-bladed wheel
model, of which the response signals are significantly noise-contaminated, the improved method performs better than
PolyMAX, while for the extruded aluminium panel, two methods generate similar results.
The method is not dedicated to identify modal shapes and may overlook some modes due to insufficiency in the number of
measured FRFs or too closely spaced modes. This, however, is not an issue for the structure treated as a SEA subsystem, since
SEA only requires band-averaged loss factors.

Acknowledgements

This work is funded by the National Key R&D Program of China (2016YFE0205200), the National Natural Science Foun-
dation of China (U1834201), and Scientific Research Foundation of State Key Laboratory of Traction Power (2017TPL_T01).

References

[1] H.J. Kang, H.S. Kim, J.S. Kim, Y.C. Lee, SEA Modelling of Ship Structures, Inter-noise & Noise-con Congress & Conference, 1993.
[2] H. A. Dande, T. Wang, J. Maxon, J. Bouriez, SEA Model Development for Cabin Noise Prediction of a Large Commercial Business Jet, SAE Technical Paper
2017-01-1764.
[3] J. Zhang, X. Xiao, X. Sheng, etc., SEA and contribution analysis for interior noise of a high speed train, Appl. Acoust. 112 (2016) 158e170.
[4] R.H. Lyon, R.G. DeJong, Theory and Application of Statistical Energy Analysis, Butterworth-Heinemann, London, 1995.
[5] E.E. Ungar, E.M. Kerwin Jr., Loss factors of viscoelastic systems in terms of energy concepts, J. Acoust. Soc. Am. 34 (1962) 954e957.
[6] Y. Zhang, D.J. Thompson, G. Squicciarini, et al., Sound transmission loss properties of truss core extruded panels, Appl. Acoust. 131 (2018) 134e153.
[7] D.A. Bies, S. Hamid, In situ determination of loss and coupling loss factors by the power injection method, J. Sound Vib. 70 (1980) 187e204.
[8] M.F. Ranky, B.L. Clarkson, Frequency average loss factors of plates and shells, J. Sound Vib. 89 (1983) 309e323.
[9] L. Wu, A.A. Gren, U. Sundback, A study of the initial decay rate of two-dimensional vibrating structures in relation to estimates of loss factor, J. Sound
Vib. 206 (1997) 663e684.
[10] A. Allen, N. Schiller, R. Cabell, M. Moeller, Loss factor estimation using the impulse response decay method on a stiffened structure, J. Reticuloendothel.
Soc. 15 (2009) 87e95.
[11] B.L. Clarkson, R.J. Pope, Experimental determination of modal densities and loss factors of flat plates and cylinders, J. Sound Vib. 77 (1981) 535e549.
[12] M.P. Norton, R. Greenhalgh, On the estimation of loss factors in lightly damped pipeline systems: some measurement techniques and their limitations,
J. Sound Vib. 105 (1986) 397e423.
[13] B.C. Bloss, M.D. Rao, Estimation of frequency-averaged loss factors by the power injection and the impulse response decay methods, J. Acoust. Soc. Am.
117 (2005) 240e249.
[14] B. Peeters, H. Auweraer, P. Guillaume, J. Leuridan, The PolyMAX frequency-domain method: a new standard for modal parameter estimation, Shock
Vib. 11 (2004) 395e409.
[15] http://www.coinv.com/product/pid/59.html. (Accessed 18 May 2019).
[16] https://www.vibetech.com/. (Accessed 18 May 2019).
[17] T. Nestorovi c, M. Trajkov, M. Patalong, Identification of modal parameters for complex structures by experimental modal analysis approach, Adv. Mech.
Eng. 8 (2016) 1e16.
[18] A. Barros, et al., New method for experimental modal analysis of hydro-generator's stator core using the excitation from the poles, J. Eng. 2019 (2019)
4341e4344.
[19] X. Sheng, A new experimental modal analysis method based on double-exponential windowing and application to lightly damped bladed wheels, J.
Sound Vib. 331 (2012) 2824e2835.
[20] X. Sheng, Q. Feng, Determination of modal parameters from off-diagonal FRFs using the double-exponential windowing method, J. Sound Vib. 333
(2014) 1644e1656.
[21] D.J. Ewins, Modal Testing: Theory and Practice, Research Studies Press LtD, 1995.
[22] S. Ahn, W. Jeong, W. Yoo, Improvement of impulse response spectrum and its application, J. Sound Vib. 288 (2005) 1223e1239.
[23] S. Ahn, W. Jeong, W. Yoo, Unbiased expression of FRF with exponential window function in impact hammer testing, J. Sound Vib. 277 (2004) 931e941.
[24] X. Sheng, Model identification and order response prediction for bladed wheels, J. Sound Vib. 323 (2009) 194e213.
[25] G. Xie, D.J. Thompson, C.J.C. Jones, A modelling approach for the vibro-acoustic behaviours of aluminium extrusions used in railway vehicles, J. Sound
Vib. 293 (2006) 921e932.
[26] L. Ji, X. Sheng, X. Xiao, Z. Wen, X. Jin, A review of mid-frequency vibro-acoustic modelling for high speed train extruded aluminium panels as well as
the most recent developments in hybrid modelling techniques, Journal of Modern Transportation 23 (2015) 159e168.