Research Article

A comparison among modal parameter extraction methods

Abdel Kader Zrayka1 · Emiliano Mucchi1

© Springer Nature Switzerland AG 2019

Abstract
An effective comparison among several methods for extraction of the modal parameters from the frequency response
functions measurements is presented in this research. In particular, several curve fitting methods, which are the peak
amplitude method, the circle fit method, the least square complex exponential method, the eigensystem realization
algorithm method and the rational fraction polynomial method were implemented in Matlab environment and com-
pared in terms of natural frequencies, modal damping and mode shapes. Measurements were performed on a carcass
of the gearbox in free–free condition. A hammer has been used with a periodic impulsive excitation signal. The natural
frequencies values obtained by all methods were very similar and the differences between the results were insignificant.
The peak amplitude and the circle fit gave good results for the damping ratios. The rational fraction polynomial method
did the best job in detecting the damping and frequency values. The results obtained by the least square complex expo-
nential method and the eigensystem realization algorithm method were reasonable for both frequency and damping.

Keywords Eigensystem realization algorithm · Least square complex exponential · Modal analysis methods · Rational
fraction polynomial · Implementation modal analysis by Matlab ®
List of symbols q Number of force input locations
ak Coefficients of the numerator polynomial of t Time
𝛼(j𝜔) Δt Simple time
A State transition matrix characterizing the [U] Orthogonal (or unitary) matrix of left singular
dynamics of the system vectors, in the SVD technique
Arpq Modal constant for mode r [V] Orthogonal (or unitary) matrix of right singular
bk Coefficients of the denominator polynomial of vectors, in the SVD technique
𝛼(j𝜔) x(t) State vector
B Input matrix y(t) Output vector
C Output matrix 𝛼pq (j𝜔) Receptance FRF (mm/N)
D The direct input–output transmission matrix 𝜓u Vector space unscaled eigenvector
fr Natural frequency of mode r (Hz) 𝜓 Mode shape in term of physical coordinate of
[H] Hankel matrix the system
hpq (t) Impulse response function corresponding to 𝜁r Damping factor of mode r
𝛼pq (j𝜔) 𝜃 Phase angle (rad)
h̃pq  Vector formed by hpq (t) elements Δ𝜃 Variation of phase angle between 𝜃i and 𝜃i+1
i Index 𝜆r System pole of mode r
j𝜔 Laplace variable 𝜔 Circular frequency (rad/s)
N Number of degrees-of-freedom of the system 𝜔r Natural frequency of mode r (rad/s)
p Number of measured response locations [𝜮] Singular value matrix

* Abdel Kader Zrayka, [email protected]; Emiliano Mucchi, [email protected] | 1Department of Engineering, University of Ferrara, Via Saragat
1, 44122 Ferrara, Italy.

SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8

Received: 21 February 2019 / Accepted: 19 June 2019 / Published online: 25 June 2019

Vol.:(0123456789)
Research Article SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8

Operators comprehension of corresponding structure. A further

[…]T Transpose matrix widely used application of these modal parameters is to
[…]−1 Inverse matrix control the health of a structure. The modal analysis meth-
F−1 Inverse Fourier trasformation ods can be separated in two leading categories which are
the time domain and the frequency domain methods [2];
Abbreviations
the earliest method works in the frequency domain and
FRF Frequency response function
is classified as an SDOF algorithm such as the peak ampli-
CFM Circle fit method
tude method (PPM) [3] and the circle fit method (CFM)
ERAM Eigensystem realization algorithm method
[4]. These methods give errors in results, particularly in
IRF Impulse response function
the damping estimation, where the modes are close to
MDOF Multi degree of freedom
each other and coupled. In addition, the SDOF algorithm
LSCEM Least square complex exponential method
should not be used when the data contains noise around
PPM Peak amplitude or pick picking method
the resonance [5]. When the MDOF algorithms operate in
RFPM Rational fraction polynomial method
the time domain and frequency domain methods, data
SDOF Single degree of freedom
are managed by assorting the FRFs in the domain. The
least square complex exponential method (LSCEM) is con-
sidered one of the fast time domain methods. [6, 7]. This
1 Introduction method uses the FRFs as an input in spite the fact that
it works in the time domain. The LSCEM uses the Least
Nowadays, technologies play a great role in human life, Square Method to find the modal parameters. The eigen-
advancing the ways of living and working but engineers system realization algorithm method (ERAM) is another
still have many issues to solve in various filed of technol- major MDOF time domain approach [8, 9]. This method
ogy. The vibration of mechanical structure remains one of uses the Singular Values Decomposition of the so-called
the most studied problems due to the difficulty to com- Hankel Matrix which is usually a matrix with a high num-
pletely avoid vibration in many applications. ber of rows and columns; for such a reason this algorithm
Currently, a large number of algorithms and literature is computationally complex [10]. One more MDOF is the
for curve fitting structural data is available; consequently, rational fraction polynomial method (RFPM) [11]; it works
the determination of the optimal method has become dif- in the frequency domain and it uses the Least Square
ficult for each situation. Hence, the main objective of this technique to minimize the error function. For more infor-
paper is the comparison among several modal parameter mation concerning modal analysis techniques can be
extraction algorithms from response vibration measure- addressed in Ref [12, 13].
ments in order to highlight not only their pros and cons. In the frequency domain method, the lower number of
Moreover, this article focuses on how these methods can order is, the more accurate the results are. However, they
be implemented in a simple way in order to find the nat- have specific problems related to the fast Fourier trans-
ural frequencies and the damping ratios to make them form (FFT) analysis [14, 15], such as the leakage [16]. Fur-
easy to use in the dynamic environment. As these authors thermore, the frequency domain methods determine the
are aware, in the literature the detailed explanation on frequency response function, but this task usually requires
the steps used to implement these methods or the tricks the input excitation data to be detailed. In time domain,
necessary to achieve good results are not provided. This signals are monitored by the time domain methods using
mostly happens when the least square complex expo- output responses only. Time domain methods are consid-
nential method, the eigensystem realization algorithm ered very useful for experimental tests due to this charac-
method and the rational fraction polynomial method are teristic. Another useful property of time domain method
used. on the other hand, the peak amplitude and the cir- is the ability to recognize two modes when they are very
cle fit methods are well clarified and achieved. Therefore, close one to each other. In general, the time domain work
it was decided to use the latters to get initial estimation of better than the frequency domain when the damping is
the modal parameters. Furthermore, the implementation low, and the frequency domain estimators work better for
of these methods highly differs between authors. In this high damping, as demonstrated in [17, 18].
work, a simplified and efficient implementation on Matlab It has always been a difficult task to define the poles
® is illustrated, in order to extract modal parameters. The of a system, including the damping and the frequency,
basic problem in experimental modal analysis is to extract from the response vibration responses. In this work, the
the natural frequency and especially the damping ratio of carcass, one of the gear box components, was sepa-
dynamical structure. With this knowledge, a theoretical rately considered for a better estimation of the FRFs
finite element models [1] can validate and gain a better [19]. The measurements were carried out after adding

Vol:.(1234567890)
SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8 Research Article

Note that 𝜔a and 𝜔b are the frequencies on the half-

power points.

2.2 Circle fit method (CFM)

PPM CFM LSCEM ERAM RFPM
The CFM in the past was known as the Kennedy–Pancu
method [4]; it is a single-input single output method and it
Fig. 1  Analysis procedure’s flowchart works in the frequency domain. A detailed study has been
established in Ref [22]. Similar to the PPM, it is designed
to separate each single mode in the system. The concept
a constrained layer to the carcass of gear box to make of this method is to consider the FRF values in the vicin-
the damping degree similar to the free–free condition. ity of the resonance as a circle in the Nyquist plot. In fact,
The structure was excited by a hammer with a periodic once the natural frequency and the damping factor were
impulse. The residues or the zeros of the system Fre- estimated, the diameter of circle is used to estimate the
quency Response Function (FRFs) can be equivalently residues. According to the algorithm, the Circle-Fit method
determined using the periodic response. The original can be described by the following sequences:
contribution of this study is related to the implemen-
tation of several methods in Maltab ® to extract their • Selection of points to be used.
modal parameters in order to then compare their • Circle fitting based on these points and calculation of the
results, discuss the limits of each method and its suitable fitting quality.
domain of employment. For the comparison, the follow- • Estimation of damping ratios and natural frequencies
ing common methods of modal analysis were used: peak • Calculation of multiple damping estimates and their
amplitude methods (PPM), circle fit method (CFM), least mean and scatter.
square complex exponential method (LSCEM), eigen sys- • Determination of the modal constant.
tem realization algorithm method (ERAM) and rational
fraction polynomial method (RFPM). The process fol-
lowed to determine the poles is shown in the flowchart
𝜔r = max(Δ𝜃) (4)
of Fig. 1. 𝜔2i+1 − 𝜔2i
𝜁r = ( ( ) ( )) (5)
𝜃 𝜃i
2𝜔r 𝜔i+1 tan i+1 2
+ 𝜔 i tan 2

2 Background and algorithms

implementation Arpq = 2|𝛼|
̃ pq 𝜔2r 𝜁 (6)

2.1 The peak picking method (PPM) The robustness of this method has been demonstrated
by several studies as given in Ref. [13–23].
The PPM, sometimes referred as a peak-amplitude
method or 3 dB method, is a single-input single output 2.3 Least Square Complex Exponential Method
(SISO) and it is the simplest of the modal parameter esti- (LSCEM)
mation methods that works in the frequency domain [20,
21]. It consists of separating each single mode in order The LSCEM was introduced in 1979 [7], it is a single-input
to determine their modal parameters. Thus, the natural multi-outpout method (SIMO) and it works in the time
frequency, damping factor and residues will be deter- domain. This method starts using the FRF receptance of a
mined as follows, respectively: general MDOF system with a general viscous damping. Then,
the impulse response function will be obtained by an inverse
𝜔r = max(FRF) (1) Fourier trasformation as follows:
𝜔 a − 𝜔b
𝜁r ≅ (2) ∑
2N Arpq F−1 ∑
2N
2𝜔r 𝛼pq (j𝜔) = ⟶ hij (t) = Arpq e𝜆r t
r=1
j𝜔 − 𝜆r r=1

Arpq = ̃ pq 𝜔2r 𝜁
2|𝛼| (3)

Vol.:(0123456789)
Research Article SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8

For the lth sample.

∑
2N
hl = Cr Vrl l = 0, 1 … , N − 1
r=1
{ (7)
Cr = Arpq
with
Vr = e𝜆r t

Equation (7) extended to the full data set of l samples,

gives:

⎡ 1 1 1 … 1 ⎤⎡ C1 ⎤ ⎡ h0 ⎤ Fig. 2  Steps of the LSCEM implemented in Matlab ®

⎢ V1 V2 V3 … V2N ⎥⎢ C2 ⎥ ⎢ h1 ⎥
⎢ ⋮ ⋮ ⋮ ⋮ ⋮ ⎥⎥⎢⎢ ⋮ ⎥⎥ ⎢⎢ ⋮ ⎥⎥
= (8)
⎢ l Eq. (10) can be solved to yield the Vr roots. Then, poles 𝜆r
⎣ V1 V2l V3l … V2N ⎦⎣ C2N ⎦ ⎣ h2N ⎦
can be calculated as.
( )
Or, in compact form: ln Vr
[ ] 𝜆r = (14)
[V ][C] = h (9) Δt

[V ] and [C] are unknowns. The mode shapes of the system can be calculated imme-
This equation can be solved by using Prony’s Method diately by substituting [V ] in Eq. (9). The challenge of this
[24], the roots 𝜆r for an underdamped system always occur method is to construct a good Hankel matrix [H] where the
in complex conjugate pairs. It always exists a polynomial in number of columns represents the number of order and the
Vr of order l with real coefficients β, (called the Autoregres- number of rows are arbitrary. In order to minimize the Least
sive coefficients) such as the following relation is verified: Square Error, some studies supposed to consider a high
number of rows [21, 25]. In this study the number of rows
𝛽0 + 𝛽1 V + 𝛽 2 V 2 + ⋯ + 𝛽 l V l = 0 (10) will be determined by increasing the number of rows one
After some steps explained in detail in the Ref [8], the at a time then the best number of rows will be found by
following equation will be obtained: observing the stabilization diagram [26]. If the stabilization
diagram obtained was not clear enough, the range of time
⎡ h0 h1 h2 … h2N−1 ⎤⎡ 𝛽1 ⎤ ⎡ h2N ⎤ should be changed. A small diagram in Fig. 2 explains the
⎢ h0 h1 h3 … h2N ⎥⎢ 𝛽2 ⎥ ⎢h ⎥
⎢ ⋮ ⎥ ⎢ ⎥ = −⎢ 2N+1 ⎥ (11) procedure of the implementation in Matlab ®.
⎢ ⋮ ⋮ ⋮ ⋮ ⎥⎢ ⋮ ⎥ ⎢ ⋮ ⎥ If it is not possible to obtain a clear stabilization diagram,
⎣ h2N−1 h2N h2N+1 … h4N−1 ⎦⎣ 𝛽2N−1 ⎦ ⎣ h4N−1 ⎦
the range of selected data must be changed and the same
procedure mentioned above will be repeated. This method
From Eq. (11), coefficients β will be determined by using suffers when the damping is high as demonstrated in Ref
the single impulse response via a Least Square Method. [27].
Instead of using a single impulse response function (IRF),
LSCEM estimates coefficients β by using several IRF’s, as 2.4 Eigensystem realization algorithm method
follows. (ERAM)

⎡ h11 ⎤ ⎡ h̃ ⎤
⎢̃⎥
11 The ERAM is a multi-input multi-output method (MIMO) and
⎢ h12 ⎥ h
⎢ ⋮ ⎥[𝛽] = −⎢⎢ 12 ⎥⎥ (12) it works in the time domain [9]. It includes information not
⎢ ⎥ ⋮ only from different output locations, but also from several
⎣ hp ⎦ ⎢ ̃ ⎥
⎣ hp ⎦ input reference points on the structure. This method suc-
ceeded in dealing with the problem of missing one of the
In compact form, vibration modes from output responses, this occasionally
happens following the application of a SIMO method. The
[ ] Singular Values Decomposition (SVD) in this method can sig-
[H][𝛽] = h̃ (13)
nificantly reduce the effect of noise [7]. The ERAM is a very
Coefficient 𝛽 will be obtained by using pseudoinverse effective method for system identification by using Hankel
technique. Now coefficients 𝛽0 , 𝛽1 , … 𝛽2N−1,are known and matrix; further details about this method can be found in
Ref [13, 28].

Vol:.(1234567890)
SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8 Research Article

Let assume a state space dynamic system as follows:

x(t + 1) = Ax(t) + Bu(t) (15)
y(t) = Cx(t) + Du(t)
Let consider an impulse force at t = 0, and the initial condi-
tion equal to zero:

⎧ u(0) = 1;
⎪
⎨ u(t) = 0, t>0
⎪D = 0
⎩
Fig. 3  Steps of the ERAM implemented in Matlab ®
Hence, by iterating the system of Eq. (11) in time, the fol-
lowing parameters will be obtained.
In modal analysis, the goal is to determine matrices  and
x(0) = 0; → y(0) = 0; ̂C, where the eigenvalues of  consist of the complex con-
x(1) = Ax(0) + Bu(0) = B → y(1) = CB; jugates poles of the system. From each pole the natural fre-
x(2) = AB; → y(2) = CAB; quency and the damping ratio can be obtained. The mode
shapes are related to matrix Ĉ . By using matrix  and by solv-
where y(0), y(1), y(2) … , y(t) are the so called Markov ing the Eigen-problem, the mode shapes can be determined
parameters. in terms of the physical coordinates of system:
By constructing the Hankel matrix H0 of the Markov [ ][ ] [ ]
parameters as: Â 𝜓u = 𝛬 𝜓u (22)

⎡ y(0) y(1) … y(p) ⎤ The transformation given by Eq. (11) must be used:
⎢ y(1) y(2) … y(p + 1) ⎥ [ ]
H0 = ⎢ ⎥ (16) 𝜓 = [C] 𝜓u
⏟⏟⏟ ⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎣ y(q) y(q + 1) … y(p + q − 1) ⎦
⏟⏟⏟ (23)
(pxq)
px1

{ } [ ]
And by apply the SVD to the matrix H0 𝛬 = 𝛬1 , 𝛬2 , … , 𝛬2N is the eigenvalues and 𝜓u is the
eigenvector. The poles will be determined by the following
⎡ ⎤ formulae:
⎢ ⎥
⎢ H(0) ⎥ = [U] [𝜮] [V]T (17) ( )
⎢⏟⏟⏟⎥ ⏟⏟⏟ ⏟⏟⏟ ⏟⏟⏟ ln 𝛬r
⎣ pxq ⎦ px2N 2Nx2N 2Nxq 𝜆r = (24)
Δt

And by shifting the Hankel matrix as follows: The ERAM was implemented in Matlab[ ®]by fixing a num-
ber of order N and generating
[ ] the matrix  by iteration. For
⎡ y(1) y(2) … y(p + 1) ⎤ ̂
each iteration, matrix A was generated as a square matrix
⎢ y(2) y(3) … y(p + 2) ⎥ (1 × 1, 2 × 2,…,N × N). Then it was possible to extract the
H1 = ⎢ ⎥ (18)
⏟⏟⏟ ⎢ ⋮ ⋮ ⋱ ⋮ ⎥ eigenvalues and the eigenvector. Furthermore, the poles
(pxq) ⎣ y(q + 1) y(q + 2) … y(p + q + 1) ⎦ were calculated by applying Eq. (24) and the physical mode
shapes by using Eq. (20). Thus, the poles and the mode
After some manipulations [13], the identified discrete shapes were determined. The final step consists of plotting
̂ B̂ and Ĉ can be written
state-space A, the stabilization diagram from which the stable poles are
extracted. Figure 3 represents the main steps used to imple-
[ ] [ ]− 1 [ ]T [ ][ ]− 1
 = 𝜮 2N 2 U 2N [H(1)] V 2N 𝜮 2N 2 (19) ment this method in Matlab ®.

[ ] [ ]1∕2 [ ]T [ ] 2.5 Rational fraction polynomial method (RFPM)

B̂ = 𝜮 2N V 2N Eq (20)
This method works in the Frequency Domain [11], and it
[ ] [ ] [ ][ ]1 is a single-input single-output method (SISO), based on a
T
Ĉ = Ep U 2N 𝜮 2N 2 (21) multi degrees of freedom approach. The Rational Polynomial

Vol.:(0123456789)
Research Article SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8

Fig. 4  Steps of the RFPM implemented in Matlab ®

Fig. 5  Measurement set-up

Method (RFPM) is a special version of the general Curve-Fit-
ting method, but it is based on the FRF which is expressed
in the Rational fraction form [13]. Thus, the FRF requires an
expression of the frequency response function as the ratio of
two polynomials, with the roots of numerator permitting to
determine the modal constant while the roots of the denom-
inator yielding the poles (frequency, damping). Generally,
the numerator and the denominator orders are independ-
ent one to each other. The denominator is considered as the
characteristic polynomial of the system. By Curve-Fitting the
FRF against the analytical form in Eq. (25), and then solving
the roots of the numerator and the characteristic polynomi-
Fig. 6  Amplitude of all the measured frequency response functions
als, the zeros and poles of the FRF can be determined.
∑2N−1
ak (j𝜔)k 3 Measurement and analysis
𝛼(j𝜔) = ∑k=0
2N (25)
k
k=0 bk (j𝜔)
Measurements were carried out on an aluminum carcass of
The curve fitting in the RFPM consists of determin- gearbox with dimension of about 150 mm by 100 mm by
ing the coefficients ak (with k = 0, … 2N − 1) and 40 mm with carcass’s thickness of 5 mm. The carcass was
bk (with k = 0, … 2N), in such a way the error between the supported by soft bungee cords fixed on a frame, in order
analytical formulae (25) and the FRF is minimized over a to consider its dynamic behavior as a free–free condition.
chosen range; more information is given in Ref [29]. Figure 5 shows the set-up of the test measurement used in
The RFPM was implemented in Matlab ® by firstly select- this study. As it can be seen, the periodic impulsive excita-
ing a small range of frequency and by “invfreqs” command tion is generated by using a hammer (PCB 068C04), while
the coefficients ak and bk were determined. Then, the poles the response is measured by using PCB piezoelectric accel-
and the residues were determined by using “residue” com- erometer (freq. range 1–10000 Hz). One accelerometer was
mand. More information about “invfreqs” and “residue” positioned, and the system was excited by a hammer in
command are included in Ref [30]. If it is not possible to 34 points. In particular, the response point was remained
obtain a clear stabilization diagram, the range of the fre- fixed during the tests, while the excitation changed from
quency will be changed. Figure 4 describes the important one measurement point to another in order to get the FRFs
steps that were used to implement the RFPM. for all the considered points [31, 32]. The measured data
This method requires more time with respect to other were recorded using software LMS Test-Lab ® [33] and the
methods to achieve a good estimation for the frequency post processing was performed in Matlab ® environment.
and damping ratio. The signals were acquired by using Nyquist frequency of
12,800 Hz and frequency resolution Δf = 1.5625 Hz com-
promising with the type of carcass gear box and the kind

Vol:.(1234567890)
SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8 Research Article

Table 1  Estimated PPM CFM LSCEM ERAM RFPM

natural frequencies, and
( ) ( ) ( )
corresponding normalized fr [Hz] fr [Hz] fr [Hz] er fr % fr [Hz] er fr % fr [Hz] er fr %
random errors, er, for the 5
methods Mode 1 1709 1707 1706 0.005 1706 0.002 1705 0.001
Mode 2 2125 2124 2125 0.067 2118 0.031 2123 0.021
Mode 3 3415 3414 3340 0.089 3413 0.013 3413 0.034
Mode 4 3964 3963 3962 0.076 3964 0.008 3967 0.001
Mode 5 4343 4343 4340 0.003 4341 0.023 4340 0.004

Table 2  Estimated damping PPM CFM LSCEM ERAM RFPM

ratios, and corresponding
( ) ( ) ( )
normalized random errors, er, 𝜁r % 𝜁r % 𝜁r % er 𝜁r % 𝜁r % er 𝜁r % 𝜁r % er 𝜁r %
for the 5 methods
Mode 1 0.27 0.23 0.29 0.08 0.28 0.616 0.29 0.042
Mode 2 0.55 0.51 0.53 0.069 0.51 0.125 0.65 0.029
Mode 3 0.44 0.53 0.45 0.073 0.3 0.190 0.25 0.131
Mode 4 0.45 0.47 0.53 0.04 0.54 0.132 0.47 0.010
Mode 5 0.74 0.74 0.72 0.029 0.76 0.012 0.84 0.008

of constraint condition. An exponential window for the

force signals is used in order to decrease the leakage. The
input autopower-spectra, output power-spectra and cross-
power-spectra are estimated and reserved for each meas-
urement location. In addition, the FRFs are determined
by using Hv estimator [33]. The coherence function is con-
trolled as an on-line check of data quality. The periodic
response was calculated by using the synchronous averag-
ing method as reported in [19] with the aim of obtaining a
good estimation of the frequency response function.
In detail, 34 periodic responses were selected to have
a better estimation of the modal parameters, i.e. natural
frequencies and damping ratios. The method used to make
the EMA is the traditional procedure in which both excita-
tion and response are measured together to obtain the
accelerance. Figure 6 shows all the FRFs measured, in total Fig. 7  Stabilization diagram extracted by LSCE
34 functions are plotted.
Five methods used for the estimation of modal param-
eters are shown in this work. At the beginning, the Peak 4 Results and discussion
Amplitude and Circle Fit were the simplest methods to
be applied to calculate the first estimation as their imple- The natural frequencies and damping ratios were
mentation is easier with respect to the other methods. extracted for the first 5 modes by the methods described
Once the results were obtained by these two methods, in Sect. 2, in the following manner. In the LSCEM, ERAM
they were compared with the results of several methods and RFPM, all modes fulfilling the stabilization criteria with
that, according to literature, are considered more accurate; tolerance 1% for the natural frequencies, 5% for the damp-
these methods are the LSCEM, ERAM, and RFPM. ing ratios and the MAC [34] values greater than 90%.
A stabilization diagram was constructed for the last In the LSCEM the stable poles that have stability in
three methods to identify the physical modes from the frequency and damping in the stabilization diagram
computational modes. Normally, a tolerance, in percent- were plotted. The LSCEM stability diagram shows that
age is given for the stability of each modal parameters. the vectors are not engaged, because the non- physical

Vol.:(0123456789)
Research Article SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8

It is evident from Table 1 that the natural frequen-

cies of all modes determined by these methods were
very similar and their errors er were very low; thus, the
difference between the method results is insignificant.
Regarding the damping factor, it can be noted in Table 2
that the results were very distant, and it is difficult to
detect a good damping estimation. The comparison
between the PPM and CFM and the other methods in
Table 2 shows that their results were slightly different,
while some similarity can be noticed on several modes.
It can be deduced that although the PPM and CFM are
generally considered simple and very basic as described
in Sect. 2, they managed to give good results in compari-
son to the other methods. Therefore, these two methods
can be very useful as a first estimation because they pro-
Fig. 8  Stabilization diagram extracted by ERAM vide good results for the major part of the modes.
In this study, the error values er is an indicator of how
good a method works. When the error is low, the stability
poles appear with the increase of the number of orders. of the poles is high in the stabilization diagram, hence
Consequently, it will be hard to reach a good estimation the method can be considered reliable. In this context,
of the mode shapes. In the ERAM the stability of vector a comparison was made between RFPM, ERAM and
and the stable poles were schemed. In order to compare LSCEM in Figs. 10 and 11; it was clear that the RFP was
the mode shapes between the iteration, the Modal Assur- the method that presented the lowest values of error
ance Criterion (MAC) was used to detect the stability of which make it the best method among the three studied.
vectors. When the RFPM method is applied, the stability The only issue here is that the error obtained by RFPM
of the vector can be rarely found because as the number was the lowest on all the modes except the third mode,
of orders increases the ill conditioning problem starts to where it has a value (0.15%) greater than the error given
appear. Therefore, its plot was eliminated in the stabiliza- by the LSCEM and almost equal to the one given by the
tion diagram of the RFPM. ERAM. This error can be due to the noise present in the
The results are summarized in Tables 1 and 2, where the experimental FRF.
average value of poles and the corresponding scatter are
listed and schemed in the Figs. 10 and 11. The scatter was
calculated as the normalized random error er (standard 5 Conclusions
deviation divided by the average value) for LSCEM, ERAM
and RFPM. It is important to mention here that the mini- This paper presents a comparison between several
mum number of pole estimates of each mode was 12. methods of experimental modal analysis for the detec-
The estimation of the stable poles can be seen in tion of the frequency and damping ratio values. The
Figs. 7, 8 and 9. Figures 7 and 8 show the stabilization results achieved during the study serve to understand
diagrams extracted by LSCEM and ERAM, respectively, the factors affecting the accuracy of the frequency and
with a number of order 80. In both cases, the presence of the damping in respect to the method in use. It can be
5 physical modes can be detected in the frequency range observed that the natural frequency values obtained
[0 4500] Hz. It can be noticed that the stable poles start by all methods were very similar and the differences
to appear when the number of order surpasses 26 for the between the results were insignificant.
LSCEM (Fig. 7). In Fig. 9 all modes were isolated, and for It was revealed that it is convenient to use the Peak
each mode a stabilization diagram was established with Picking Method and Circle Fit Method as a first estima-
a number of order 20. When the frequency bandwidth is tion; these methods succeeded in giving good estima-
narrowed around the Peaks, these modes immediately tion of the natural frequency and damping. They are also
appear when the number of order exceed one, and their simple to use and fast to implement.
estimation becomes more accurate.

Vol:.(1234567890)
SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8 Research Article

(a) (b)

(c) (d)

(e)

Fig. 9  a Stabilization diagram close to the first mode; b second modes; c third mode; d fourth mode; e fifth mode

er [%] Regarding the other methods, if the stability of poles

0.089

0.076

0.1 dispersed in the stabilization diagram was taken into

0.067

0.08 consideration, the Rational Fraction Polynomial method

0.034

0.06 did the best job in detecting the damping and frequency
0.031

0.023
0.021

values. This method presented the least scatter on all

0.013

0.04
0.008
0.005

0.004
0.003
0.002
0.001

0.001

0.02 modes except on the third, probably due to the noise

0
MODE 1 MODE 2 MODE 3 MODE 4 MODE 5
present in the measured frequency response func-
LSCEM ERAM RFPM tion. The results obtained by the LSCEM and the ERAM
were reasonable for both frequency and damping. The
Fig. 10  Scatter of frequency for each mode, determined by LSCEM,
ERAM and RFPM

Vol.:(0123456789)
Research Article SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8

er [%] 9. Juang JN, Pappa RS (1985) An eigensystem realization algorithm

0.313
0.4 for modal parameter identification and model reduction. J Guid
0.35
0.3 Control Dyn 8:620–627

0.19
0.25 10. El-Kafafy M, Guillaume P, Peeters B, Marra F, Coppotelli G (2012)

0.132
0.131
0.125
0.2 Advanced frequency- domain modal analysis for dealing with

0.073
0.069
0.08

0.15 measurement noise and parameter uncertainty. In: Confer-

0.042

0.029

0.029
0.04

0.012
0.008
0.1

0.01
ence proceeding of the 30 th IMAC, a conference on structural
0.05
0 dynamics, Bethel, CT, USA, 5. Springer, pp 179–198
MODE 1 MODE 2 MODE 3 MODE 4 MODE 5 11. Richardson MH, Formenti DL (1982) Parameter estimation from
LSCEM ERAM RFPM frequency response measurements using rational fraction poly-
nomials. In: Proceedings of the 1st IMAC conference, Orlando,
Florida, November 1982, pp 167–181
Fig. 11  Scatter of damping for each mode, determined by LSCEM,
12. Brincker R, Ventura CE (2015) Introduction to operational modal
ERAM and RFPM
analysis. Wiley, London
13. Silva JMME (1978) Measurements and application of structural
mobility data for the vibration analysis of complex structures.
scatter was found slightly greater in the ERAM than in Ph.D. thesis, Imperial College of Science and Technology, Lon-
don, UK
the LSCEM. 14. Davis D, Abrams M, Brault J (2001) Fourier transform spectrom-
In addition, it has been noted by following this proce- etry, 1st edn. Academic Press, ebook 9780080506913
dure of implementation. The fast method to extract the 15. Mallat S (2008) A wavelet tour of signal processing, 3rd edn.
natural frequencies and the damping ratios was the ERAM Academic Press, ebook 9780080922027
16. Douglas L (2009) The discrete Fourier transform, part 4: spectral
next come the LSCEM then the RFPM. leakage. J Object Technol 8:7
17. Fahey SO’F, Pratt J (1998) Frequency domain modal estimation
techniques. Exp Tech 22:33–37
Compliance with ethical standards 18. Fahey SO’F, Pratt J (1998) Time domain modal estimation tech-
niques. Exp Tech 22:45–49
19. Ford R, Randall RB, Wardrop T (2002) Updating modal properties
Conflict of interest The authors declare that they have no conflict of from response-only measurements on a rail vehicle. In: Interna-
interest regarding the publication of this paper. tional conference on noise and vibration engineering, Leuven,
September 16–18
20. Ewins DJ (2000) Modal testing: theory, practice and application,
2nd edn. Wiley, New York
21. Fasana A, Marchesiello F (2006) Meccanica Delle Vibrazioni. C. L.
References U. T, Torino
22. Maia NMM (1985) Interference criteria in modal analysis identi-
1. Rieger NF (2003) The relationship between finite element analy- fication. M.Sc. Thesis, Technical University of Lisbon, Portuguese
sis and modal analysis, stress technology incorporated. Roches- 23. Robb DA, Ewins DJ, Maia NMM (1985) Modal tests on the end
ter, New York windings of an electric motor. Imperial College of Science and
2. Zhou W, Chelidze D (2008) Generalized eigenvalue decompo- Technology, London
sition in time domain modal parameter identification. J Vib 24. Prony R (1795) Essai expérimental et analytique sur les lois de
Acoust ASME 130(1):011001. https:​ //doi.org/10.1115/1.277550
​ 9 la dilatabilité des fluides de la vapeur de l’alkool, à différentes
3. Cakir F, Habib U (2015) Experimental modal analysis of brick températures. Journal de l’École Polytechnique Floréal et Plairial,
masonry arches strengthened prepreg composites. J Cult Herit 1, 22:24–76
16:284–292 25. Helyen W, Lammens S, Sas P (1998) Modal analysis theory and
4. James MWB, Alison R, James B, Alessandro A, Emma H, Peter D testing. Katholieke University, Leuven
(2018) Experimental modal analysis of British rock lighthouse. 26. Allemang RJ, Philips AW (2004) The unified matrix polynomial
Mar Struct 62:1–22 approach to understanding modal parameter estimation an
5. Diogo M, Julio MMS (2014) A contribution to the modal identi- update structure dynamic. Research laboratory, University of
fication of the damping factor based on the dissipated energy Cincinnati, Cincinnati
.In: Proceedings of the 9th international conference on struc- 27. Brillhart SR, Mikulcik EC (1988) Comparison of modal parameter
tural dynamics Portugal 30 June–2 July estimation methods for highly damped structures. In: Proceed-
6. LMS International (2005) The LMS theory and background ings of IMAC VI, Kissimmee, Florida, pp 705–711
book—analysis and design. Manual of test. Lab revision 5. 28. Juang JN, Pappa RS (1986) Effect of noise on modal parameters
LMS International, Leuven identified by eigensystem realization algorithm. J Guid Control
7. Brown DL, Allemang RJ, Zimmerman RD, Mergeay M (1979) Dyn 9(3):294–303
Parameter estimation modal techniques for modal analysis. 29. Francisco M, Marriaga ME, Pérez TE, Piñar MA (2019) Coherent
SAE technical no. 790221, 15–24 pairs of bivariate orthogonal polynomials. J Approx Theory
8. Juang JJ (1994) Applied system identification. Prentice-Hall, 245:40–63
Englewood Cliffs 30. Code of Matlab Software’s “invfreqs” https​://it.mathw​orks.com/
help/signal​ /ref/invfre​ qs.html and “residue” https:​ //it.mathwo
​ rks.
com/help/matla​b/ref/resid​ue.html function, Mathworks

Vol:.(1234567890)
SN Applied Sciences (2019) 1:781 | https://doi.org/10.1007/s42452-019-0806-8 Research Article

31. Mucchi E (2013) On the sweet spot estimation in beach ten- 34. Allemang RJ (2003) The modal assurance criterion: twenty years
nis rackets. Measurement 46(4):1399–1410. https ​ : //doi. of use and abuse. Sound Vib 37:14–23
org/10.1016/j.measu​remen​t.2012.12.014
32. Mucchi E (2012) Experimental evaluation of modal damping in Publisher’s Note Springer Nature remains neutral with regard to
automotive components with different constraint conditions. jurisdictional claims in published maps and institutional affiliations.
Mecc Int J Theor Appl Mech AIMETA 47(4):1035–1041
33. LMS Test.Lab (2004) Theory and background book. LMS Inter-
national, Dubai

Vol.:(0123456789)