Available online at www.sciencedirect.

com
Available
Available online
online at www.sciencedirect.com
at www.sciencedirect.com

ScienceDirect
Procedia
Procedia Engineering
Engineering 00 (2017)
199 (2017) 000–000
1002–1007
Procedia Engineering 00 (2017) 000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

X International Conference on Structural Dynamics, EURODYN 2017

X International Conference on Structural Dynamics, EURODYN 2017
Output
Output only
only system
system identification
identification based
based on
on synchrosqueezed
synchrosqueezed
transform
transform
a b,∗
Smita
Smita Kaloni
Kalonia ,, Manish
Manish Shrikhande
Shrikhandeb,∗
a National Institute of Technology Uttarakhand, Srinagar (Garhwal)–246174, India
a
b DepartmentNational Institute
of Earthquake of Technology
Engineering, Uttarakhand,
Indian Institute ofSrinagar (Garhwal)–246174,
Technology India
Roorkee, Roorkee–247667, India
b Department of Earthquake Engineering, Indian Institute of Technology Roorkee, Roorkee–247667, India

Abstract
Abstract
A new blind source separation formulation for modal identification has been developed based on synchrosqueezed transform.
A new blind source separation formulation for modal identification has been developed based on synchrosqueezed transform.
It is shown that by using acceleration measurements there is a greater chance of identifying higher modes than with displace-
It is shown that by using acceleration measurements there is a greater chance of identifying higher modes than with displace-
ment/velocity measurements. An elegant and robust procedure for estimation of mode shapes and modal damping is presented
ment/velocity measurements. An elegant and robust procedure for estimation of mode shapes and modal damping is presented
which avoids tedious computations required for other blind source separation formulations, such as the independent component
which avoids tedious computations required for other blind source separation formulations, such as the independent component
analysis and second order blind identification. A numerical example study based on the recorded earthquake response of the UCLA
analysis and second order blind identification. A numerical example study based on the recorded earthquake response of the UCLA
Factor building has been used to validate the proposed formulation.
Factor
c 2017building has been used to validate the proposed formulation.
 The Authors. Published by Elsevier Ltd.
©c 2017
2017 The
TheAuthors.
Authors.Published
Published byElsevier
ElsevierLtd.
Ltd.
Peer-review under responsibilityby

of the organizing committee of EURODYN 2017.
Peer-review under responsibility of the organizing committee of EURODYN
EURODYN 2017.
2017.
Keywords: Blind source separation; Output only system identification; Synchrosqueezed transform; Time-frequency analysis
Keywords: Blind source separation; Output only system identification; Synchrosqueezed transform; Time-frequency analysis

1. Introduction
1. Introduction
System identification from the analyses of vibration signatures of structural systems holds a lot of promise as a
System identification from the analyses of vibration signatures of structural systems holds a lot of promise as a
diagnostic tool in maintenance of infrastructure and lifeline facilities. It is an inverse problem to ascertain the system
diagnostic tool in maintenance of infrastructure and lifeline facilities. It is an inverse problem to ascertain the system
parameters from a set of system input-output, or only output data [13]. Considering the difficulty of measuring input
parameters from a set of system input-output, or only output data [13]. Considering the difficulty of measuring input
data in several instances, various output-only identification techniques have been developed [10,11,16]. A new class
data in several instances, various output-only identification techniques have been developed [10,11,16]. A new class
of output-only identification technique, namely, blind source separation (BSS) has been used for modal identification
of output-only identification technique, namely, blind source separation (BSS) has been used for modal identification
in recent times [12] The BSS technique involves separation of a set of independent signals from their mixture w,ithout
in recent times [12] The BSS technique involves separation of a set of independent signals from their mixture w,ithout
any prior information about the mixing process. The separation of independent sources from the observed mixtures is
any prior information about the mixing process. The separation of independent sources from the observed mixtures is
achieved by using some statistical properties of the underlying process with techniques like independent component
achieved by using some statistical properties of the underlying process with techniques like independent component
analysis (ICA) [5,9], second order blind identification (SOBI) [2] and complexity pursuit (CP) [1]. These procedures
analysis (ICA) [5,9], second order blind identification (SOBI) [2] and complexity pursuit (CP) [1]. These procedures
are based on certain idealizations about the statistical properties of the signals which are not always consistent with the
are based on certain idealizations about the statistical properties of the signals which are not always consistent with the
practical cases and the results are often found to be sensitive to the choice of algorithmic parameters. The independent
practical cases and the results are often found to be sensitive to the choice of algorithmic parameters. The independent
sources can be identified only within a scaling factor of true modal contributions thereby necessitating specialized
sources can be identified only within a scaling factor of true modal contributions thereby necessitating specialized
techniques for the estimation of mode shapes. In view of these difficulties with the prevalent BSS formulations, we
techniques for the estimation of mode shapes. In view of these difficulties with the prevalent BSS formulations, we

∗ Corresponding author. Tel.: +91-1332-28-5437 ; fax: +91-1332-27-3560.

∗ Corresponding
E-mail address:author. Tel.: +91-1332-28-5437 ; fax: +91-1332-27-3560.
mshrifeq@iitr.ac.in
E-mail address: mshrifeq@iitr.ac.in
1877-7058 c 2017 The Authors. Published by Elsevier Ltd.
1877-7058 c 2017 The Authors. Published by Elsevier Ltd.
Peer-reviewunder responsibility of the organizing committee of EURODYN 2017.
1877-7058 © 2017responsibility
Peer-review under The Authors. Published
of the by committee
organizing Elsevier Ltd.
of EURODYN 2017.
Peer-review under responsibility of the organizing committee of EURODYN 2017.
10.1016/j.proeng.2017.09.235
 Smita Kaloni et al. / Procedia Engineering 199 (2017) 1002–1007 1003
2 Smita Kaloni and Manish Shrikhande / Procedia Engineering 00 (2017) 000–000

explore the use of a wavelet based synchrosqueezed transform (SST) [3,7,15] for separation of independent harmonic
components from the recorded structural response data for subsequent identification of modal parameters, namely,
natural frequencies, mode shapes and modal damping.

2. Blind Source Separation and Modal Identification

The blind source separation (BSS) problem can be stated as:

x(t) = As(t) + n(t) (1)

where, x(t) is the vector of observed output, s(t) is the vector of source signals, A is the mixing matrix, and n(t)
represents the measurement noise. For a comparison, the linear dynamic response of a structural system can be given

by a linear combination of the normal modes of the system: x = Φq(= ni=1 φ i qi ) with q as modal coordinates,
and Φ as the mode shape matrix. Kerschen et al. proposed that modal coordinates can be viewed as independent
sources, if the ratios of system natural frequencies are not integers [12]. Thus the modal expansion of system response
can be seen as being analogous to the BSS statement Eq. (1) with modal coordinates q being proportional to virtual
sources s regardless of the number and type of the physical excitation force and, the mixing matrix A is similar to
the mode shape matrix Φ. The modal parameters (natural frequencies and modal damping) are identified from the
post-processing of the identified source time histories, and the mode shapes are provided by the columns of the mixing
matrix.
For a base excited structure, the equations of motion are given in terms of the relative displacement of the mass
with respect to the base:
Mẍ + Cẋ + Kx = −MRẍg (2)
where ẍg represents the ground acceleration, R is matrix of rigid body influence coefficients, and x denotes the dis-
placement of mass-points relative to the base. The solution by mode superposition proceeds with the modal expansion.
However, the relative motion can not be measured directly and the inertial accelerometers used for building instrumen-
tation pickup total acceleration of the point [4]. The governing equation of motion in terms of the total accelerations
can be given by:
ẍt = (ẍ + Rẍg ) = −M−1 (Cẋ + Kx) = −M−1 (CΦq̇ + KΦq) (3)
Since the damping term is generally negligible in comparison to the restoring force term, we get an approximate
relation for total acceleration as:
ẍt ≈ −M−1 KΦq = −ΦΛq (4)
where, M−1 K = ΦΛΦ−1 has been substituted from generalized eigenvalue problem of structural dynamics, and Λ is
the diagonal matrix containing eigenvalues. This is now in a form suitable for using BSS with total acceleration mea-
surements, ẍt . The weighting of modal coordinates by the respective eigenvalues enhances the energy of higher mode
components which helps in the identification of higher modes as well—an important factor for structural health moni-
toring applications. We now explore the possibility of using a new wavelet-based technique, namely, synchrosqueezed
transform (SST) for extracting the sources from the recorded total accelerations in an earthquake excited building.

3. Synchrosqueezed Transform

Synchrosqueezing technique was initially introduced in the context of audio signal and further analysed as an
alternative to empirical mode decomposition algorithm [7]. The synchrosqueezed transform (SST) is a time-frequency
analysis tool that can identify different harmonic components in a time series comprising of several harmonics. The

time series y(t) may be given in terms of constituent parts as: y(t) = m i=1 si (t) + n(t) with each source component
(si (t)) being approximately harmonic, like si (t) = Ai (t) cos θi (t), and m is the number of components in a signal, Ai
and θi (t) are the instantaneous amplitude and (slowly varying) phase of the ith component respectively, and n(t) is the
associated measurement noise. Synchrosqueezed transform for estimation of individual source components is a three
step process: (i) estimation of continuous wavelet transform coefficients, (ii) synchrosqueezing of the CWT coefficient
followed by (iii) reconstruction of individual components.
1004 Smita Kaloni et al. / Procedia Engineering 199 (2017) 1002–1007
Smita Kaloni and Manish Shrikhande / Procedia Engineering 00 (2017) 000–000 3

3.1. Continuous Wavelet Transform

  
The continuous wavelet transform (CWT) coefficient Wy (a, b) can be estimated as: Wy (a, b) = √1a y(t)ψ∗ t−b
a dt
where, ψ∗ (t) is the complex conjugate of the mother wavelet ψ(t) [6]. The parameters a and, b control the scale and
time-shift of the transform process. The continuous wavelet transform (CWT) provides the representation of a signal
in the time-scale plane from which instantaneous frequencies can be estimated [7]. The instantaneous frequency
ω(a, b) of the signal y for any point in time-scale plane (a, b) for which Wy (a, b)  0 is given by:

 −i ∂[Wy (a,b)]
 Wy (a,b) ∂b ; |Wy (a, b)|  0

ω(a, b) = 
 (5)
∞; |Wy (a, b)| = 0

3.2. Synchrosqueezing of CWT Coefficients

The mapping of CWT coefficients from time-scale plane to the time-frequency plane, i.e., (b, a) → (b, ω(a, b))
defines the synchrosqueezing operation. Let us assume that the CWT coefficient Wy (a, b) is evaluated at discrete
values of the scale parameter, ak , with (∆a)k = ak − ak−1 and its synchrosqueezed transform T y (ω, b) is determined at
the centre ωl of the frequency interval [ωl − ∆ω
2 , ωl + 2 ], with ωl − ωl−1 = ∆ω, by adding different contributions [7]:
∆ω

1 
T y (ωl , b) = Wy (ak , b)ak −3/2 (∆a)k (6)
∆ω
ak :|ω(ak ,b)−ωl |≤ ∆ω
2

This facilitates the refinement of the of signal representation in time-frequency plane around frequency, ωl .

3.3. Signal Reconstruction from SST

The original signal can be reconstructed from its synchrosqueezed transform as:
 
 1  
y(t) =   T y (ωl , t) (7)
Cψ l

where, (·) denotes the real part of a complex quantity, and the normalizing constant Cψ is related to the mother
∞ ∗
wavelet ψ(t) as: Cψ = 12 0 ψ̂ ξ(ξ) dξ with ψ̂(ξ) being the Fourier transform of ψ(t) and the superscript (·)∗ indicating
complex conjugate. It is possible to extract individual harmonics in the signal by using ideal bandpass filter with
unit gain in the passband to extract the relevant information from the time-frequency plane which is then transformed
to time-domain by using Eq. (7). The reallocation and reconstruction of individual components through the SST is
ideally suited for use as a source separation technique in BSS and offers a viable alternative to other techniques like
ICA, SOBI or, CP.
Starting from the BSS statement in structural dynamics, we consider decomposing the measured response into
constituent parts by using SST. These individual parts are then processed to estimate the modal frequencies, damping
and the mode shapes. Let us consider the response of a generic n-degree of freedom system and its constituent parts
obtained from SST:
xk (t) = sk1 (t) + sk2 (t) + . . . + skm (t) (8)
where, xk (t) represents the response of kth degree of freedom, and ski (t) denotes the ith SST identified component of
the kth signal. From the modal expansion theorem (with mode truncation), we have:

xk (t) = φk1 q1 (t) + φk2 q2 (t) + . . . + φkm qm (t) (9)

where, we assume that only m modes contribute significantly to the system response. Comparing Eqs. (8 and 9) and
assuming the mode shapes to be normalized to unity at nth degree of freedom, it follows that the identified constituent
sources ski (t) are proportional to the modal coordinates qi (t), i = 1, 2, . . . , m. The natural frequency and modal
damping can be estimated by post-processing of the identified modal response. The mode shape coefficients at other
 Smita Kaloni et al. / Procedia Engineering 199 (2017) 1002–1007 1005
4 Smita Kaloni and Manish Shrikhande / Procedia Engineering 00 (2017) 000–000

degrees of freedom are estimated as time-domain average of the instantaneous ratios of the SST identified modal
coordinates:  
ski (t)
φki = mean (10)
t sni (t)
For the case of base excitation and acceleration measurements, the operating governing equation is Eq. (4) and the
identified constituent components correspond to modal responses weighted by respective eigenvalues, ω2i , and lead to
a comparatively better representation of higher modes than displacement/velocity measurements.
We consider the SST-based modal identification of the UCLA Doris and Louis Factor building (UCLAFB) for
validation of the proposed procedure. This building is one of the most densely permanently instrumented buildings
in North America and has provided a wealth of strong motion recordings over a decade for system identification
and health monitoring studies [14] For this validation exercise, we consider the recorded building response during
the 28 September 2004 Parkfield, CA earthquake (Mw ∼ 6.0) which had epicenter located about 260 km from the
UCLAFB. The recorded peak acceleration at the roof level was 0.0025 g in the EW direction. The same set of data
has been used in several other studies providing a good pool of independent results for comparison. The building
response has also been computed analytically via a finite element model. A good agreement between the two sets of
time histories was obtained thereby affirming the fidelity of the finite element model. The modal identification of the
UCLAFB was carried out using SOBI and SST formulations for blind source separation. Morlet wavelet was chosen
as the mother wavelet for the estimation of individual source components via synchrosqueezed transformation. Two
of the individual sources, normalized by their respective peak values, identified by SOBI and SST procedures are
shown in Fig. 1 along with the computational estimate of corresponding modal acceleration (normalized with respect
to peak value). The agreement between identified BSS sources and the modal acceleration from finite element model

(a) First Source (b) Second Source

Fig. 1: Comparison between the computed modal acceleration and identified sources by using SOBI and SST

is reasonably good considering that the SST- and SOBI-based procedures are designed to selectively extract signal in
narrow frequency bands. The natural frequencies are estimated by the peak-picking method using Fourier transform
of the identified sources. The estimated natural frequencies of first 15 modes, considering the measured building
response in EW and NS directions and the torsion component, are tabulated in Table 1. The free vibration results of
the finite element model are tabulated as Computed, while the SST-based and SOBI-based modal identification results
are tabulated under the columns SST and SOBI, respectively. The WMCC results are adapted from Hazra et al. for
wavelet based modified cross correlation method using the same data set [8]. While the estimated natural frequencies
by different BSS procedures are comparable to each other, those are consistently higher than the frequencies predicted
by the finite element model. This discrepancy might be due to limitations in the modeling of skeletal structural frame
while ignoring the stiffness contributions of non-structural components. The mode shape vectors were obtained by
taking ratios of the sources identified from different floor levels as described in the preceding section. The consistency
1006 Smita Kaloni et al. / Procedia Engineering 199 (2017) 1002–1007
Smita Kaloni and Manish Shrikhande / Procedia Engineering 00 (2017) 000–000 5

Table 1: Natural frequencies (in Hz) of UCLA Factor building

Mode Direction Computed SOBI WMCC SST

1 EW 0.39 0.45 0.47 0.46
2 NS 0.40 0.54 0.51 0.54
3 Torsion 0.56 0.65 0.68 0.70
4 EW 1.16 1.49 1.5 1.50
5 NS 1.22 1.68 1.70 1.71
6 Torsion 1.64 2.37 2.37 2.38
7 EW 1.95 2.67 2.67 2.67
8 NS 2.13 2.85 2.88 2.85
9 Torsion 2.74 3.82 3.83 3.78

of the identified mode shapes is checked by computing the modal assurance criterion (MAC) for the first nine modes
of finite element model with those estimated by the SST-based modal identification and also between SST and SOBI
estimates and are found to be in the range [1.0–0.80]. Large MAC values confirm the consistency of estimated mode
shapes from the vibration data. The mode shapes identified by SST- and SOBI-based procedures are in very good
agreement and reinforce the validity of the simple mode shape estimation procedure reported in this study.
For typical earthquake excitation, the ground motion amplitudes in the tail portion are very small, and therefore, the
system response shall be primarily dominated by the free vibration component. It is therefore appropriate to extract
a small stretch of tail portion of the identified sources as representative of the free vibration response and estimate
the modal damping by finding a suitable decay exponent of the envelope. We explore the possibility of fitting an
exponential decay function of the form e(t) = ρe−2ζn ωn t to the tail data y(t) = x2 (t), where, x(t) denotes the source
identified by SST procedure, ζn is the modal damping, ωn is the natural frequency (in rad/s) for the respective mode,
and ρ denotes the scaling factor of the envelope. Figure 2 shows the damped exponential fit for the extracted tail data
(normalized with respect to peak value) for the first two modes identified by the SST procedure. A similar damped
exponent fit for the tail data has also been obtained for sources identified via SOBI procedure. The modal damping

1.4 1.2
Mode 1 Mode 2
1.2 Natural frequency = 0.46 Hz 1.0 Natural frequency = 0.54 Hz
Normalized Amplitude

Normalized Amplitude

Damping = 5% Damping = 4.8%

1.0
0.8
0.8
0.6
0.6
0.4
0.4

0.2 0.2

0.0 0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0
Time (sec) Time (sec)

Fig. 2: Fitting of SST-based identified tail portion of sources with damped free vibration envelope

estimates for sources identified by SST-based formulation are compared with those for SOBI and WMCC schemes
(as reported in [8]) in Table 2. The estimates from different procedures are generally consistent with each other for
the first nine modes and estimated damping ratios are range bound. Considering that the modal damping is always an
elusive parameter to estimate, these estimates are used more as an indicator of the reliability of the modal parameter
identification for that mode. For higher modes, the estimates of modal damping are generally much higher indicating
the diminishing reliability of the identified modal parameters from identified sources with very small amplitudes.

4. Conclusions

A new technique for robust implementation of modal identification via blind source separation has been proposed
for base excited structures. The use of acceleration response measurement allows for greater fidelity in identification
 Smita Kaloni et al. / Procedia Engineering 199 (2017) 1002–1007 1007
6 Smita Kaloni and Manish Shrikhande / Procedia Engineering 00 (2017) 000–000

Table 2: Modal damping (%)

Mode Direction SOBI WMCC SST

1 EW 4.0 3.4 5.0
2 NS 4.5 3.2 4.8
3 Torsion 4.8 3.4 4.3
4 EW 2.0 3.3 4.0
5 NS 3.2 3.4 3.0
6 Torsion 2.2 4.3 2.0
7 EW 2.5 2.8 3.5
8 NS 3.8 3.3 4.5
9 Torsion 2.0 3.0 2.0

of higher modes due to enhanced participation. The use of synchrosqueezed transform allows for easy extraction of in-
dependent harmonic sources that are true to their amplitudes and the corresponding mode shape vector can be obtained
by the ratio of identified harmonic sources from measured data at different floor levels with the one corresponding to
the roof level. A simple scheme for estimation of modal damping for earthquake excited systems is proposed based
on free-vibration idealization of the tail portion of the response. The proposed scheme has been validated through the
identification of modal parameters for UCLA Factor building and the results are found to be in good agreement with
those obtained from computational model and from other implementations of BSS modal identification. The proposed
procedure is found to be robust and efficient in implementation for BSS modal identification of structural systems.

References

[1] J. Antoni, R. Castiglione, and L. Garibaldi. Interpretation and generalization of complexity pursuit for the blind separation of modal contribu-
tions. Mechanical Systems and Signal Processing, 85:773–788, 2017.
[2] A. Belouchrani, K. Abed-Meraim, J.-F. Cardoso, and E. Moulines. A blind source separation technique using second-order statistics. IEEE
Transactions on Signal Processing, 45(2):434–444, 1997.
[3] E. Brevdo, N.S. Fuckar, G. Thakur, and H.-T. Wu. The synchrosqueezing algorithm: a robust analysis tool for signals with time-varying
spectrum. Arxiv preprint, 2011.
[4] M. Çelebi. Seismic instrumentation of buildings (with emphasis on federal buildings). Technical Report 0-7460-68170, United States Geolog-
ical Survey, Menlo Park, California, U.S.A., 2002.
[5] P. Comon. Independent component analysis, a new concept? Signal Processing, 36(3):287–314, 1994.
[6] I. Daubechies. Ten Lectures on Wavelets, volume 61. Society for Industrial and Applied Mathematics, 1992.
[7] I. Daubechies, J. Lu, and H.T. Wu. Synchrosqueezed wavelet transforms: An empirical mode decomposition-like tool. Applied and Computa-
tional Harmonic Analysis, 30(2):243–261, 2011.
[8] B. Hazra and S. Narasimhan. Wavelet-based blind identification of the ucla factor building using ambient and earthquake responses. Smart
Materials and Structures, 19(2):025005, 2009.
[9] A. Hyvärinen and E. Oja. Independent component analysis: algorithms and applications. Neural Networks, 13(4-5):411–430, 2000.
[10] S.R. Ibrahim. Random decrement technique for modal identification of structures. Journal of Spacecraft and Rockets, 14(11):696–700, 1977.
[11] J.-N. Juang and R.S. Pappa. An eigensystem realization algorithm for modal parameter identification and model reduction. Journal of Guidance,
Control, and Dynamics, 8(5):620–627, 1985.
[12] G. Kerschen, F. Poncelet, and J.C. Golinval. Physical interpretation of independent component analysis in structural dynamics. Mechanical
Systems and Signal Processing, 21(4):1561–1575, 2007.
[13] M. Shrikhande. Finite Element Method and Computational Structural Dynamics. PHI Learning Pvt. Ltd., New Delhi, 2014.
[14] D. Skolnik, Y. Lei, E. Yu, and J.W. Wallace. Identification, model updating, and response prediction of an instrumented 15-story steel-frame
building. Earthquake Spectra, 22(3):781–802, 2006.
[15] G. Thakur, E. Brevdo, N.S. Fučkar, and H.T. Wu. The synchrosqueezing algorithm for time-varying spectral analysis: Robustness properties
and new paleoclimate applications. Signal Processing, 93(5):1079–1094, 2013.
[16] P. Van Overschee and B.L. De Moor. Subspace Identification for Linear Systems: Theory–Implementation–Applications. Springer Science &
Business Media, Boston, U.S.A., 1996.