Numerical Investigations on the Accuracy

of an Automated Modal Identification

Technique

Carlo Rainieri and Giovanni Fabbrocino

Abstract Systems and techniques for fast damage detection play a fundamental
role in the development of effective Structural Health Monitoring systems. Modal-
based damage detection algorithms are well-known techniques for structural health
assessment but they need reliable and accurate automated modal identification and
tracking procedures in order to be effective. In this chapter, the performance of a
recently developed algorithm for automated output-only modal parameter esti-
mation is assessed. An extensive validation of the algorithm for continuous
monitoring applications is carried out based on simulated data. Different levels of
damping are considered. The numerical study demonstrates that the algorithm
provides fairly robust, accurate and precise estimates of the modal parameters,
including damping ratios.

Keywords: Automated operational modal analysis
Damping measurement

1 Introduction

Several algorithms for automated identification [1] and tracking [2] of modal
parameters based on Operational Modal Analysis (OMA) methods have been
developed in recent years. As a consequence, modal based damage detection
techniques [3] are again gaining in popularity, even if they suffer limitations in
terms of damage localization and quantification, as well as drawbacks related to
sensitivity to measurement quality and to environmental and operational factors.

C. Rainieri (&)
G. Fabbrocino

StreGa Lab, DiBT Department, University of Molise, Via Duca degli Abruzzi 86039
Termoli, Italy
e-mail: carlo.rainieri@unimol.it
G. Fabbrocino
e-mail: giovanni.fabbrocino@unimol.it

G. Dalpiaz et al. (eds.), Advances in Condition Monitoring of Machinery 465

in Non-Stationary Operations, Lecture Notes in Mechanical Engineering,
DOI: 10.1007/978-3-642-39348-8_40,  Springer-Verlag Berlin Heidelberg 2014
466 C. Rainieri and G. Fabbrocino

Nevertheless, continuous monitoring of modal parameters has a large potential in

performance and health assessment of civil engineering structures [4]. Applica-
tions range from prompt detection of damage and degradation phenomena to post-
earthquake health assessment and emergency management.
The available solutions for automated output-only modal identification show
some drawbacks [1] which can affect their reliability and accuracy in the frame-
work of continuous monitoring:
• threshold based peak and physical pole detection;
• need of a preliminary calibration phase at each new application;
• static settings of thresholds and parameters which may be unsuitable to track the
natural changes in modal properties of structures due to damage or environ-
mental effects;
• sensitivity to noise, problems of false or missed identification.
Moreover, a number of algorithms do not provide damping estimates; whenever
they are able to estimate modal damping, the resulting values are usually very
scattered. The fairly large scatter associated to damping estimates, compared with
natural frequency and mode shape estimates, is well documented in the literature.
Even if it can be partially addressed to inherent limitations of the estimators and
the adoption of an equivalent viscous damping model [5], appropriate data pro-
cessing procedures have to be adopted in order to minimize the estimation error.
As a consequence, a careful design of the algorithm is fundamental to enhance
reliability, robustness and accuracy of automated modal identification procedures.
A thorough performance assessment of automated modal identification algo-
rithms is rarely reported in the literature but it is a fundamental step in view of
proper post-processing of modal parameters for damage detection and perfor-
mance evaluation purposes.
In the present chapter, the performance of a recently developed procedure for
fully automated output-only modal identification is analyzed. The performance of
the algorithm is assessed through the analysis of the modal identification results
obtained from four case studies, each one consisting of 10,000 simulated datasets.
The analysis of the obtained results and the comparisons with the reference values of
the modal properties show that the algorithm is characterized by a very high success
rate and that it is able to provide accurate estimates of the modal parameters,
including damping. The performance assessment of the algorithm is still in progress.
However, the herein illustrated results are definitely promising, so the algorithm has
a potential in the continuous vibration based monitoring of civil structures.

2 The Automated OMA Algorithm

The proposed algorithm (Fig. 1) is based on the combination of different OMA

techniques in order to faclitate the analysis and interpretation of the stabilization
diagram. The method is based on Stochastic Subspace Identification (SSI) [6] and
Numerical Investigations 467

Fig. 1 Flowchart of the

automated OMA algorithm START

DATA LOADING

JAD (SOBI)

SSI OF THE SOURCES

CLUSTERING
(physical and noise modes)

MODE SHAPES

TRACKING/REPORT

FINISH

END

the selection of physical poles in the stabilization diagram by clustering tech-

niques, but it takes advantage of the Blind Source Separation (BSS) [7], operated
at a preliminary stage according to the Second Order Blind Identification (SOBI)
[8, 9]) procedure, in order to simplify the interpretation of the stabilization
diagram.
In fact, as a result of the BSS phase, the raw data associated to the measured
structural response are transformed into sources which can be well-separated (they
show the contribution of a single mode to the structural response), not well-
separated (noise or minor contributions from other modes could be superimposed
to the contribution of the main mode) or just noise sources [9]. Thus, the BSS
simplifies the analysis of the data and the interpretation of the stabilization dia-
gram, allowing the extraction of the modal information from the individual sources
and not the multivariate time series of raw data. The interpretation of the stabil-
ization diagram, therefore, becomes easier since it basically reports information
about only one mode at the time. The sources are analyzed one-by-one according
to the SSI method and the physical poles are separated from the spurious ones by
means of clustering techniques [10] and mode validation criteria. The final values
of the natural frequency and damping ratio for the identified modes are obtained by
a sensitivity analysis with respect to the number of block rows in SSI, for a fixed
value of the maximum model order in the stabilization diagram. The cluster
characterized by the minimum variance of the estimates when i ranges in a certain
468 C. Rainieri and G. Fabbrocino

Fig. 2 The benchmark 4-DOF system

interval with a certain step Di is then selected as the one providing the best
estimate of the modal parameters for a given structural mode [11]. Mode shape
estimates are finally obtained from Singular Value Decomposition (SVD) of the
output Power Spectral Density (PSD) matrix at the previously estimated frequency
of the mode [12].
The key feature of the algorithm is the absence of any analysis parameters to be
tuned at each new monitoring application. In fact, it has been shown [13] that the
results are insensitive to the setting of the two parameters governing the Joint
Approximate Diagonalization (JAD) [14] for source extraction, while the influence
of the number of block rows [6] is taken into account by sensitivity analyses.
Moreover, the parameters governing the JAD play a primary role in the control of
response time and computational burden [13]. This is relevant, in particular, for
SHM applications in seismically prone areas [2].

3 Case Studies and Analysis of Results

The performance of the proposed algorithm in terms of accuracy and reliability of

estimates has been investigated through a statistical analysis of the results obtained
from simulated data continuously generated through the excitation of a 4-DOF
system by a Gaussian white noise.
The mass and stiffness properties of the system are reported in Fig. 2. Rayleigh
damping is adopted. The modal properties of the system are reported in Table 1.
The four case studies differ for the assumed values of damping or the signal-to-
noise ratio (SNR).
The system matrices and, therefore, the associated modal parameters are the
same in all runs, since only the uncertainties associated to inherent limitations of the
estimator are the key focus of the investigation. The performance of the method in
the presence of uncertain system matrices, that is when modal parameters can
slightly change at each run as an effect of the deviation of the system matrices from
their nominal values, is out of the scope of the present paper. Specific numerical
analyses are under development and will be discussed elsewhere.
The system response to Gaussian white noise N(0, 1) has been simulated 10,000
times. The input has been applied at DOF #1. Each simulated dataset consisted of
Numerical Investigations 469

Table 1 Modal properties of the simulated 4-DOF system

Mode Natural frequency Damping ratio [%]
# [Hz]
Case study #1 Case study #2 Case study #3 Case study #4
(SNR = 5 dB) (SNR = 15 dB) (SNR = 5 dB) (SNR = 5 dB)
I 0.668 1.00 1.00 2.00 2.00
II 1.137 0.88 0.88 1.31 1.76
III 1.526 0.92 0.92 1.09 1.84
IV 1.879 1.00 1.00 1.00 2.00

Table 2 Success rate of automated modal identification over 10,000 runs

Mode # Success rate [%]
Case study #1 Case study #2 Case study #3 Case study #4
(SNR = 5 dB) (SNR = 15 dB) (SNR = 5 dB) (SNR = 5 dB)
I 99.79 100.0 99.71 99.87
II 99.96 100.0 99.97 100.0
III 99.95 100.0 99.98 100.0
IV 100.0 100.0 100.0 100.0

four measurement channels; the total record length was 3,600 s and the sampling
frequency was 10 Hz. Gaussian white noise has been added to the system response
in order to simulate the effect of measurement noise. The SNRs are reported in
Table 1. Each dataset has been then processed using the described algorithm in
order to automatically extract the modal parameters of the system. The analysis of
the simulated datasets has been carried out considering a number of block rows
i ranging between 20 and 80 with Di = 2 and considering a maximum model order
of 16 in the construction of the stabilization diagram for each analyzed source.
The analysis of the obtained results has pointed out that the algorithm carries
out automated output-only modal identification in a very robust way. In fact, a
success rate [15] larger than 99 % has been obtained for all modes (Table 2). Just
in a few runs missed identification of the modal properties of a mode occurred.
This was probably due to a combined effect of weak excitation and low SNR,
which affected the quality of separation and stabilization.
The results in terms of natural frequency and damping estimates are summa-
rized in Tables 3 and 4. Very accurate natural frequency estimates, associated also
to a low standard deviation r, have been obtained. The error in natural frequency
estimates is much lower than 1 % in the 95 % of the runs for all case studies. The
accuracy of estimates slightly improves when the SNR increases. The analysis of
the coefficient of variation c associated to the natural frequency estimates in the
cluster selected by the sensitivity analysis with respect to the number of block rows
in each run points out that the estimates are also very precise (cf, cluster much lower
than 0.1 % in the 95 % of the runs). When damping increases, the error associated
to the estimates also increases, but the results are still very accurate.
470 C. Rainieri and G. Fabbrocino

Table 3 Automated modal identification results: natural frequencies

Case Mode fnominal fav rf [Hz] Df [%] 50th Df [%] 95th cf,cluster [%] 95th
study # [Hz] [Hz] centile centile centile
1 I 0.668 0.668 0.000730 0.072 0.216 0.0254
1 II 1.137 1.137 0.000911 0.052 0.156 0.0285
1 III 1.526 1.526 0.001157 0.047 0.143 0.0269
1 IV 1.879 1.879 0.001568 0.048 0.151 0.0363
2 I 0.668 0.668 0.000665 0.067 0.196 0.0319
2 II 1.137 1.137 0.000879 0.050 0.152 0.0234
2 III 1.526 1.526 0.001090 0.047 0.140 0.0232
2 IV 1.879 1.879 0.001425 0.047 0.149 0.0296
3 I 0.668 0.668 0.001275 0.121 0.359 0.0598
3 II 1.137 1.137 0.001324 0.070 0.219 0.0509
3 III 1.526 1.526 0.001309 0.054 0.166 0.0345
3 IV 1.879 1.879 0.001469 0.048 0.148 0.0346
4 I 0.668 0.668 0.001368 0.117 0.358 0.0729
4 II 1.137 1.137 0.001636 0.086 0.278 0.0814
4 III 1.526 1.526 0.002054 0.083 0.260 0.0798
4 IV 1.879 1.879 0.002825 0.085 0.277 0.0991

Table 4 Automated modal identification results: damping ratios

Case Mode nnominal nav and (mode, rn Dn [%] 50th Dn [%] 95th cn,cluster [%]
study # [%] median) [%] [%] centile centile 95th centile
1 I 1.00 1.02 (1.0, 1.02) 0.108 7.2 21.6 7.0
1 II 0.88 0.89 (0.9, 0.89) 0.080 6.0 18.0 4.4
1 III 0.92 0.93 (0.9, 0.93) 0.072 5.1 15.3 2.5
1 IV 1.00 1.01 (1.0, 1.01) 0.075 4.9 14.9 3.1
2 I 1.00 1.01 (1.0, 1.01) 0.099 6.7 19.4 2.2
2 II 0.88 0.88 (0.9, 0.88) 0.075 5.7 16.6 1.6
2 III 0.92 0.92 (0.9, 0.92) 0.073 5.3 15.5 1.7
2 IV 1.00 1.00 (1.0, 1.00) 0.074 4.7 14.6 2.1
3 I 2.00 2.03 (2.0, 2.03) 0.179 5.9 17.9 7.7
3 II 1.31 1.32 (1.3, 1.32) 0.109 5.4 16.7 5.1
3 III 1.09 1.10 (1.1, 1.10) 0.081 4.9 14.7 2.4
3 IV 1.00 1.01 (1.0, 1.00) 0.075 4.9 14.9 2.8
4 I 2.00 2.03 (2.0, 2.02) 0.183 6.0 18.3 6.7
4 II 1.76 1.78 (1.8, 1.78) 0.138 5.2 15.6 4.6
4 III 1.84 1.85 (1.8, 1.85) 0.129 4.5 13.9 3.0
4 IV 2.00 2.02 (2.0, 2.02) 0.143 4.3 14.0 3.6

Damping estimates are fairly accurate and characterized by moderate uncer-

tainty (r lower than 0.2 %). In particular, the variability of estimates slightly
increases when the nominal damping values increase. Larger errors are associated
to damping estimates with respect to natural frequencies. However, the scatter with
respect to the nominal values is lower than 10 and 20 % in the 50 and 95 % of the
runs, respectively. The errors slightly decrease when the SNR increases. Damping
Numerical Investigations 471

estimates are also fairly precise (cn, cluster much lower than 10 % in the 95 % of the
runs). The mean, mode and median of the identified damping ratios after 10,000
runs are very close each other and to the nominal values of modal damping ratios.
Taking into account the uncertainty associated to damping estimates, the mode of
damping values is given with one decimal place only.
Marginal refinements can be obtained by outlier removal, confirming the
robustness and accuracy of the algorithm.

4 Conclusions

A hybrid automated OMA algorithm has been described in the present paper. Its
performance has been assessed against simulated data generated by a 4-DOF for
different damping levels and SNRs. The system was excited by a Gaussian white
noise. The results obtained from 10,000 runs have been analyzed and robustness,
accuracy and precision of the algorithm have been checked. Encouraging results
have been obtained, in particular as the possibility to estimate damping ratios in an
accurate and fully automated way is concerned. Further investigations are in
progress to assess the performance of the algorithm in the case of uncertain system
matrices in view of continuous, long term vibration based SHM applications.

Acknowledgments The present work is carried out in the framework of AT2—LR 2—Task 3 of
the ReLUIS-DPC Project 2010-2013, whose support is gratefully acknowledged.

References

1. Rainieri C, Fabbrocino G (2010) Automated output-only dynamic identification of civil

engineering structures. Mech Syst Signal Process 24:678–695
2. Rainieri C, Fabbrocino G, Cosenza E (2011) Near real-time tracking of dynamic properties
for standalone structural health monitoring systems. Mech Syst Signal Process 25:3010–3026
3. Doebling SW, Farrar CR, Prime MB, Shevitz DW (1996) Damage identification and health
monitoring of structural and mechanical systems from changes in their vibration
characteristics: a literature review. Technical Report LA-13070-MS, UC-900, Los Alamos
National Laboratory, New Mexico 87545
4. Rainieri C, Fabbrocino G, Cosenza E (2010a) Integrated seismic early warning and structural
health monitoring of critical civil infrastructures in seismically prone areas. Struct health
monit—Int J 10:291–308
5. Rainieri C, Fabbrocino G, Cosenza E (2010b) Some remarks on experimental estimation of
damping for seismic design of civil constructions. Shock Vib 17:383–395
6. Van Overschee P, De Moor B (1996) Subspace identification for linear systems: theory—
implementation—applications. Kluwer Academic Publishers, Dordrecht
7. Ans B, Hérault J, Jutten C (1985) Adaptive neural architectures: detection of primitives.
COGNITIVA’85:593-597
8. Belouchrani A, Abed-Meraim K, Cardoso JF, Moulines E (1997) A blind source separation
technique using second-order statistics. IEEE Trans Signal Process 45:434–444
472 C. Rainieri and G. Fabbrocino

9. Poncelet F, Kerschen G, Golinval JC, Verhelst D (2007) Output-only modal analysis using
blind source separation techniques. Mech Syst Signal Process 21:2335–2358
10. Tan P-N, Steinbach M, Kumar V (2006) Introduction to data mining. Pearson Addison-
Wesley, Reading, MA
11. Rainieri C, Fabbrocino G, Cosenza E (2010c) On damping experimental estimation. In:
Proceedings of 10th international conference on computational structures technology,
Valencia
12. Brincker R, Zhang L, Andersen P (2000) Modal identification from ambient responses using
frequency domain decomposition. In: Proceedings of 18th SEM international modal analysis
conference, San Antonio
13. Rainieri C, Fabbrocino G (2012) A hybrid automated modal identification algorithm for
Structural Health Monitoring: a comparative assessment. In: Proceedings of international
conference on noise and vibration engineering ISMA2012, Leuven
14. Cardoso JF, Souloumiac A (1996) Jacobi angles for simultaneous diagonalization. SIAM J
Matrix Anal Appl 17:161–164
15. Magalhaes F, Cunha A, Caetano E (2012) Online automatic identification of the modal
parameters of a long span arch bridge. Mech Syst Signal Process 23:316–329