European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Contents lists available at SciVerse ScienceDirect

European Journal of Control

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

Structural monitoring of a tower by means of MEMS-based sensing and

enhanced autoregressive models
Roberto Guidorzi a,n, Roberto Diversi a, Loris Vincenzi b, Claudio Mazzotti c, Vittorio Simioli d
a
DEI, University of Bologna, viale Risorgimento 2, 40136 Bologna, Italy
b
DIMEC, University of Modena e Reggio Emilia, via Vignolese 905, 41125 Modena, Italy
c
DICAM, University of Bologna, viale Risorgimento 2, 40136 Bologna, Italy
d
TELECO SpA, Via E. Majorana 49, 48022 Lugo (RA), Italy

art ic l e i nf o a b s t r a c t

Article history: Structural Health Monitoring (SHM) methodologies are taking advantage of the development of new
Received 9 March 2012 families of MEMS sensors and of the available network technologies. Advanced systems rely on
Accepted 24 June 2013 intelligent bus-connected sensing units performing locally data filtering, elaboration and model
Recommended by A. Alessandro
identification. This paper describes a family of enhanced multivariate autoregressive models that can
be used in SHM-oriented identification procedures and the implementation of a new advanced SHM
Keywords: system in the tower of the Engineering School of Bologna University. It describes also the results given by
SHM systems the considered procedure and a comparison of the implemented MEMS-based system with a traditional
Multivariate identification solution based on piezoelectric seismic accelerometers.
AR models
& 2013 European Control Association. Published by Elsevier Ltd. All rights reserved.
AR+noise models

1. Introduction with active seismic response control systems [31,13] are even
more demanding since they require a proper monitoring concern-
Civil infrastructures, like highways, bridges, airports, seaports, ing the whole operating life of the structures.
railroads, water management systems, oil and gas pipelines, are of Structural identification is, in general, considered as an applied
paramount importance for economic and industrial development. methodology for characterizing a structural system using mea-
These systems are characterized by high costs, strong impact on surements describing how the structure behaves under loading
the safety and quality of life for large communities and long [3,33]. Structural Health Monitoring (SHM) relies on structural
operative lives. Their proper management requires, consequently, identification techniques to perform comparisons between the
the adoption of carefully selected policies developed taking into reference dynamical behavior of the monitored structures and the
account the delicate balance between potentially conflicting current one [5,38,10]. Due to the large number of uncertainties
requirements like, for instance, achieving high safety standards and the unavoidable simplifications performed during the model-
and limiting maintenance costs [1]. Moreover, some specific ing processes (see for instance wind or seismic loading conditions,
events like earthquakes, floods or tornadoes can lead to very behavior of materials and the contributions of non-structural
critical decisions in ascertaining the integrity of surviving struc- elements) the real structural behavior is always different from
tures and their suitability to fulfill their intended role [30]. Similar the descriptions obtained with structural models; it is thus
problems afflict the evaluation of the state of structures built appropriate to verify that the real structural performances match
during the last century (e.g. many bridges in the United States) and with the model response [32]. In addition, the structural system
of ancient buildings inside large cities, exposed to the stress can be subjected to changes (damages) of the material/geometric
caused by the increase of surface and underground urban trans- properties, including changes of the boundary conditions and
port systems. The relevance of these problems is not limited, system connectivity, which adversely affect the system's perfor-
however, to the evaluation of the state of structures potentially mance [10]. The SHM process involves the observation of a system
damaged by traumatic events; the advanced technologies imple- over time using periodically sampled dynamic response measure-
mented in the realization of new projects, for instance, buildings ments from an array of sensors with a suitable sampling rate.
Traditional SHM systems are essentially composed of a certain
n
number of analog sensors (strain gauges, accelerometers, tem-
Corresponding author. Tel.: +39 0512093027
E-mail addresses: [email protected] (R. Guidorzi),
perature sensors) connected, through signal conditioning units,
[email protected] (R. Diversi), [email protected] (L. Vincenzi), to multichannel data loggers. The measures obtained from
[email protected] (C. Mazzotti), [email protected] (V. Simioli). these sensors are the structure response to external or internal

0947-3580/$ - see front matter & 2013 European Control Association. Published by Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ejcon.2013.06.004

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
2 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎

excitations due to wind and other meteorological phenomena, be processed and stored. To avoid the off-line manipulation of all
vehicle traffic, seismic events, mass movements [36] or use of data, it is preferable to implement automated procedures
specific excitation hardware like mechanical shakers [26]. The [30,14,27] or to work only on some form of concentrated informa-
evaluation of the information contained in the data is then usually tion extracted from the data, like dynamic models obtained by
performed off-line by experts relying on suitable models of the means of identification techniques. These techniques allow not
structure to be analyzed and on the compliance of these models only for a very large condensation of information but can also be
with the measured data. Damage-sensitive features and the effectively used to separate the information contained in the
statistical analysis of these features are then used to determine acquired data sets from the observation errors due to the intrinsic
the current state of the system health. The application of SHM noise of the sensors and to the inevitable misfit between the
techniques leads to a periodically updated information regarding considered class of models and the real process to be described.
the ability of the structure to perform its intended function in the Thus the power spectrum associated with an identified model will
presence of the inevitable aging and degradation resulting from look as (and will be) a smoothed version of that directly obtained
operational environments. After extreme events, such as earth- by applying FFT techniques to the measured sequences that would
quakes, SHM is used for rapid condition screening and aims to also lead to many spurious lines due to the additive noise.
provide, in near real time, reliable information regarding the In typical SHM applications, the process input, i.e. the excita-
integrity of the structure [30]. The state of the structure can be tion applied to the structure, is only seldom measured (usually this
evaluated by comparing the responses obtained in reference happens only when artificial inputs are applied for test purposes
(integrity) conditions with the current ones [10,9]; these compar- by means of hammers, mechanical shakers or other methods); in
isons could be (and, sometimes, are) performed by directly almost all permanent implementations of SHM systems the
extracting from the collected time series information depending excitation is given by vehicle traffic or natural phenomena as
only on the structure [11,29,22,39]. In other words, Structural seismic events and wind pressure and is not directly measured.
Health Monitoring aims to give, at every moment during the life of The available data are thus given by sequences of observations,
a structure, a diagnosis of the “state” of the different parts, and of yð1Þ; yð2Þ; …; yðLÞ, where y(t) denotes the vector of acceleration
the full assembly of these parts constituting the structure as a measures, i.e.
whole. The state of the structure must remain in the designed
yðtÞ ¼ ½y1 ðtÞ y2 ðtÞ … ym ðtÞT : ð1Þ
range despite the deterioration due to usage, the action of the
environment and accidental events. A class of models frequently used to represent observations of this
In this complex and demanding context, SHM methodologies kind is given by multivariate AR processes, described by the
are taking advantage of the unprecedented development of relation
sensors, microelectronics and microprocessor technologies [6,28],
yðtÞ ¼ Q 1 yðt1Þ þ Q 2 yðt2Þ þ ⋯ þ Q μ yðtμÞ þ eðtÞ ð2Þ
wireless communications [25,21,37,23], and are moving towards
their life-long integration in new projects while still playing the where the matrices Qi (i ¼ 1; …; μ) are square (m  m) coefficient
role of advanced analysis tools for evaluating the state of struc- matrices,
tures not endowed with permanent monitoring systems. In 2 3
q11i q12i … q1mi
particular, the introduction of MEMS sensors [6] allows the 6 7
6 q21i q22i … q2mi 7
realization of systems that conjugate a reduced cost with perfor- Qi ¼ 6
6 ⋮
7; ð3Þ
mances suitable for SHM applications. The new generation of SHM 4 ⋮ … ⋮ 75
systems relies on designs that integrate advanced sensor technol- qm1i qm2i … qmmi
ogies with distributed computational power as well as efficient
the integer μ denotes the memory of the model and
implementations of identification methodologies. These advanced
systems rely on intelligent sensors that elaborate local models and eðtÞ ¼ ½e1 ðtÞ e2 ðtÞ … em ðtÞT ð4Þ
exchange data, information and models on a local network under
the supervision of a control/storage unit accessible also in the is a vector whose elements ei(t) (i ¼ 1; …; m) are white processes
Internet. with null expected value, E½ei ðtÞ ¼ 0, and with variances s2ei ; these
The purpose of this paper is twofold. A first contribution processes can be mutually correlated so that the covariance matrix
concerns the use of AR+noise models [7] in applications that fit of e(t) is not necessarily diagonal. By denoting with z1 the unitary
their stochastic environment like in the case of this application. A delay operator, model (2) can also be written in the compact
second contribution concerns the description of the practical polynomial form:
results given by a new SHM system developed at Bologna Q ðz1 ÞyðtÞ ¼ eðtÞ ð5Þ
University and engineered by Teleco SpA whose advanced archi-
1
tecture, based on a digital bus and on MEMS sensors, is basically where Q ðz Þ is the polynomial matrix
1
different from traditional ones based on radial analog connections Q ðz Þ ¼ IQ 1 z1 ⋯Q μ zμ : ð6Þ
of piezoelectric sensors. The data sets used for identification and
validation purposes have been collected on the experimental Fig. 1 shows a possible interpretation of multivariate AR models
system installed in the Tower of the School of Engineering of that can be seen as filters driven by the input vector e(t) with
Bologna University during small seismic events. Further data have transfer matrix Q ðz1 Þ1 and output vector y(t).
been collected from temporarily installed traditional piezoelectric Model (2) belongs to the family of equation error models and
accelerometers and have been used for a comparative analysis. its optimal predictor (minimal variance and whiteness of the
prediction error on every output) is given by [17]
^ ¼ Q 1 yðt1Þ þ Q 2 yðt2Þ þ ⋯ þ Q μ yðtμÞ
yðtÞ ð7Þ
2. Modeling data in SHM applications

All dynamic SHM implementations rely on data, measured with

a suitable sampling rate, acquired by a certain number of accel-
erometers mounted on the structure to be monitored. SHM
systems usually generate enormous amounts of data that must Fig. 1. Interpretation of multivariate AR models.

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3

^ ¼ eðtÞ coincides with the

and its prediction error εðtÞ ¼ yðtÞyðtÞ explained by the previous observation on the nature of real
equation error. By denoting with θ a generic set of parameters of processes.
the model, i.e. a generic set of entries of the matrices Qi, the The choice of the algorithms to estimate the parameters of
prediction error obtained by using this parameterization in pre- multivariate AR models is not limited to Least Squares; another
^ θÞ; it coincides with
dictor (7) will be denoted as εðt; θÞ ¼ yðtÞyðt; possibility concerns the use of Yule–Walker equations or of the
e(t) only when the entries of θ are the exact parameters, θn , of the wide class of Instrumental Variable (IV) algorithms (in fact Yule–
AR process that has generated the data. The model parameters are Walker equations constitute a subcase of the IV approach where
usually estimated by minimizing the Euclidean norm of the past outputs are used as instruments). These options can be used
prediction error εðt; θÞ: to compensate a possible lack of whiteness in the equation error
sequences (by avoiding the use of the first low order equations
1 L 1 L 1 m L
JðθÞ ¼ ∑ ‖εðt; θÞ‖22 ¼ ∑ εðt; θÞT εðt; θÞ ¼ ∑ ∑ ε ðt; θi Þ2 when using Yule–Walker equations, by selecting suitable instru-
N t ¼ μþ1 N t ¼ μþ1 N i ¼ 1 t ¼ μþ1 i
ments when relying on IV approaches) but lead, in general, to a
ð8Þ larger uncertainty on the parameter estimates, i.e. to covariance
where N ¼ Lμ, θi is the vector of the coefficients appearing in the matrices of the estimates larger than the LS one. Other possibilities
i-th row of Q ðz1 Þ concern the use of on-line algorithms, typically on-line least
squares (weighted or not) to update a model as long as new
θi ¼ ½qi11 …qim1 …qi1μ …qimμ T ; ð9Þ measures are performed. The Levinson algorithm offers an elegant
and efficient way to compute increasing-order AR models from
and εi ðt; θÞ ¼ yi ðtÞy^ i ðt; θÞ. The minimization of JðθÞ can be easily
data covariances. A sequence of increasing-order AR models can
performed by means of the Least Squares (LS) method [17]. For
also be estimated directly from the data by means of least squares
this purpose, define the Hankel matrix of output samples
2 3 approaches.
yi ð1Þ yi ð2Þ … yi ðμÞ It has been shown that reliable procedures for modal identifi-
6 7
6 yi ð2Þ yi ð3Þ … yi ðμ þ 1Þ 7 cation can be used to develop an efficient modal-based Structural
Hðyi Þ ¼ 6
6 ⋮
7;
7 ð10Þ Health Monitoring system using, for example, the AR (or ARX)
4 ⋮ … ⋮ 5
yi ðNÞ yi ðN þ 1Þ … yi ðL1Þ coefficients as damage-sensitive parameters [11,29,34]. Moreover,
unlike other dynamic models, AR models are unaffected by over-
the matrix parameterization. When these algorithms are applied to records,
including the structural response to a ground motion, they can
H ¼ ½Hðy1 Þ Hðy2 Þ … Hðym Þ ð11Þ
lead to unreliable results due to the fact that the hypothesis about
and the vector of output samples the input (white noise) could not be fulfilled by the earthquake
spectra. It is worth noting that the near-fault ground motion
yoi ¼ ½yi ðμ þ 1Þyi ðμ þ 2Þ…yi ðLÞT : ð12Þ
spectra are significantly different from those obtained in a far-
Then, under suitable excitation conditions (non-singularity of field condition [2] in that usually a near-fault earthquake can be
ðH T HÞ), the LS estimate of θi is given by viewed as an impulse; moreover, intensity, ground motion spatial
variations and local site conditions can influence significantly the
θ^ i ¼ ðH T HÞ1 H T yoi ði ¼ 1; …; mÞ: ð13Þ
earthquake spectra. For these reasons, in some cases the ground
All previous steps can be easily performed on the basis of a set of motion spectra can be assumed as flat at least in the range of
observed process sequences but requires a previous choice of the frequencies of interest. In these cases, the input of the process to
model memory, μ (the model order is n ¼ deg det Q ðz1 Þ ¼ mμ). Of be identified can be assumed as white.
course, when the observations are generated by a true multivariate Once that a multivariate AR model has been identified, it is also
AR process, μ and n should assume their true values and could be possible to obtain equivalent representations to fulfill specific
estimated by applying suitable order selection criteria like FPE needs; control applications could call, for instance, for state-space
(Final Prediction Error), AIC (Akaike Information Criterion), MDL models. Other representations frequently used are the transfer
(Minimum Description Length) or others [17,24]. These criteria are matrix between the driving noise and the output, Q ðz1 Þ1 , the
usually formulated for the univariate case but can be easily impulse response (AR models do not consider any measurable
extended to the multivariate one. While all previous criteria give input; the input pulse is considered on the components of e(t)) and
correct results for data generated by true AR processes, real the model power spectra. When the models must be used for fault
processes are intrinsically distributed so that the correct model diagnosis applications, as in SHM, the choice of the representation
memory should be infinite and different criteria can lead to to be used can be critical. Consider, as an example, a non-
different evaluations. minimally parameterized model; its parameters could exhibit
A reliable criterion that can be applied in the identification of large but mutually compensated variations also in the absence of
real processes and that can be used not only to select a proper significant process changes. It is thus important to observe
model order but also to validate the whole identification proce- possible changes, to look at model properties reflecting actual
dure consists in checking the whiteness of the prediction errors variations of the identified process; possible choices could concern
εi ðt; θÞ; this happens only if the model order is sufficient and the the parameters of minimally parameterized models, model poles,
description of the considered process by means of an AR model is frequency responses, power spectra and cross-spectra.
acceptable. A good strategy can thus consist in starting with μ ¼ 1 Another desirable feature usually absent in identified models
and increasing μ checking, at every step, the whiteness of the concerns the physical significance of the models; the models
sequences εi ðt; θÞ; as soon as all these sequences satisfy a proper obtained by means of identification techniques can be very
whiteness test (for instance a χ 2 test with a number of degrees of accurate but usually lack, differently from those obtained by
freedom equal to 2–3 times the model memory and a confidence means of traditional modeling techniques, a direct physical mean-
level of, say, 99%), a suitable model memory has been reached. It ing. This requirement and the previous one lead often to the use,
must, however, be observed that the use of higher values, while in SHM applications, of the spectra and cross-spectra associated
leading to overparameterized models and to higher uncertainty with the identified multivariate models. This information reflects
levels for the parameters, does not lead, usually, to a crash of the well defined physical properties of the structures and can be easily
identification procedure or to worse results and this can be easily linked to project-level evaluations.

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
4 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Remark 1. Relation (2) is universally considered as the standard identification problem into an EIV identification scheme, more
definition of multivariate AR models. This definition is, however, precisely into the Frisch scheme that allows estimating, by means
afflicted by conceptual limitations because of the implicit assump- of a search procedure, both model parameters and the observation
tion that all channels have the same memory; thus the order of the and process noise variances [18]. The estimate of AR+noise models
processes described by these models can assume only values by means of a Frisch-scheme approach has been described in [7,8]
multiples of the model memory. More general and minimally for the univariate case but can be extended to the multivariate
parameterized representations of multivariate systems have been context. An approach of this kind has the advantage of leading to a
described in [15–17] and could also be used in the SHM context to congruent solution and of being intrinsically suitable for fault
obtain more physically precise descriptions of complex structures. diagnosis; a possible disadvantage concerns the fact that the
Good results can usually also be obtained by using basic AR models stability of the obtained model is not assured. Another way to
like (2) and (5). solve the problem could rely on the use of compensated least
squares schemes, like BELS algorithms [40,41]. These algorithms
3. Advanced AR modeling: AR+noise representations are iterative and, usually, fast but they assure neither congruence
nor convergence.
Traditional AR models are endowed with many advantages that An approach suggested by filtering techniques applied in
range from the easy estimation of their parameters by means of speech enhancement relies on the separate estimate of the
unbiased and efficient algorithms like LS to the stability of the variance of the additive observation noise from sequences col-
associated optimal predictors (independent from the stability of lected in the absence of signals (silent frames). This estimate is
the model). These models can be interpreted according to the then used to compensate the presence of the observation noise
scheme reported in Fig. 1 where the equation error e(t) is reducing thus the AR+noise estimation problem to the estimation
considered as input of a filter; in these models the equation error of an AR model. A procedure of this kind can also be adopted in the
e(t) is the only tool available to describe the different causes for multivariate case and effectively applied in the SHM context. It
misalignment between the model and the data (non-linearities, allows also the extension to more general contexts where not all
process noise, observation errors, non-stationarity, etc.). A more observation errors are independent and this can be of practical
sophisticated way to manage these inevitable misalignments relevance in SHM. To illustrate this two-step procedure, consider,
consists in introducing a specific description of the observation for an AR+noise process, the covariance matrix
errors, separating these errors from those due to other causes. AR H nT H n
+noise models consist in AR models whose output is considered to Σ n ¼ lim ð18Þ
N-1 N
be affected by an additive observation error (Fig. 2).
AR+noise models are thus described by the equations: where Hn has the same structure as H and is constructed with
samples, yn ðtÞ, of the AR part of the model. Because of the relation
yn ðtÞ ¼ Q 1 yn ðt1Þ þ ⋯ þ Q μ yn ðtμÞ þ eðtÞ ð14Þ yðtÞ ¼ yn ðtÞ þ wðtÞ and of the assumption of non-correlation
between e(t) and w(t), and, consequently, between yn ðtÞ and
yðtÞ ¼ yn ðtÞ þ wðtÞ ð15Þ w(t), it follows that
where
HT H
T
Σ ¼ lim ¼ Σ n þ Σ oe ð19Þ
wðtÞ ¼ ½w1 ðtÞ w2 ðtÞ … wm ðtÞ ð16Þ N-1 N

is a vector whose elements wi(t) (i ¼ 1; …; m) are white processes where Σ oe denotes the covariance matrix of observation errors
mutually uncorrelated, uncorrelated with the entries of e(t), with
Σ oe ¼ diag½s2w1 I μ …s2wm I μ : ð20Þ
null expected value, E½wi ðtÞ ¼ 0, and with variances s2wi ; the
n
covariance matrix of w(t) is thus diagonal If Σ oe is known, it is possible to deduce Σ from (19) and,
consequently, to reduce the problem to the identification of an
Σ w ¼ diag½s2w1 s2w2 …s2wm : ð17Þ
AR process by substituting H T H with NΣ. In practical applications
More general AR+noise schemes could consider additive colored relation (19) will be applied to the available sample quantities by
noise on the observations and/or the presence of correlations means of the relation
between the observation noises. The interest, in SHM implemen-
H nT H n ¼ H T HNΣ oe ð21Þ
tations, of the first extension is modest while the second one, as it
will be shown in the sequel, can be necessary for a realistic and, under the assumption of non-correlation between e(t) and
description of some sensors. w(t), an asymptotically unbiased estimate of the AR model para-
The identification of AR+noise models is more complex than meters is
the identification of AR models because it is necessary to estimate
θ^ i ¼ ðH T HNΣ oe Þ1 H T yoi ði ¼ 1; …; mÞ: ð22Þ
not only the parameters of Q ðz1 Þ and the covariance matrix of e(t)
but also the covariance matrix of w(t) and, in this stochastic An estimate of Σ w and, consequently, of Σ oe can be obtained by
context, LS would lead to biased estimates. computing the sample covariance matrix of output sequences that
The parameters of AR+noise models could be estimated by do not contain any useful information (silent frames); this can be
means of IV algorithms; the disadvantage of this solution concerns verified by means of a whiteness test on the components of y(t).
the uncertainty of the estimates and the fact that the variances of
the equation errors and of the observation noise are not estimated. Remark 2. The subtraction from the main diagonal of Σ of the
Another approach could rely on mapping the AR+noise diagonal elements of Σ w (in blocks of μ elements) can lead to non-
positive definite matrices ðH T HN Σ oe Þ and/or to estimate unstable
models. The reasons derive both from the approximation asso-
ciated with the use of sample quantities and from the assumption
of zero off-diagonal elements in Σ oe . When this happens it is
possible to modify relation (22) as follows:

Fig. 2. Structure of AR+noise models. θ^ i ¼ ðH T HkNΣ oe Þ1 H T yoi ði ¼ 1; …; mÞ ð23Þ

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 5

where 0 ok o1 is chosen in order to respect the condition

ðH T HkNΣ oe Þ 40 and the stability constraint.

4. The SHM system in the tower of the Engineering School of

Bologna University

The building where the tower is located has been designed by

the Italian architect Giuseppe Vaccaro and was built between 1933
and 1935 (Fig. 3). The tower is actually an archive capable of
holding over 60,000 volumes, arranged on movable metal shelves.
It is approximately 45 m high and its structure is characterized by
four rectangular columns which support nine concrete slabs. The
measures are performed by means of a prototype of the advanced
SHM system developed at the University of Bologna and engi-
neered by Teleco, the SHM602 [35], compliant with the recom-
mendations reported in [12,19].
The main components of this system consist in a controller/
storage unit TSD10 and in intelligent sensing units TSM02 con-
nected to the controller by means of a serial bus. Every sensing
unit can send the measures of the acceleration on two orthogonal
axes and that of the temperature; the sampling frequency can be
selected by the user at 20 Hz, 40 Hz or 80 Hz. These units rely on
MEMS sensors that, because of the used local digital filtering
techniques and oversampling rates, exhibit a noise floor almost
one order of magnitude lower than that of other MEMS based
systems. The dynamic behavior of the tower is monitored by
means of four TSM02 units (eight accelerometers) installed in four
different floors (M1–M4 in Figs. 4 and 5). Their locations have been
carefully chosen in order to avoid nodal points (zero response
points) on the first several vibration mode shapes. Four piezo-
electric single-axis accelerometers (denoted as A1–A4 in
Figs. 4 and 5) have been temporarily installed in two of the
previous locations for comparison purposes. A first set of measures
(in mg) has concerned the evaluation of the signal variances in the
absence of excitation; this can be easily performed since the
building is located in a quiet area, outside traffic patterns. Fig. 4. Accelerometer locations in the tower.

The obtained sample variances and covariances are reported in

Table 1 where it can be observed that the noises on the x- and y-axes reported in Table 1) are, on the contrary, negligible. This observation
of the same unit exhibit a non-negligible correlation; the covari- can be easily explained since the accelerometers of every TSM02 unit
ance values associated with accelerometers of different sensors (not are physically allocated on the same MEMS chip. It can also be
observed that the obtained variances are perfectly aligned with the
nominal values of TSM02 units with the exception of the fourth
sensor whose noise level is approximately 20% lower.
The stochastic environment of the considered application and
the possibility of constructing a sample noise covariance matrix on
the basis of the measures in the absence of excitation suggest the
use of AR+noise models instead of AR models. It can also be
observed that, from a spectral point of view, AR+noise models are
equivalent to ARMA models [20] that, however, would not allow
exploiting the available information about the observation noise.

4.1. Results of the identification experiments

Several models using all eight available channels and charac-

terized by different memory values have been tested; of course
increasing the memory model leads to higher resolutions of the
associated power spectra. The model described in the following is
limited to four channels in order to comply with space constraints
without omitting any feature of interest of the adopted procedure;
its memory is μ ¼ 80 in order to allow a fair comparison between
the model spectrum and the spectrum directly obtained from the
observations. The measures considered in the construction of this
model are reported in Table 2.
The identification has been performed by using AR+noise
Fig. 3. The Engineering School building. models and the covariance matrix of the observation noise has

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
6 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎

Fig. 5. Accelerometer locations in the selected floors.

100
Table 1
Variances and covariances of measured noise. 80
Data set 1
Sensors s2x s2y sxy 60

40
Sensor M1 0.1087 0.0952  0.0421
Sensor M2 0.0993 0.0986  0.0379 20
Sensor M3 0.1061 0.0965  0.0373
Sensor M4 0.0773 0.0771  0.0315 0
0 1 2 3 4 5 6 7 8 9 Hz

Fig. 6. Normalized power spectra of y1; data-based (continuous line) and model-
based (dashed line).

Table 2 and congruent variations have been introduced in Σ oe that

Model outputs. assumes the following form:
2 3
Sensors Channel 1 Channel 2 Channel 3 Channel 4 0:1087I μ 0:0421I μ 0 0
y1 y2 y3 y4 6 7
6 0:0421I μ 0:0952I μ 0 0 7
Σ oe ¼ 6
6
7:
7 ð25Þ
Sensor M1 x-axis y-axis – – 4 0 0 0:0986I μ 0 5
Sensor M2 – – y-axis – 0 0 0 0:0965I μ
Sensor M3 – – – y-axis
Sensor M4 – – – –
The data set used for the identification has been recorded on
December 5, 2010 and concerns a small seismic event with
magnitude 3.2 observed at a depth of 15 km in the area of Castel
San Pietro Terme, at a distance of 28 km from Bologna. The most
been constructed on the basis of the measures reported in Table 1. suitable k values to be used in (23) have been obtained from tests
Since two channels of the same sensor have been inserted in the on the positive definiteness of ðH T HkNΣ oe Þ and on the stability of
model, considering a diagonal covariance matrix Σ w for the the AR+noise model. A comparison between the power spectra of
additive observation noise would not be congruent with the the four observed sequences and those computed by means of the
measured covariances; thus the evaluation of Σ w that has been identified model is reported in Figs. 6–9 where the “data-based”
actually used is PSDs have been obtained from the FFT of the observed signals
2 3 (computed by means of the classical Cooley–Tukey algorithm).
0:1087 0:0421 0 0
6 0:0421 0:0952 Moreover, the Matlab smooth function and a scaling normalization
6 0 0 7 7
Σw ¼ 6 7 ð24Þ have been used for an easier comparison.
4 0 0 0:0986 0 5 The significant peak frequencies obtained from these models
0 0 0 0:0965 are reported in Table 3. It can be observed that the main resonance

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 7

100 100

80 80
Data set 1 Data set 2
60 60

40 40

20 20

0 0
0 1 2 3 4 5 6 7 8 9 Hz 0 1 2 3 4 5 6 7 8 9 Hz

Fig. 7. Normalized power spectra of y2; data-based (continuous line) and model- Fig. 10. Normalized power spectra of y1; data-based (continuous line) and model-
based (dashed line). based (dashed line).

100
100
80
80
Data set 1
60 Data set 2
60
40
40
20
20
0 0
0 1 2 3 4 5 6 7 8 9 Hz 0 1 2 3 4 5 6 7 8 9 Hz
Fig. 8. Normalized power spectra of y3; data-based (continuous line) and model- Fig. 11. Normalized power spectra of y2; data-based (continuous line) and model-
based (dashed line). based (dashed line).

100 100

80 80
Data set 1 Data set 2
60 60

40 40

20 20

0
0 0 1 2 3 4 5 6 7 8 9 Hz
0 1 2 3 4 5 6 7 8 9 Hz
Fig. 12. Normalized power spectra of y3; data-based (continuous line) and model-
Fig. 9. Normalized power spectra of y4; data-based (continuous line) and model-
based (dashed line).
based (dashed line).

100

Table 3 80
Peak frequencies (model 1/model 2). Data set 2
60
Channel 1 1.60/1.60 1.75/1.80 – 2.45/2.50 3.725/3.85
40
Channel 2 – – 2.20/2.20 2.40/2.40 –
Channel 3 – – 2.20/2.20 2.40/2.40 – 20
Channel 4 – – 2.20/2.20 2.40/2.40 –
0
0 1 2 3 4 5 6 7 8 9 Hz

Fig. 13. Normalized power spectra of y4; data-based (continuous line) and model-
based (dashed line).
frequencies along the x-axis are approximately 1.60 and 1.80 Hz
while a third frequency is around 3.8 Hz; the resonance frequen-
cies along the y-axis are approximately 2.2 Hz and 2.40 Hz. 100

80
4.2. Validation of identification results Data sets 1 and 2
60

Validating a model consists, essentially, in verifying whether 40

the model is suitable for the purpose that has suggested its
20
construction; it is thus possible to adopt different validation
criteria in different applications. In this case the most relevant 0
0 1 2 3 4 5 6 7 8 9 Hz
information associated with the identified models concerns their
PSD and, consequently, it makes sense to compare the PSDs of Fig. 14. Cross spectra between y2 and y1; model 1 (continuous line) and model 2
(dashed line).
models obtained from different data sets because their congruence
can confirm the absence of damages and the model suitability in
describing the structure's behavior. A second model has thus been is reported in Figs. 10–13. The associated significant peak frequen-
obtained from data recorded on December 6, 2010 concerning cies, reported in Table 3, are congruent with those of the previous
another small seismic event with magnitude 3.0 observed at a model. The cross-spectra between y2 and y1, y3, y4, computed with
depth of 24 km in the same area as the previous one. The the models identified from the considered data sets, are reported
comparison between the power spectra of the four observed in Figs. 14–16 that show good agreements. Similar results can be
sequences and those computed by means of the identified model observed on remaining cross-spectra.

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
8 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎

100 100

80 80
Data sets 1 and 2 Data set 2
60 60

40 40

20 20

0 0
0 1 2 3 4 5 6 7 8 9 Hz 0 1 2 3 4 5 6 7 8 9 Hz

Fig. 15. Cross spectra between y2 and y3; model 1 (continuous line) and model 2 Fig. 17. Data-based PSD of y1: data from MEMS (continuous line) and piezoelectric
(dashed line). accelerometers (dashed line).

100

80
Data sets 1 and 2 100
60
80
40 Data set 2
60
20
40
0
0 1 2 3 4 5 6 7 8 9 Hz 20

Fig. 16. Cross spectra between y2 and y4; model 1 (continuous line) and model 2 0
(dashed line). 0 1 2 3 4 5 6 7 8 9 Hz

Fig. 18. Data-based PSD of y2: data from MEMS (continuous line) and piezoelectric
accelerometers (dashed line).
A further check of the validity of the identified models has been
performed by comparing their power spectra with those obtained
with piezoelectric accelerometers placed in the same positions. 100
Figs. 17 and 18 show that, even if the noise level of the MEMS-
80
based sensing units is greater than that of piezoelectric acceler-
Data set 2
ometers, the obtained PSDs are strongly congruent. This confirms 60
the suitability of the identified models and of the SHM602 system
40
for the performed task.
Finally, in order to confirm the good results obtained from the 20

AR+noise model, a comparison with a classic frequency domain 0

technique has been performed. The Enhanced Frequency Domain 0 1 2 3 4 5 6 7 8 9 Hz

Decomposition (EFDD) technique is one of the most used identi- Fig. 19. Power spectra of y1: data-based (continuous line) and obtained with the
fication techniques in civil engineering applications and it can be EFDD technique (dashed line).
considered as an extension of the so-called Peak–Picking method.
The modal parameters can be estimated from the output power
spectral density (PSD), in the condition of white noise input and 100
lightly damped structures. A Singular Value Decomposition (SVD) 80
is carried out for the PSD matrix for each frequency and all modes Data set 2
60
contributing to the vibration information at a given frequency are
separated into principal value and orthogonal vectors [4]. In the 40
proximity of each PSD peak, the power of the measured signals
20
can be attributed to a limited subset of modes. The first singular
values correspond to these modes, while the other singular values 0
0 1 2 3 4 5 6 7 8 9 Hz
that are not associated with any mode consist of decomposed
noise contained in the signals. This means that the decomposition Fig. 20. Power spectra of y2: data-based (continuous line) and obtained with the
EFDD technique (dashed line).
corresponds to a Single Degree of Freedom identification of the
system for each singular value.
The singular value plot of the spectral density matrix concen-
trates information from all spectral density functions. Moreover, if superior to a user-specified parameter (which is called the MAC
the assumptions are fulfilled, this technique simply provides a rejection level) are kept, thus forming an equivalent single degree of
modal decomposition of the vibration information, and the modal freedom spectral bell. Then, by inverse fast Fourier transform (IFFT)
information for each mode – even in the case of closely spaced of that spectral bell, the resulting auto-correlation function can be
modes and noise – can be extracted easily and accurately, for used to reevaluate the frequency by counting the number of zero
example by means of the peak-picking method. Once that the crossings in a finite time interval. Damping ratios are also estimated
natural frequencies have been identified by peak-picking, the mode using the logarithmic decrement of the auto-correlation function
shapes have been estimated by using the singular vector matrices [30].
and the equivalent single degree of freedom “spectral bells” are After identifying the main modes, the PSD for each channel can
identified for each mode by comparing the estimated mode shape be computed by means of the following equation:
of interest with all vectors previously estimated throughout the
Pyy ðjωÞ ¼ GðjωÞT Pxx ðjωÞGðjωÞ ð26Þ
spectrum by SVD of the PSD matrices. A comparison of the mode
shapes is then carried out by computing the modal assurance where Pxx ðjωÞ is the Power Spectra matrix of the inputs that is
criterion (MAC). All singular values corresponding to a MAC value constant because the inputs are assumed as white noise processes

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 9

100 [4] R. Brincker, L. Zhang, P. Andersen, Modal identification of output-only systems

using frequency domain decomposition, Smart Materials and Structures 10
80 (2001) 441–445.
Data set 2 [5] F.K. Chang (Ed.), in: Proceedings of the First, Second and Third International
60
Workshops on Structural Health Monitoring, Stanford University, Stanford, CA,
40
CRC Press, New York, 1997, 1999 and 2001.
[6] H.C. Chung, T. Enotomo, K. Loh, M. Shinozuka, Real-time visualization of bridge
20 structural response through wireless MEMS sensors, in: Proceedings of SPIE—
The International Society for Optical Engineering, 2004.
0 [7] R. Diversi, R. Guidorzi, U. Soverini, Identification of autoregressive models in
0 1 2 3 4 5 6 7 8 9 Hz the presence of additive noise, International Journal of Adaptive Control and
Signal Processing 22 (2008) 465–481.
Fig. 21. EFDD technique: first singular value in the decomposed power spectra. [8] R. Diversi, U. Soverini, R. Guidorzi, A new estimation approach for AR models
in presence of noise, in: Preprints of the 16th IFAC World Congress, Prague,
Czech Republic, 2005.
and GðjωÞ is the Frequency Response Function matrix. The Fre-
[9] S.W. Doebling, C.R. Farrar, M.B. Prime, D.W. Shevitz, A review of damage
quency Response Function GðjωÞ can be written, using a partial identification methods that examine changes in dynamic properties, Shock
fraction expansion, in the following pole-residue form: and Vibration Digest 30 (1998) 91–105.
! [10] S.W. Doebling, C.R. Farrar, M.B. Prime, D.W. Shevitz, Damage Identification and
N Rk Rk Health Monitoring of Structural and Mechanical Systems from Changes in
GðjωÞ ¼ ∑ þ ð27Þ Their Vibration Characteristics: A Literature Review, Los Alamos National
k ¼ 1 jωλk jωλ k Laboratory Report LA-13070-MS, Los Alamos National Laboratory, Los Alamos,
qffiffiffiffiffiffiffiffiffiffiffi
NM 87545, 1996.
where the poles λk ¼ ξk ωk þ jωk 1ξ2k and the residues Rk ¼ ϕk γ k [11] C.R. Farrar, S.W. Doebling, D.A. Nix, Vibration-based structural damage
contain the identified parameters; “ ” denotes the complex identification, Philosophical Transactions of the Royal Society: Mathematical,
conjugate. Physical & Engineering Sciences 359 (2001) 131–149.
[12] Fib Task Group 5.1, Monitoring and Safety Evaluation of Existing Concrete
Figs. 19 and 20 report a comparison between the PSD of the Structures, State-of-the-Art Report, Final Draft, 2002.
recorded data and the estimated ones. It can be observed that [13] Y. Fujino, A. Nishitani (Eds.), in: Proceedings of the Fifth World Conference on
the estimated PSD (dashed curves) does not always fit well the Structural Control and Monitoring (5WCSCM), Tokyo, 2010.
[14] Y. Gao, B.F. Spencer, Online damage diagnosis for civil infrastructure employ-
experimental ones. This is mainly due to a poor estimation of the ing a flexibility-based approach, Smart Materials and Structures 15 (2006)
damping ratios and to the fact that the selection of the eigen- 9–19.
frequencies can be a critical task (see, for example, Fig. 21). [15] R. Guidorzi, Equivalence, invariance and dynamical system canonical model-
ling. Part I, Kybernetika 25 (1989) 233–257.
[16] R. Guidorzi, Equivalence, invariance and dynamical system canonical model-
ling. Part II, Kybernetika 25 (1989) 386–407.
5. Concluding remarks [17] R. Guidorzi, Multivariable System Identification: From Observations to Models,
Bononia University Press, Bologna, Italy, 2003.
This paper has described some families of multivariate models [18] R. Guidorzi, R. Diversi, U. Soverini, The Frisch scheme in algebraic and dynamic
identification problems, Kybernetika 44 (2008) 585–616.
that can be used in SHM-oriented identification procedures and, in [19] ISO/DIS 18649:2002(E). Mechanical Vibration—Evaluation of Measurement
particular, the extension of AR models to consider the presence of Results from Dynamic Tests and Investigations on Bridges, 2002.
additive measurement noise. It describes also the implementation, [20] S.M. Kay, The effects of noise on the autoregressive spectral estimator, IEEE
Transactions on Acoustics, Speech and Signal Processing 27 (1979) 478–485.
in the tower of the Engineering School of Bologna University, of a [21] S. Kim, Wireless Sensor Networks for Structural Health Monitoring, Master's
new advanced MEMS-based SHM system developed at Bologna Thesis, University of California at Berkeley, May 2005.
University and engineered by Teleco. Two data sets collected by [22] H.S. Kim, H. Melhem, Damage detection of structures by wavelet analysis,
Engineering Structures 26 (2004) 347–362.
this system during small seismic events have been used to test the [23] Q. Ling, Z. Tian, Y. Yin, Y. Li, Localized structural health monitoring using
suitability for SHM applications of multivariate AR+noise models. energy-efficient wireless sensor networks, IEEE Sensors Journal 9 (2009)
The congruence and effectiveness of the identified models have 1596–1604.
[24] L. Ljung, System Identification—Theory for the User, Prentice Hall, Englewood
been verified by means of comparisons between the models
Cliffs, NJ, 1999.
obtained from different sets of data including the data obtained [25] J.P. Lynch, A. Sundararajan, K.H. Law, A.S. Kiremidjian, T. Kenny, E. Carryer,
from temporarily installed seismic sensors. Embedment of structural monitoring algorithms in a wireless sensing unit,
Structural Engineering and Mechanics 15 (2003) 285–297.
The identified multivariate AR+noise models exhibit a strong
[26] C. Mazzotti, L. Vincenzi, Structural identification of a steel structure by forced
congruence; a high level of congruence can also be observed vibrations, in: ICOSSAR—International Conference of Structural Safety and
between the power spectral densities of these models and those of Reliability, Rome, Italy, 2005.
the data used for their identification. The suitability of the new [27] A. Mita, T. Inamura, S. Yoshikawa, Structural health monitoring for buildings
with automatic data management system, in: 4th International Conference on
system for SHM applications has been verified by comparing the Earthquake Engineering, Taipei, Taiwan, 2006.
power spectral densities of the sequences obtained from this [28] T. Nagayama, B.F. Spencer, Structural Health Monitoring Using Smart Sensors,
system with those of traditional (piezoelectric) accelerometers. NSEL Report, Series 001, University of Illinois Urbana-Champaign, 2007
〈https://www.ideals.uiuc.edu/handle/2142/3521〉.
The conclusion that can be deduced from the tests is that [29] K.K. Nair, A.S. Kiremidjian, K.H. Law, Time series-based damage detection and
MEMS-based systems can be as effective as traditional ones, localization algorithm with application to the ASCE benchmark structure,
despite the intrinsically higher level of noise, in SHM applications Journal of Sound and Vibration 291 (2006) 349–368.
[30] C. Rainieri, G. Fabbrocino, E. Cosenza, Structural health monitoring systems as
and that multivariate AR+noise models offer a useful tool for a tool for seismic protection, in: Proceedings of the 14th World Conference on
implementing modal-based SHM analyses. Earthquake Engineering, Beijing, China, 2008.
[31] F. Ricciardelli, A. Occhiuzzi, P. Clemente, Semi-active tuned mass damper
control strategy for wind-excited structures, Journal of Wind Engineering and
References Industrial Aerodynamics 88 (2000) 57–74.
[32] M. Savoia, L. Vincenzi, Differential evolution algorithm for dynamic structural
[1] A.E. Aktan, F.N. Catbas, K.A. Grimmelsman, C.J. Tsikos, Issues in infrastructure identification, Journal of Earthquake Engineering 12 (2008) 800–821.
health monitoring for management, Journal of Engineering Mechanics 126 [33] D. Skolnik, E. Yu, E. Taciroglu, J.W. Wallace, System identification and health
(2000) 711–724. monitoring studies on two buildings in Los Angeles, in: 6th International
[2] A. Bayraktar, A.C. Altunsik, B. Sevim, M.E. Kartal, T. Trker, Y. Bilici, Comparison Workshop of Structural Health Monitoring, Stanford University, Stanford, CA,
of near and far fault ground motion effects on the nonlinear response of dam- 2007.
reservoir-foundation systems, Nonlinear Dynamics 58 (2009) 655–673. [34] H. Sohn, C.R. Farrar, Damage diagnosis using time series analysis of vibration
[3] A. Bonelli, O. Bursi, R. Ceravolo, S. Santini, N. Tondini, Dynamic identification signals, Smart Materials and Structures 10 (2001) 446–451.
and structural health monitoring of a twin deck curved cable-stayed foot- [35] Teleco SHM Systems site: 〈www.telecoshmsystems.com〉.
bridge: the “Ponte del Mare” of Pescara in Italy, in: Proceedings of the 5th [36] L. Vincenzi, C. Mazzotti, M. Savoia, Modal identification of a TAV viaduct using
European Conference on Structural Health Monitoring, Sorrento, Italy, 2010. subspace models, in: 2nd International FIB Congress, Naples, Italy, 2006.

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i
10 R. Guidorzi et al. / European Journal of Control ∎ (∎∎∎∎) ∎∎∎–∎∎∎

[37] Y. Wang, J.P. Lynch, K.H. Law, A wireless structural health monitoring system [40] W.X. Zheng, A least-squares based method for autoregressive signals in the
with multithreaded sensing devices: design and validation, Structure and presence of noise, IEEE Transactions on Circuits and Systems—II 46 (1999)
Infrastructure Engineering 3 (2007) 103–120. 81–85.
[38] K. Worden, C.R. Farrar, G. Manson, G. Park, The fundamental axioms of [41] W.X. Zheng, Fast identification of autoregressive signals from noisy observa-
structural health monitoring, Proceedings of the Royal Society A 463 (2007) tions, IEEE Transactions on Circuits and Systems—II 52 (2005) 43–48.
1639–1664.
[39] G.J. Yuna, S.G. Lee, J. Carletta, T. Nagayama, Decentralized damage identifica-
tion using wavelet signal analysis embedded on wireless smart sensors,
Engineering Structures 33 (2011) 2162–2172.

Please cite this article as: R. Guidorzi, et al., Structural monitoring of a tower by means of MEMS-based sensing and enhanced
autoregressive models, European Journal of Control (2013), http://dx.doi.org/10.1016/j.ejcon.2013.06.004i