IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO.

7, APRIL 1, 2014 1655

Modal Analysis With Compressive Measurements

Jae Young Park, Michael B. Wakin, Senior Member, IEEE, and Anna C. Gilbert, Member, IEEE

Abstract—Structural Health Monitoring (SHM) systems are Although the details of SHM systems vary, some features are
critical for monitoring aging infrastructure (such as buildings or shared among many methods [4], [5]. A typical SHM system
bridges) in a cost-effective manner. Such systems typically involve monitors an in-service structure in real-time. To do so, it makes
collections of battery-operated wireless sensors that sample vibra-
tion data over time. After the data is transmitted to a central node, use of a network of sensors installed on the structure to collect
modal analysis can be used to detect damage in the structure. In vibration data for damage detection. This may include strain
this paper, we propose and study three frameworks for Compres- data, acceleration data, velocity data, or displacement data. The
sive Sensing (CS) in SHM systems; these methods are intended to acquired data from each sensor is transmitted over a network to
minimize power consumption by allowing the data to be sampled the central data repository where damage detection algorithms
and/or transmitted more efficiently. At the central node, all of
these frameworks involve a very simple technique for estimating
are run to detect, localize, or classify possible damage in the
the structure’s mode shapes without requiring a traditional CS structure.
reconstruction of the vibration signals; all that is needed is to An important part of damage detection is a process called
compute a simple Singular Value Decomposition. We provide modal analysis. This process is used to infer properties such as
theoretical justification (including measurement bounds) for each the modal frequencies, mode shapes, and modal damping ratios
of these techniques based on the equations of motion describing
of the structure. Such modal parameters describe the vibrational
a simplified Multiple-Degree-Of-Freedom (MDOF) system, and
we support our proposed techniques using simulations based on characteristics when external forces such as wind, earthquakes,
synthetic and real data. or vehicle loadings are applied to the structure. For example,
Index Terms—Compressive sensing, modal analysis, singular
if a structure is forced to vibrate close to a modal frequency,
value decomposition, structural health monitoring. the shape of the structure’s vibration will be dominated by the
corresponding mode shape. This vibration will eventually die
out in the absence of external force, and the modal damping
I. INTRODUCTION ratio will determine rate of decay.
Many damage detection algorithms make use of modal pa-
A. Structural Health Monitoring Systems rameters to detect, localize, and assess the severity of damage.

O VER the past decade, more than 5 million commercial

buildings [1], 130 million housing units [2], and 0.6 mil-
lion bridges [3] have been built in the United States. In any struc-
Briefly speaking, these methods rely on the notion that when
a structure is damaged, its modal parameters will change. As-
suming that one has the modal frequencies or mode shapes from
ture, damage caused over time by continuous use is inevitable. the time when the structure was healthy, for example, these can
In order to maintain safely operable structures for as long as pos- be compared to the current estimates of the modal frequencies
sible, periodic inspections are a must. When damage is detected, or mode shapes to judge whether or not damage has occurred.
some structures can be repaired, while others must be taken out Modal frequencies are global properties of a structure, although
of service immediately. they can change when a structure undergoes even local damage.
Due to the quantity, size, and complexity of structures, the By comparing the mode shapes before and after damage has oc-
task of inspection is labor intensive, costly, and time consuming. curred, however, one can identify and locate which part of the
Consequently, there have been significant efforts in the struc- structure has changed [6]. A comprehensive survey of damage
tural engineering community to automate this process. Struc- detection methods is presented in [4], [5].
tural Health Monitoring (SHM) systems are precisely designed
to address this issue.
B. Wireless SHM Systems

Manuscript received July 08, 2013; revised November 04, 2013; accepted In the early designs of SHM systems, sensors were linked via
December 23, 2013. Date of publication January 27, 2014; date of current ver- coaxial cables that provided reliable communication to the cen-
sion March 06, 2014. The associate editor coordinating the review of this man- tral data repository. Power cables were coupled with the coaxial
uscript and approving it for publication was Dr. Ashish Pandharipande. This
work was supported in part by NSF Grants CCF-1161233 and CIF-0910765,
cables to provide the power to run the sensors. Unfortunately,
AFOSR Grant FA9550-09-1-0465, and NSF CAREER Grant CCF-1149225. it was often impractical to install a dense network of sensors
J. Y. Park and A. C. Gilbert are with the Department of Electrical Engineering along with all of the requisite cables. When only a few sensors
and Computer Science, University of Michigan, Ann Arbor, MI 48109 USA
(e-mail: jaeypark@umich.edu; annacg@umich.edu).
could be installed, the accuracy of damage detection and anal-
M. B. Wakin is with the Department of Electrical Engineering and Computer ysis would be sacrificed.
Science at the Colorado School of Mines, Golden, CO 80401 USA (e-mail: As a way to overcome this issue, researchers have proposed
mwakin@mines.edu).
Color versions of one or more of the figures in this paper are available online
to deploy wireless sensors on structures to acquire and transmit
at http://ieeexplore.ieee.org. data to the central repository. With the rapid advancement of
Digital Object Identifier 10.1109/TSP.2014.2302736 wireless technology and the ability to build sensors at low cost,

1053-587X © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1656 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

it has become possible to quickly deploy a much denser network that must be wirelessly transmitted to the central repository is
of sensors for a given budget. minimized.
The challenges of wireless sensor networks in SHM are quite Mathematically, denoting the continuous-time displacement
different from those of wired sensor networks. In particular, a signal at sensor as , where , and denotes
primary challenge in a wireless network is to maximize the life the number of nodes, a CS architecture employing non-uni-
expectancy of the batteries that power the sensors. From this form sampling (which is just one possible CS protocol [10])
perspective, there are important questions that should be con- would simply sample1 at time points and transmit the
sidered when designing a wireless sensor network for SHM: resulting samples . The compression would
Should we compress the time data before we transmit it to the come from the fact that is smaller than the number of Nyquist
central data repository? This would allow us to transmit less data samples obtained over an appropriately chosen total sampling
but would require additional power for processing. Or, would time span.2 Typically, one would choose randomly.
we save more battery life by sending the raw data itself without Typically, when we undersample a signal compared to its
processing it at all? Nyquist rate, we must solve an underdetermined system of equa-
In [7], the authors presented thorough answers to these tions in order to reconstruct the original signal (this requires a
questions in real world applications involving a certain wireless sparse model for the signal in some basis). There have been sev-
sensor. The authors found that on average, significant savings in eral papers involving the application of CS to SHM systems.
battery can be expected by locally compressing and processing In [11], the authors implement a prototype wireless sensor that
the data first compared to sending the raw time data. The main computes the compressed measurements locally after the wire-
reason for this was that the radio drains much more power less sensor has acquired the time data. By sending both com-
(almost 3 more) than the on-board processor. As long as the pressed measurements and the raw time measurements to the
execution time of the local algorithm is reasonably short, it central node, the authors evaluate the performance of CS re-
would be more beneficial to first process and compress, then construction of the raw data. It is shown that a number of mea-
transmit less data. surements is needed for an accurate reconstruc-
Another factor that contributes to the draining of the battery is tion of the raw data, where represents the length of the orig-
the sampling process. The Nyquist-Shannon sampling theorem inal Nyquist-rate sample vector. Once the time data is recon-
states that when sampling a signal, the sampling rate should be structed, conventional modal analysis techniques are carried out
proportional to the maximum frequency content in the signal. that aid in damage detection. In [12], the authors also recon-
For the purpose of modal analysis the highest frequency con- struct the original time data from CS measurements obtained
tent is dictated by the highest modal frequency of the underlying at each sensor. Similar to [11], the authors claim that a rela-
structure. Intuitively, structures that are stiff and light will gener- tively large number of measurements are required for accurate
ally possess high modal frequencies, and for such structures we reconstruction.
may have to sample at relatively high rates. Unfortunately, faster The main reason that the above methods require so many
sampling requires more power. The total length of the sampling measurements is because the signals are simply not sparse
time span also plays an important role in optimizing the battery enough in the Discrete Wavelet Transform (DWT) or Discrete
life. Obviously, it would be best to sample only for a certain Fourier Transform (DFT). This suggests that the model of
amount of time and turn the sensor off once a sufficient amount sparsity may not be sufficiently strong for the individual signals
of data has been collected. that arise in modal analysis.
At this point it is worth asking whether signal reconstruction
is necessary at all when employing CS in wireless SHM sys-
C. Compressive Sensing in Wireless SHM Systems tems. The only reason that the above methods attempt to re-
construct the original signals is that conventional modal anal-
In light of these observations, we believe that wireless sen-
ysis tools are designed to operate on signal samples captured
sors equipped with Compressive Sensing (CS) architectures will
at or above the Nyquist rate. There are two reasons to question
be a perfect fit to improve the efficiency and accuracy of wire-
whether this approach is optimal. The first reason is the potential
less SHM systems. The theory of CS has in recent years offered
loss of performance. The frequency content within the signals
the great promise of efficiently capturing essential signal infor-
plays an important role in the subsequent modal analysis. For
mation with low-rate sampling protocols, often below the min-
example, a popular modal analysis algorithm proposed in [13]
imum rate required by the Nyquist sampling theorem [8]–[10].
is very sensitive to the accuracy of each frequency component
The use of CS can dramatically reduce both the complexity of a
of the signal. As we anticipate noise in our acquired signals, the
sensor and the amount of data that must be stored and/or trans-
reconstruction of each signal will also be noisy. These errors
mitted downstream.
will propagate to the modal analysis step, which could poten-
Employing CS in wireless SHM systems would allow
tially lead to misidentification and errors in the estimated modal
compressed data to be acquired directly without the need for
parameters.
local processing. This enables power savings in several ways:
First, because the data can be collected at a rate lower than the 1In this paper, we assume that the time samples obtained at different sensors

Nyquist rate the physical demands on the sensing hardware will are all synchronized in time.
2We use the word “compression” to reflect the fact that fewer samples are
be lower. Second, because there is no need for local processing,
transmitted. A deeper analysis—beyond the scope of this paper—would con-
there is no computational burden at the sensor. Third, because sider specific binary encodings of the raw and CS data and quantify the actual
the data is acquired in compressed form, the amount of data compression savings in terms of bits.
PARK et al.: MODAL ANALYSIS WITH COMPRESSIVE MEASUREMENTS 1657

The second reason is the added computational complexity. accurate than the mode shape estimates obtained using an
Taking the above method for example, using an off-the-self equivalent number of uniform time samples.
CS reconstruction algorithm presented in [14], the total com- • When time samples are collected uniformly in time but
putational complexity for signals will scale as . then multiplied by a compressive matrix, we quantify the
Added to this will be the computation required for the actual accuracy of the estimated mode shapes in terms of the
modal analysis. One can view the reconstruction step as being number of compressive measurements (the number of
somewhat wasteful as it is carried out only to enable the use of columns of ). This analysis provides a mechanism for
conventional modal analysis algorithms. reducing the amount of data that must be transmitted even
when conventional high-rate uniform samplers are used to
record the data at the individual sensors.
D. Contributions As noted above, our theoretical results are currently limited
to MDOF structures with no damping and in free vibration. We
The main objective of this paper is to study a simple method also assume the samples are noiseless. However, at the end of
for directly extracting the mode shapes from CS measurements this paper, we present promising simulation results showing that
without the need for reconstructing the individual time signals. our methods can accurately estimate the mode shapes using a
Unlike previously proposed CS-based methods, we do this by number of samples or measurements that is only a small frac-
processing the sample vectors jointly. tion of the original signal length. These simulations include real
The method we study is as simple as computing the Singular vibration data collected from a bridge in Ypsilanti, MI. Despite
Value Decomposition (SVD) of the signal matrix obtained by the presence of noise in this data and the fact that the bridge does
stacking each sample vector have damping and is not necessarily in free vibration mode, the
into a matrix denoted as . Here, the SVD-based technique is able to successfully estimate the mode
superscript denotes the conjugate transpose operator (we shape vectors. This suggests that our theoretical findings may
explain the use of complex-valued data in Section III-A). As be extendable to more complicated scenarios.
we discuss, this method is inspired by the equations of motion
describing a simplified Multiple-Degree-Of-Freedom (MDOF) II. BACKGROUND
structure with no damping and in free decay. For structures
In this section, we give an introduction to the frequently
of this type, we evaluate the performance of this method both
used mathematical model that governs the motion of structures.
when are chosen deterministically with uniform
These equations form the basis of the proposed CS methods.
spacing and when are chosen uniformly at random.
We begin with the simple Single-Degree-Of-Freedom (SDOF)
We also evaluate the performance when we compute the SVD
system and then move on to the MDOF system. Following
of which is formed by choosing as uniform
what is standard in the structural dynamics community, we use
deterministic time points to form and subsequently mul-
to denote a vector and to denote a matrix . We
tiplying by an ( ) random matrix
denote the th entry of as , and the entry of in
such that . For each sampling method, we give
the th row and th column as . Furthermore, we reserve
sufficient conditions on the required sampling rate, the total
to denote the imaginary unit.
sampling time span, and the total number of measurements for
accurate recovery of mode shape vectors. A. Single-Degree-of-Freedom System
Although it has not been considered in the context of CS, the
use of the SVD for estimating mode shapes has also appeared An SDOF system under no external force can be described
under the name of the Proper Orthogonal Decomposition (POD) by the following differential equation:
in the structural dynamics literature. In fact, it has previously (1)
been shown that, under certain assumptions, the singular vec-
tors (also known as proper orthogonal modes) returned by such where , , and denote the mass, damping, and stiffness pa-
a technique converge to the true mode shapes asymptotically as rameters of the underlying system. To solve for the displace-
the number of observations goes to infinity [15], [16]. We be- ment signal that satisfies the above equation, let us assume
lieve that our work is new in three respects: a solution of the form , where . Then,
• When time samples are collected uniformly in time from , , , and by plugging
the sensors, we quantify the accuracy of the estimated in these expressions into (1) we get .
mode shapes in terms of the sampling rate, total obser- Since this needs to be satisfied for all , it must be that
vation time, and spacing of the modal frequencies. This , and it is easy to see that . In the
provides finite (non-asymptotic) sample bounds for the structural dynamics community it is customary to rewrite this as
SVD/POD technique that, to the best of our knowledge,
have not appeared in the literature. (2)
• When time samples are collected randomly in time from
the sensors, we quantify the accuracy of the estimated where and represent the natural frequency
mode shapes in terms of the number of samples, total and damping ratio, respectively. As we can see, the natural fre-
observation time, and spacing of the modal frequencies. In quencies and damping ratios will always be positive, and de-
some cases, we show that mode shape estimates obtained pending on the value of , may be real or complex and there
using a given number of random samples can be more may be one or two possible solutions.
1658 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

In this paper, we restrict ourselves to the case when there is solve the above problem, as in the conventional eigenvalue
no damping, i.e., , and thus . From (2) we can problem, one starts off by computing the generalized eigen-
see that when , we have two purely imaginary solutions value that satisfies
, and . Thus, , and
are both eligible solutions to (1). In fact, any solution (5)
to the above equation can be expressed as a linear combination
Assuming that this does not vanish as a function of and that
of and , such that the general solution can be written
is full rank, the left hand side of (5) will represent an th
as
order polynomial and there will be generalized eigenvalues
as roots of this polynomial. For simplicity in this paper, we
assume these generalized eigenvalues are distinct.4 Each
when plugged back in to (4) will have a corresponding
Furthermore, noting that and , generalized eigenvector . Thus, there will be general-
it follows that . ized eigenvalues , which are known as the modal
Since we want to deal with real valued signals , this de- frequencies, and corresponding generalized eigenvectors
mands that . Denoting , we can once again , which are known as the mode shape vectors.
rewrite the solution as . Fi- We also assume that the mass matrix is proportional to
nally, because any linear combination of sines and cosines with the identity (i.e., that all masses are equal). If this is not the
the same frequency is also a sine wave with the same frequency, case, each displacement signal can be normalized by the square
we can rewrite this as , where root of the mass at that node (see [15], [16] for details); this
yields an equivalent set of dynamic equations having a mass
, matrix equal to the identity. Under this assumption, each modal
frequency will be real and positive, and the mode shape vectors
. will be orthonormal to one another. The orthonormality of the
(3) true mode shape vectors plays an important role in our analysis.
It is clear that each , , will be a
B. Multiple-Degree-of-Freedom System
valid solution to the MDOF system equation. Furthermore, as in
Similar to the SDOF system, an -degree MDOF system3 the SDOF system will also be a valid solution
can be formulated as and thus for each a complete solution will be of the form
. We can guarantee this solution is
real by ensuring that is real (there exists a real eigenvector
that satisfies (4) given that the mass and stiffness matrices are
where is an diagonal mass matrix, is a symmetric
real and symmetric) and . Thus, as with the SDOF case
damping matrix, is an symmetric stiffness
we can rewrite this solution as , where
matrix, and is an 1 vector of displacement signals.
and are as defined in (3). Finally, it is easy to see that all
Note that , and each
linear combinations of these solutions are also valid solutions
, , is a displacement signal. One can view
to the MDOF system equation and thus the general solution is
as the signal being observed at the th sensor node.
of the form
Again we consider an undamped system and set .
This simplifies the above equation to
. Let us assume to be a solution (6)
to this equation, where . Here, is an 1 spa-
tial vector that is independent of time; we define it to have unit In the structural dynamics community this is known as the
energy, i.e., (we can assume this as the normal- modal superposition equation.
ization can be absorbed into the scalar variable ). Plugging
in the appropriate derivative to the above expression we get III. PROBLEM FORMULATION
. Since this must hold for
all values of it must be that A. The Analytic Signal of
Our framework and analysis will involve sampling what is
(4)
known as the analytic signal of [17].
The above represents a generalized eigenvalue problem and Definition 1 (Definition 1.2.1, [17]): A signal is said to
our objective is to find pairs of and that satisfy this be analytic iff , where is the Fourier
equation. Notice the similarity to the conventional eigenvalue transform of .
problem, which corresponds to the case when . To An analytic signal can be obtained by removing all negative
frequencies in a given signal. The analytic signal is a frequently
3Theoretically, a structure will have infinitely many degrees of freedom.
However, the number of mode shapes that one can detect is equal to the number 4A structure with a repeated modal frequency may have an entire subspace
of sensors placed on the structure. In the following, whenever we deal with an of mode shape vectors associated to that frequency; an interesting question for
-degree MDOF system, we are implicitly assuming that we have sensor future work would be how well these subspaces can be recovered using the
nodes deployed on the structure. techniques we propose in this paper.
PARK et al.: MODAL ANALYSIS WITH COMPRESSIVE MEASUREMENTS 1659

used representation in mathematics, signal processing, and com- denotes a diagonal matrix, and
munications; in some problems (such as ours) it can simplify the
mathematical manipulations.
To discuss the analytic signal of , let us examine ..
each entry in , i.e., . Based on the derivation in .
Section II-B, each can be written as
denotes a 1 modal coordinate vector.
In order to see how the SVD could be useful in extracting
the modal parameters, let us suppose that we sample each com-
ponent of at distinct points in time . We
Thus, the analytic signal of , represented as , is simply assume and denote the resulting data matrix as

.. .. .. .. (8)
. . . .
and the analytic signal of the entire vector , denoted as
, can be written as The sampling of at implies the sampling of
at the exact same time points which leads us to define
(7)

.. .. .. ..
Note that is no longer real but complex. . . . .
Obtaining an analytic signal in practice involves the applica-
tion of a Hilbert transform. However, detailed discussion of this
and allows us to write the matrix of samples as
matter is out of scope of this paper and we will refer interested
readers to [18] for more information. For the remainder of this (9)
paper, we will assume that we have successfully extracted the
analytic signal from each . Thus, all derivations from here Equation (9) makes explicit the relationship between the SVD
onwards will be in terms of . and the modal parameters. We know that is a square matrix
with orthonormal columns, and is a diagonal matrix. Hy-
pothetically, if happened to be a matrix with orthonormal
B. The Relationship to the SVD
rows, then (9) would precisely describe the SVD of . (Tech-
If certain modal frequencies have not been excited, it is pos- nically, in a valid SVD the entries of the diagonal matrix must
sible that some of the complex amplitudes may be zero. We be real-valued and positive; however, any nonzero phases in the
let denote the number of nonzero amplitudes. While in gen- entries of could be absorbed into the corresponding rows of
eral may be as large as , in some cases may be smaller .) In that case, one could obtain the mode shape vectors
than and it will not be possible to estimate the modal pa- simply by computing the left singular vectors of ! To be clear,
rameters corresponding to the zero amplitudes. Without loss of this statement is only true up to the uniqueness of the SVD. Any
generality, we also assume the complex amplitudes are ordered left singular vector of can be multiplied by a phase rotation
from large to small, and thus we will have term (for any value of ) as long as the corresponding right
. (We do not assume any particular or- singular vector is multiplied by the same phase rotation term,
dering to the modal frequencies.) and the resulting SVD will still be valid. Similarly, however,
We can write the modal superposition (7) in matrix-vector mode shape vectors arise as vectors in the nullspace of
multiplication format as a particular matrix (recall (4)) and are also unique only up to
multiplication by a phase rotation term, although we have noted
in Section II-B that a real solution exists under certain condi-
tions. For these reasons, our accuracy bounds on the estimation
where denotes an matrix of are stated in terms of the absolute correlation between
formed from the mode shape vectors corresponding to the and its estimate; this quantity is invariant to any phase ro-
nonzero amplitudes (note that the columns of are or- tation of either vector.
thonormal due to our assumption that the mass matrix is As an example, the rows of would be perfectly orthogonal
proportional to the identity), if they happened to equal distinct length- DFT vectors.
This could be ensured if we sampled at uniform times
, where is a sampling interval, ,
and , but it would require the modal frequencies to lie
.. .. .. .. on a grid such that . If
. . . .
these conditions were satisfied, the SVD of would exactly
1660 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

recover the mode shape vectors up to a possible phase rotation. sampling, and uniform time sampling followed by a random
Unfortunately, this is an unrealistic model for the purpose of matrix multiplication—and for each measurement scheme we
modal analysis because the modal frequencies will typically not provide a sufficient condition for the accurate recovery of mode
lie on a grid. shapes via the SVD. Proofs of all of our results appear in the
If we drop the assumption that the modal frequencies lie on a Appendix.
grid, the problem becomes much more complicated, and in gen- In the results that follow, we use and to de-
eral the rows of will not be orthogonal. However, our main note lower and upper bounds on the minimum and maximum
results, presented in Section IV, rely on characterizing sampling separation of the modal frequencies. In other words, we as-
strategies that ensure the rows of will still be nearly orthog- sume that , where the minimum is taken
onal and showing in these situations that the mode shape vectors over all with , and that
can be accurately estimated by computing the SVD of . , where the maximum is also taken over all
. Furthermore, we use to denote the
C. SVD in Modal Analysis total sampling time span. Finally for , the quantity
is
Among the many techniques that have been proposed for known as the binary information divergence, or the Kullback-
modal analysis, we briefly mention a few that also make use of Leibler divergence [21].
the SVD. The Ibrahim Time Domain (ITD) [19] method shares
a number of similarities with the SVD-based estimation method
we study. This method also begins with the MDOF model and
A. Uniform Time Sampling and Random Time Sampling
sets up an equation that relates the modal parameters to the ob-
servations akin to (6). After further algebraic manipulations, a
Our proposed method for recovering the mode shape vectors
set of equations reveals that one can extract the modal param-
from uniform time samples or from random time samples is very
eters via an eigendecompsition. Based on this observation, the
simple and is described in Algorithm 1.
ITD method obtains estimates of the modal parameters by com-
puting the eigendecomposition of a matrix that is a function of
the observed data matrix. In the Frequency Domain Decompo- Algorithm 1: Pseudo-code for mode shape estimation
sition (FDD) [13] method, the observations are used to compute Input: Data matrix as defined in (8)
cross power spectral density estimates. These cross power spec- Output: (left singular vectors of )
tral density estimates are collectively a 3D data cube that con-
sists of a 2D cross spectral matrix at each frequency. Given these
estimates, the SVD is used to extract the singular vectors of the
In particular, one simply computes the SVD of
2D cross spectral matrices. These singular vectors provide esti-
and returns the truncated matrix of left singular vectors
mates of the mode shape vectors. The Eigensystem Realization
as estimates of the true mode shape
Algorithm (ERA) [20] method makes use of ideas in control
matrix . (Although the full-size SVD
theory and sets up the problem with a state-space equation of an
of the matrix will return left singular vectors,
MDOF system. The main use of the SVD in this method is to
we use only the first of these as estimates of the mode
decompose the Hankel matrix that can be constructed from mea-
shape vectors.) As we discussed in Section I-D, the use of the
sured impulse response data. The singular vectors and singular
SVD for estimating mode shapes has also appeared under the
values of the Hankel matrix are then manipulated in order to
name of the Proper Orthogonal Decomposition (POD) in the
form a system matrix describing the underlying system. The es-
structural dynamics literature [15], [16]. However, to the best
timates of the modal parameters are then computed via an eigen-
of our knowledge, the POD has not previously been considered
decomposition on this system matrix.
in the context of CS, and the finite sample accuracy bounds we
All of these methods implicitly assume that the observations
present throughout the following sections are all new.
are sampled at uniform intervals at or above the Nyquist rate.
1) Uniform Time Sampling: Let us now suppose that we
Thus, these methods may not be directly applicable when the
sample uniformly in time with a uniform sampling in-
observations are sampled in a random fashion. Furthermore, to
terval denoted by . The sampling times are given by
the best of our knowledge none of these methods are accompa-
, . We are therefore sampling within
nied by error analysis or instructions on how long to sample the
the time span , where . We can es-
vibration signal. In the next section, we present our proposed
tablish the following theorem.
measurement schemes along with detailed analysis providing
Theorem 2: Let be as given in (9) (de-
error bounds and sufficient conditions on how to sample in order
scribing an -degree-of-freedom with no damping, equal
to guarantee faithful recovery of the mode shape vectors using
masses, and nonzero amplitudes in free decay) sampled
the SVD.
according to the uniform sampling scheme described above.
For , suppose we sample for a total time span of at
IV. MAIN RESULTS least

In this section, we present our main results. We propose three

measurement schemes—uniform time sampling, random time (10)
PARK et al.: MODAL ANALYSIS WITH COMPRESSIVE MEASUREMENTS 1661

with sampling interval and ensure that . larger than . This suggests that for the purpose of mode shape
Then, the mode shape estimates obtained via Algorithm 1 extraction, we can potentially sample at a rate lower than the
satisfy Nyquist rate and still accurately recover the mode shapes. What
we are describing here is analogous to how one can sample a
bandpass signal below the naive Nyquist sampling rate; the rel-
evant measure of the signal complexity is the width of the oc-
cupied spectral band, not the absolute highest frequency. In our
(11) case, it is important to note that in order to sample with the pre-
for , where scribed interval we must know in advance the maximum sep-
aration between modal frequencies. In scenarios where is
unknown it would be more reasonable to sample at a sufficiently
small interval to ensure the sampling conditions are
satisfied.
In (11), note that because and are normalized to Finally, recall that . If we set
have unit magnitude, a correlation indi- (the largest permitted by Theorem 2), then in order to satisfy the
cates a perfect match up to a possible phase rotation of either lower bound on given in (10), we would need to collect at
vector. One can convert (11) into an distance bound between least
and but must account for this phase ambiguity. In
particular, for any value of , one can write (12)

total uniform time samples. If we were to reduce the sampling

interval we would need to take more total samples to cover
the same time span . Finally, note that the necessary con-
dition (12) on is fairly satisfactory in its logarithmic depen-
dence on and its linear dependence on (this assumes the
left hand term dominates in (12)). However, it also scales with
For whatever is specified as the “true” mode shape vector
the ratio , which depends on the structure. For some struc-
and whatever is returned by the SVD of , there ex- tures this ratio could in fact be large, and in the absence of addi-
ists a such that tional information about the structure, one may need to assume
, which is the quantity that is bounded in (11). this ratio is large. This motivates our second sampling strategy,
Therefore, if is close to 1, some phase rotated which appears below.
version of is guaranteed to be close to in distance. 2) Random Time Sampling: Let us now suppose that we
In (11), we also see that the error in the th estimated mode sample at time points chosen uniformly
shape vector mainly depends on and what is essentially the at random in the time interval . We do not assume that
minimum separation between and all other . The vari- these sample times are ordered (i.e., that ),
able controls how close the rows of are to being orthog- although in practice there is no harm in generating them ran-
onal; a small choice of implies more orthogonal rows and leads domly and then sorting them. We can establish the following
to a better preservation of the mode shapes but requires more theorem.
samples. Furthermore, the bigger the separation between Theorem 3: Let be as given in (9) (de-
and all other , the better our estimate. Note that the ampli- scribing an -degree-of-freedom with no damping, equal
tudes are dependent on the underlying structure and on the masses, and nonzero amplitudes in free decay) sampled
input that excited the structure prior to free vibration observa- according to the random sampling scheme described above.
tion period, and thus these parameters may be beyond our con- Suppose we sample for a total time span of at least
trol. In order to guarantee a small error in the th mode shape
when is close to some other , one would need to make (13)
smaller.
Turning our attention to the sampling parameters, the above and within this time span suppose we take a number of measure-
theorem essentially tells us that we need to sample for a time ments satisfying
span that is inversely proportional to the minimum spacing be-
tween the modal frequencies. Thus, the smaller the minimum (14)
spacing between the modal frequencies, the longer we must
sample to get an accurate estimate. Also, since the theorem re-
where
quires that , the maximum spacing between modal
frequencies determines how fast we need to sample. Comparing
this sampling interval to the Nyquist sampling interval,5 which
would be , it is interesting to note that can be
5This is the Nyquist sampling interval for the analytic signal of ; the Then with probability at least all of the mode shape esti-
Nyquist sampling interval for itself would be half as long. mates obtained via Algorithm 1 will satisfy (11).
1662 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

This result for random time sampling looks somewhat similar where represents the distortion factor of (see Definition
to Theorem 2 for uniform time sampling. The recovery guar- 4). Let and let denote the left singular vectors
antee is the same and the required time span differs only by of . Then, with probability exceeding ,
a constant. A critical difference, however, is that the requisite
number of samples no longer depends on the ratio .
However, when is large and is small, the denominator in
(14) will scale like , and so the requisite number of samples (15)
will scale like . This represents a stronger dependence for , where
on compared to what appears in (12), but only by a loga-
rithmic factor (because the right hand term in (12) scales like
). It also represents a stronger dependence on compared
to what appears in (12). Ultimately, we see that in some cases and are the singular values of .
random time sampling could provide a significant reduction in The theorem above prescribes a condition under which the
the number of samples for systems where is large or un- th left singular vector of the compressed matrix will
known. For a given problem, the better choice between uniform be highly correlated with the th left singular vector of
and random time sampling may depend on the particular cir- the original uniformly sampled data matrix . Assuming
cumstances and the parameters of the system under study.
is generated following the conditions required by Theorem 2,
the bound (11) will guarantee that is highly correlated
B. Uniform Sampling Followed by Random Matrix
with the true mode shape vector .6 We note that if both
Multiplication
and are close to 1, one can guar-
The last measurement scheme that we consider involves antee that is also close to 1. Unfortunately, the
taking uniform time samples and compressing these via mul- “triangle inequality” for correlations is not as simple to state as
tiplication by a random matrix. More specifically, let us form the one for distances, and so we defer a statement of this in-
following the uniform sampling scheme as discussed in equality to Appendix D.
Section IV-A.1 with and satisfying the constraints listed In Theorem 5, the required number of columns in the random
in Theorem 2. Subsequently, we construct a random matrix is again dependent on the number of nonzero ampli-
matrix and compute the matrix of tudes . For the purpose of computing a bound, in the worse
compressed measurements. We are specifically interested in case one could assume . Theorem 5 would be particu-
cases where , i.e., when has fewer columns than larly useful if happened to be generated with much larger
. We also note that can be applied individually to each than is required by Theorem 2. For example, suppose we are
row of and the resulting measurements can be concatenated dealing with a structure that has high and small (or
to form . This means that this CS measurement scheme can suppose that we do not know these quantities and so we conser-
be performed sensor-by-sensor in a SHM system. vatively suppose they are large and small, respectively). This
To state our results, we write the truncated SVD of anal- means that we need to take a large number of uniform sam-
ogously to that of as . The matrices ples using a small sampling interval over a long time duration
, and will be , , and . In such scenarios, one could choose to post-process the
, respectively. We also require the following definition. signals to reduce the number of measurements by multiplying
Definition 4: A random matrix is said to satisfy each sample vector with . Another similar scenario is when
the distributional JL property if for any fixed and we have a conventional uniform sensor over which we do not
any distortion factor , have control over the sampling interval. Again, we may use
as a way to reduce the number of measurements. In Section V-A,
we demonstrate the usefulness of this sampling scheme with a
synthetic example.
where is a constant depending only on .
For most random matrices satisfying the distributional JL V. EXPERIMENTAL RESULTS
property, the functional dependence on , , is quadratic in
as . There are a variety of random matrix construc- A. Experiments With Synthetic Data
tions known to possess the distributional JL property. Notably,
random matrices populated with independent and identically We begin by demonstrating the effectiveness of the various
distributed (i.i.d.) subgaussian entries will possess this property methods using an idealized synthetic dataset. The system
that we consider is illustrated in Fig. 2. It is a simple 4-de-
[22]. Subgaussian distributions include suitably scaled Gaussian
gree-of-freedom system consisting of four objects of masses
and Bernoulli random variables.
arranged in a line. Springs connect each adjacent
Our next theorem follows from [23, Theorem 1].
pair of masses, where the springs have stiffness values of
Theorem 5: Suppose is an matrix having rank
. We do not consider damping in this system; thus
and left singular vectors , and . For the following simulation, we simulate the above
suppose . Let represent an random matrix system under free vibration with , i.e.,
that satisfies the distributional JL property with , and with , , , ,
6Technically, one could also apply this to a data matrix generated with
random time sampling if the conditions of Theorem 3 were met.
PARK et al.: MODAL ANALYSIS WITH COMPRESSIVE MEASUREMENTS 1663

Fig. 1. Absolute inner product of ground truth mode shapes and mode shape estimates using uniform (black) and random (light blue) time sampling schemes.
The four subfigures correspond to the four mode shapes, and the results are plotted as a function of the total sampling time . (a) Results using an initial set of
modal frequencies with rad/s. (b) Results using a set of modal frequencies with smaller minimum separation of rad/s. These figures
illustrate that the mode shape estimates via uniform and random sampling schemes perform comparatively, and the effect has on for accurate recovery
of mode shape vectors.

In our first experiment, we demonstrate the uniform and

random time sampling methods by plotting the absolute
inner-product of the true and the estimated mode shapes ob-
Fig. 2. A simple 4-degree-of-freedom system made up of 4 box cars each of
mass . The box cars are connected via springs of stiffness tained from the SVD of the respective sampled matrix .
. In the synthetic simulations, we simulate this particular system with For the uniform time sampling scheme, we use a sampling
, and , , , interval of . This is just slightly faster than ,
, . which is the upper bound of what our theorem prescribes.
Using this fixed rate, we collect samples over a total time span
and . The stiffness matrix corresponding to the of duration , and we repeat the experiment for values of
above system can be written as (the value of therefore increases with
). For each value of and for , we plot
the absolute inner product between the ground
truth mode shape vector and the corresponding estimated
vector . The results are shown in the black curves in
Fig. 1(a). For the random time sampling scheme, we use the
same values of and sample uniformly at random
within the interval . For each value of , the total
number of samples we collect is chosen to equal the corre-
sponding value of used for uniform time sampling above.
Since , the mode shape vectors are simply the The absolute inner product value of the true and the estimated
eigenvectors of and we use these mode shapes throughout mode shapes are shown in the light blue curves in Fig. 1(a). We
this subsection. The modal frequencies are simply the square see that overall, the performance of the two sampling schemes
roots of the eigenvalues of , which are , is comparable. This is in agreement with Theorems 2 and 3,
, , rad/s. We also use the fol- as they both suggest the same reconstruction guarantees given
lowing matrix of amplitudes throughout this subsection: that we satisfy the sampling conditions.
Our second experiment highlights the role played by
(the minimum separation of the modal frequencies). We con-
sider two distinct systems. The first system is the same as above.
We simulate a second system that uses the exact same mode
shape vectors and matrix as above but has a slightly different
Collecting everything together, the modal superposition equa- set of modal frequencies, such that , ,
tion can be written as , rad/s.7 Note that in the
7Note that this system may not have the same setup as the one illustrated in
Fig. 2. However, in order to focus on the effect has on the mode shape
recovery we have developed this hypothetical scenario where we keep all pa-
rameters fixed except in order to achieve a smaller . Similarly, in the
simulations that follow, where necessary, we will change the modal frequencies
but always employ the same mode shapes and matrix.
1664 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

Fig. 4. The relative layout of 18 vibration sensors installed on a bridge

in Ypsilanti, MI. From these sensors, we acquire vibration signals
(18).

time span of . For these modal frequencies, in order

to avoid aliasing, the uniform sampling interval would need
to satisfy . We first obtain a data matrix
by sampling with (a sub-Nyquist rate) over the
total sampling time span . This gives us in total
64 samples. We then obtain a second data matrix by sampling
with (a super-Nyquist rate) and then multiplying
the length-2001 sample vector by a random Gaussian matrix
to produce measurements. We compute the left
Fig. 3. Absolute inner product of ground truth mode shapes and mode shape singular vectors for each of these data matrices to estimate the
estimates using uniform (black) and random (light blue) time sampling schemes.
The results are plotted as a function of the total number of samples . For each
mode shapes. For the first matrix, due to severe aliasing we
value of , the for random sampling is chosen to be 3 seconds longer see absolute inner product values of ,
than the for uniform sampling. , , and
. For the second matrix, we
first system, and in the second system. Based on see absolute inner products of ,
Theorems 2 and 3, we anticipate the need to sample for a longer , , and
time span (larger ) when is smaller. The absolute inner . These results illustrate the utility of
product values using the second set of modal frequencies are random matrix multiplication for reducing the dimensionality
plotted in Fig. 1(b). Comparing to the results from the first set, of a uniform sample vector.
we see that a longer sampling duration is indeed needed to
achieve comparable accuracy in estimating the mode shapes. B. Experiments With Real Data
For uniform time sampling, when is fixed, increasing We conclude by presenting simulation results using vibration
will automatically increase the number of samples to be ac- data collected from a bridge in Ypsilanti, MI. On this bridge,
quired. For random time sampling, however, and can be there are wireless nodes, each of which is equipped
chosen independently of one another. While in Fig. 1 for each
with an accelerometer. The relative layout of the sensors is
we have always used the same value of for random sam-
shown in Fig. 4. Each sensor measures acceleration data and
pling as we used for uniform sampling, this is not actually neces-
sary. In fact, our Theorem 3 suggests that, for random sampling transmits it to the central node for analysis. Because this data
when is small, we can increase without increasing contains noise and is collected from a real bridge—which does
the number of samples . To demonstrate this, we conduct a have damping, is not necessarily in free vibration mode, and
third experiment, and for this we use the second (more closely has unknown and unequal node masses (to which we have not
spaced) set of modal frequencies above. For several values of normalized the displacement signals)—we feel this represents
, we collect samples using both uniform time sampling an interesting test for our methods when the assumptions of our
(for which will be determined by ) and random time current theoretical results are violated.
sampling. For each value of with random time sampling, The data that is available to us from each sensor is a set of
however, we choose to be 3 s longer than the value of real-valued collected uniformly in time
used for uniform sampling with the same value of . The re- at a rate faster than the Nyquist rate. We stack this data into
sults are shown in Fig. 3. We see that simply by increasing a real-valued data matrix we call (note that we do not as-
without affecting , the random time sampling scheme can ac- sume samples of the analytic signals are available), and we
commodate the decreased value of . test the effectiveness of multiplying by a random matrix
To motivate our fourth experiment, consider a scenario and then computing the SVD of the compressed matrix
where we are limited in the number of samples we can transmit . For a point of comparison, we test a method that
to the central data repository. If our sensor is limited to col- is similar to one presented in [11], [12]. That method, which we
lecting uniform time samples, then we may not be able to refer to as “CS+FDD,” involves reconstructing each length-
sample and transmit at a fast enough rate to avoid aliasing and
signal from the random Gaussian measurements
accurately recover the mode shape vectors. In a scenario such
. These signals are reconstructed one-by-one and then
as this, one way to improve the performance would be to first
sample uniformly at a high rate but then to multiply the high fed into the Frequency Domain Decomposition (FDD) method
rate sample vector by a random compressive matrix so that [13] for modal analysis. The reconstruction of each is ac-
the amount of transmitted data is reduced. To illustrate this, complished by solving
we set the modal frequencies to be , ,
, and rad/s, and we sample for a total
PARK et al.: MODAL ANALYSIS WITH COMPRESSIVE MEASUREMENTS 1665

Fig. 5. Comparison of FDD on the original data matrix with CS+FDD and our proposed method SVD( ). The results were evaluated by computing the absolute
inner product between each mode shape returned by FDD and the corresponding estimated mode shapes via CS+FDD and SVD( ). The results were as follows:
(a) CS+FDD: 0.94, SVD( ): 0.99 (b) CS+FDD: 0.52, SVD( ): 0.99, and (c) CS+FDD: 0.87, SVD( ): 0.97.

where represents a DWT matrix (we saw similar results Finally, we note that while our focus in this paper has been
with the DFT), and we let the reconstructed . on recovering mode shape vectors, one can also consider the
For both the proposed SVD method and the CS+FDD problem of estimating modal frequencies. In fact, the SVD of
method, we take measurements of each using a data matrix can also be used for this purpose. Referring
the same measurement matrix for all . Since we do not to (9), we see that each row of contains a sampled com-
know the true mode shapes of the structure, we use as a bench- plex exponential corresponding to one of the modal frequen-
mark the three dominant mode shapes returned when FDD cies. Supposing the time samples are collected uniformly, one
is applied to the original uncompressed data matrix . The simple way to extract each frequency is to compute the Fourier
results are presented in Fig. 5. As we can see, the mode shapes transform for each row of and to identify the frequency
estimated using CS+FDD (plotted in red) are not particularly with maximum magnitude.8 (Recall that each row of cor-
close to the mode shapes returned when FDD is applied on responds to a right singular vector of .) This technique has
the original data (plotted in black). This is apparently because been successfully applied in the structural dynamics literature
it is difficult to accurately reconstruct any individual signal [24]. Building on our results for estimating mode shape vectors
from just 50 random measurements. In contrast, using the SVD, our aim for future work is to provide a rigorous
however, when we apply the SVD to the compressed data analysis of the problem of estimating modal frequencies in CS
matrix , the estimated mode shapes (plotted in light blue) settings, including scenarios where a random matrix multiplica-
provide much better approximations to the true FDD mode tion is applied to uniform time samples.
shapes.
APPENDIX
VI. CONCLUSIONS In this appendix we provide proofs of the main results. To do
In this paper, we have provided a rigorous analysis of so, we take a perturbation theoretic viewpoint. We first describe
a simple SVD-based technique for estimating mode shape how we can formulate our problem as a perturbation problem
vectors in three different measurement regimes: one where and then provide separate proofs for each theorem.
uniform time samples are collected of the displacement signals,
one where random time samples are collected, and one where A. Perturbation Analysis
a compressive matrix multiplication is applied to each vector We start with the equation , and we allow the
of uniform time samples. Our theoretical results quantify the sample times to be arbitrary. To carry out perturba-
tradeoffs between these three measurement strategies, and ex- tion analysis let us note that ,
perimental results on synthetic data (in Section V-A) illustrate where is an matrix containing the true mode shape
the effectiveness of the various methods. vectors as orthonormal columns, and is a matrix
Our simulation results on real data (in Section V-B) support with entries
our theoretical results that the SVD of a compressed data ma-
trix can return accurate estimates of the true mode shape ,
vectors. We emphasize again that the dataset in this simulation .
is real-valued, contains noise, and is collected from a real struc-
ture with damping; technically, none of this is covered by the
assumptions of our current theoretical results. The fact that the Thus, we can decompose this product as ,
SVD-based method was nevertheless able to successfully esti- where contains the off-diagonal entries of . Then
mate the mode shape vectors is very encouraging and suggests
that our theoretical findings may be extendable to more com- (16)
plicated scenarios. We plan to explore such scenarios in future 8In scenarios where the samples are collected randomly, one could employ a
work. nonuniform FFT.
1666 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

The above expression allows us to view as the summa- Let us examine the term which appears in (20) and is defined
tion of an “original” matrix and a “perturbation” in (18). Note that
matrix . Noting that the eigenvectors of the
original matrix are given by , our goal is to
show that the eigenvectors of (which equal the left sin-
gular vectors of ) are close to those of this original matrix.
At this point, it is helpful to define

(17)
In the second to last line we let and
note that . Our Theorems 2 and 3 rely on
proving that . We prove this fact first in the context
In this equation, we can view as the sum of an original ma-
of random sampling (in Appendix B) and then in the context of
trix and a perturbation .
uniform sampling (in Appendix C).
Recall that we write the SVD of as . Since the
Once we have established that , we can plug
rank of must be or less—and in fact we explain below
this inequality into (20) and obtain the bound
conditions under which it is guaranteed to be equal to —we
can consider the truncated SVD of where has size
. Therefore, note that
(21)
We note that (21) depends both on the original eigenvalues of
(appearing as ) and on the eigenvalues of the
Since the column spans of and are the same, is perturbed matrix (appearing as ). We can relate
a orthonormal matrix that contains the eigenvectors of these two collections of eigenvalues as follows. First, note that
. The following theorem allows us to relate the eigenvectors the eigenvalues of are the same as those of . Further-
of to those of . more, the eigenvalues of are the same as those of .
Theorem 6 (Theorem 1, [25]): Let and We focus on relating the eigenvalues of to those of .
be positive definite Recall that we can write . Thus, for
matrices. Assume that and are unitary and that any ,
and are diagonal.
Let , and assume

(18)
Let us denote and use this to write
where . Then for any and for any
set T not containing we have,

(19) If we can show that , this means

T
T , giving us
We apply Theorem 6 with and as defined in (17),
allowing us to take , , and
. Since and have or-
Applying Lemma 1 in [26] allows us to conclude that
thonormal columns and identical column spans, for any fixed
, we will have .
(22)
Therefore, assuming (19) holds, we will have
for . We note that, in order to apply Theorem 6
above, it is necessary that and be positive definite. The
requirement that be positive definite will be satisfied by our
assumption that the amplitudes are nonzero. The
requirement that be positive definite will then be satisfied
as a consequence of (22). Note that this latter requirement can
(20) only be satisfied if .
PARK et al.: MODAL ANALYSIS WITH COMPRESSIVE MEASUREMENTS 1667

Supposing that (22) is satisfied, let us examine the term in the above theorem as . As a result,
appearing in (21). Recall that is fixed. For a given , will be i.i.d. positive semi-definite matrices, i.e., ,
an upper bound on this expression is given by of rank 1 with . We wish to
compute

Plugging this bound into (21) and using the fact that
for , we obtain the inner product bound
appearing in (11).

B. Proof of Theorem 3 (Random Sampling) or the appropriate lower and upper bound on the above quanti-
To compute an upper bound for , let us note that ties. Note that
, where we use
to denote the th largest eigenvalue of . If we can
find upper and lower bounds on the eigenvalues of ,
such that holds for all , then we can
see that , which in turn implies ,
that . .
Therefore, our strategy is to bound from below
and above in order to bound . In this section, we estab- The eigenvalues can also be written as
lish such a bound for the case where are chosen , where is the off-diag-
uniformly at random in the time interval . onal matrix of . According to Gershgorin’s circle
Theorem 7: Given that we sample with satisfying (13) theorem [27] we know that every eigenvalue of must lie
and satisfying (14), then with probability at least we within at least one Gershgorin disk. As has zero diagonal
will have for all . entries, every Gershgorin disk must be centered at zero. Thus,
Proof: To bound we use a slightly modified the radius of the largest disk will provide a bound on all eigen-
version of a theorem that appeared in [21]. values of . It follows that every eigenvalue of will
Theorem 8: [[21], Theorem 5.1] Consider a sequence obey the following bound:
of independent, -dimensional,
random, self-adjoint matrices that satisfy and
almost surely. Then for any and
such that

we have (25)

where we have used to denote the index of the median modal

frequency among (when is even we can let de-
(23)
note the index of either the th or st largest modal
for and
frequency), and we have used the fact that if the modal frequen-
cies were to be ordered, any adjacent pair would be separated
by at least . The summation term in the above bound is also
(24) known as the Harmonic number. We can simplify the above ex-
for . pression by using the following bound on the Harmonic number.
In order to apply Theorem 8, let us write Theorem 9 ([28], Theorem 1): For any natural number
, the following inequality is valid:

..
.
The constant is known as Euler’s constant. The
We define the vector , constants and are the best possible, and
where . Let us set the matrix that appears equality holds only for .
1668 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

Applying this theorem to (25), we have when . The fraction of sinusoids in the above equation
is known as the periodic sinc function or the Dirichlet function
and is defined as . More specifically,

,
, .

As its name implies, the function is a periodic function

where the period is equal to when is odd, and when
is even. Every time is equal to an integer multiple of ,
Collecting everything together, we will have for all , the function will reach its maximum value.
Again, we bound the eigenvalues of using Gershgorin’s
disk theorem. Since every Gershgorin disk will again be cen-
tered at zero, every eigenvalue of must obey the following
bound:

(26)

We can guarantee the evaluation of the function to be

Supposing that (13) is satisfied, we have that small by restricting ourselves to only certain values of . We
compute an upper bound on the function by noting that
the denominator can be lower bounded by a linear function for
a certain range of . In particular, for we have
Note that and . Then, according to the that . Applying this to (26), we
above theorem, inequality (24) will hold for any have
, which will always include . Similarly, inequality
(23) will hold for any , which will always
include . Substituting the appropriate values of
, , and into Theorem 8, with probability at least
we will have

By choosing to satisfy (14), we can ensure both that

and that , and therefore the when for all and . In the last line we have used
overall failure probability will be less than . the fact that . To ensure that for
all and , we choose the sampling interval such that .
C. Proof of Theorem 2 (Uniform Sampling) Following the same arguments we used in the proof of Theorem
3, this will give us
For the uniform sampling scenario, our strategy is again to
bound from below and above in order to bound
.
Theorem 10: Suppose we sample with a total time span
satisfying (10) with sampling interval and ensure
that . Then we establish the following bound on the
eigenvalues of : . Now, note that and if we choose so that
Proof: The off-diagonal matrix of , denoted as ,
has the following entries: when , and

we will have that In summary, when we sample

within a sampling interval satisfying and a time span
satisfying , we will have
PARK et al.: MODAL ANALYSIS WITH COMPRESSIVE MEASUREMENTS 1669

D. “Triangle Inequality” for Correlations [5] S. W. Doebling, C. R. Farrar, M. B. Prime, and D. W. Shevitz, “Damage
identification and health monitoring of structural and mechanical sys-
Consider three unit-norm vectors , , and . We tems from changes in their vibration characteristics: A literature re-
view,” Los Alamos National Lab., Los Alamos, NM, USA, 1996.
would like to bound in terms of and [6] A. K. Pandey, M. Biswas, and M. M. Samman, “Damage detection
. Let denote the orthogonal projection matrix from changes in curvature mode shapes,” J. Sound Vibrat., vol. 145,
no. 2, pp. 321–332, Mar. 1991.
onto the span of . Then we can write [7] J. P. Lynch, “An overview of wireless structural health monitoring for
civil structures,” Philosoph. Trans. Roy. Soc. A, vol. 365, no. 1851, pp.
345–372, 2007.
[8] D. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52,
no. 4, pp. 1289–1306, Apr. 2006.
[9] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles:
Exact signal reconstruction from highly incomplete frequency infor-
mation,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, Feb.
Therefore, 2006.
[10] M. Wakin, S. Becker, E. Nakamura, M. Grant, E. Sovero, D. Ching,
J. Yoo, J. Romberg, A. Emami-Neyestanak, and E. Candés, “A non-
uniform sampler for wideband spectrally-sparse environments,” IEEE
J. Emerg. Sel. Topics Circuits Syst., vol. 2, no. 3, pp. 516–529, Sep.
2012.
[11] D. Mascarenas, D. Hush, J. Theiler, and C. Farrar, “The application
of compressed sensing to detecting damage in structures,” presented at
the 8th Int. Workshop Structural Health Monitor., 2011.
[12] B. Yuequan, J. L. Beck, and L. Hui, “Compressive sampling for ac-
celerometer signals in structural health monitoring,” Structural Health
Monitor., vol. 10, no. 3, pp. 235–246, 2011.
[13] R. Brincker, L. Zhang, and P. Andersen, “Modal identification from
ambient responses using frequency domain decomposition,” presented
at the 18th Int. Modal Anal. Conf., 2000.
[14] D. Needell and J. A. Tropp, “CoSaMP: Iterative signal recovery from
incomplete and inaccurate samples,” Appl. Comput. Harm. Anal., vol.
26, no. 3, pp. 301–321, 2009.
[15] B. F. Feeny and R. Kappagantu, “On the physical interpretation of
proper orthogonal modes in vibrations,” J. Sound Vibration, vol. 211,
no. 4, pp. 607–616, 1998.
(27) [16] G. Kerschen and J.-C. Golinval, “Physical interpretation of the proper
orthogonal modes using the singular value decomposition,” J. Sound
Vibration, vol. 249, no. 5, pp. 849–865, 2002.
where the first inequality above follows from the conventional [17] B. Boashash, Time Frequency Signal Analysis and Processing: A Com-
prehensive Reference. Amsterdam, The Netherlands: Elsevier, 2003.
triangle inequality and the second inequality above uses the [18] F. W. King, Hilbert Transforms. Cambridge, U.K.: Cambridge Univ.
Cauchy-Schwarz inequality. Overall, the inequality (27) can be Press, 2009.
[19] S. R. Ibrahim and E. C. Mikulcik, “A method for the direct identifica-
viewed as a type of triangle inequality for correlations between tion of vibration parameters from the free response,” Shock Vibration
unit-norm vectors. Given lower bounds on and Bull., 1977.
[20] J. N. Juang and R. S. Pappa, “An eigensystem realization algorithm
, one can use (27) to obtain a lower bound on for modal parameter identification and model reduction,” J. Guidance,
. It is also possible to obtain a lower bound using Contr., Dynam., vol. 8, no. 5, pp. 620–627, 1985.
[21] J. A. Tropp, “User-friendly tail bounds for sums of random matrices,”
and , although we do not present this here. Found. Comput. Math., vol. 12, no. 4, pp. 389–434, 2012.
[22] M. Davenport, “Concentration of measure and sub-Gaussian distribu-
tions,” 2009 [Online]. Available: http://cnx.org/content/m32583/latest/
[23] A. C. Gilbert, J. Y. Park, and M. B. Wakin, “Sketched SVD: Recov-
ACKNOWLEDGMENT ering spectral features from compressive measurements,” Arxiv, 2012
[Online]. Available: arXiv:1211.0361, to be published
The authors would like to thank S. O’Connor and [24] B. F. Feeny, “On proper orthogonal coordinates as indicators of modal
activity,” J. Sound Vibration, vol. 255, no. 5, pp. 805–817, 2002.
Prof. J. P. Lynch at the University of Michigan for helpful [25] R. Mathias and K. Veselić, “A relative perturbation bound for positive
discussions on modal analysis and for providing us with real definite matrices,” Linear Algebra Appl., vol. 270, pp. 315–321, 1998.
[26] J. Barlow and J. Demmel, “Computing accurate eigensystems of scaled
datasets to carry out the experiments presented in this paper. diagonally dominant matrices,” Tech. Rep., 1980.
The authors would also like to thank H. J. Cho at the University [27] S. Gerschgorin, “Über die abgrenzung der eigenwerte einer matrix,”
Bulletin de l’Académie des Sci. de l’URSS. Classe des Sciences Math-
of Michigan and Prof. M. Davenport at Georgia Tech for their ématiques et na, pp. 749–754, 1931.
helpful input and the anonymous reviewers for their helpful [28] M. B. Villarino, “Sharp bounds for the harmonic numbers” Arxiv, 2005
[Online]. Available: arXiv:math/0510585, to be published
comments and suggestions.

REFERENCES
[1] Commercial Buildings Energy Consumption Survey
(CBECS)—Overview of Commercial Buildings Characteristics, Jae Young Park received the BEng degree in elec-
[Online]. Available: http://www.eia.gov/consumption/commer- trical and electronic engineering in 2006 from Impe-
cial/data/2003/pdf/a5.pdf rial College London, London, UK. Subsequently, he
[2] American FactFinder [Online]. Available: http://factfinder2.census. obtained both M.S. and Ph.D. degrees in electrical
gov/faces/tableservices/jsf/pages/productview.xhtml?pid=ACS_11_ engineering in 2008 and 2013, respectively, from the
5YR_DP04 University of Michigan, Ann Arbor, MI. Currently,
[3] 2013 Report Card on America’s Infrastructure [Online]. Available: he is a research scientist at Schlumberger, Houston,
http://www.infrastructurereportcard.org/fact-sheet/bridges TX.
[4] H. Sohn, C. R. Farrar, F. Hemez, and J. Czarnecki, “A review of struc-
tural health monitoring literature 1996–2001,” Los Alamos National
Lab., Los Alamos, NM, USA, 2001.
1670 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 62, NO. 7, APRIL 1, 2014

Michael B. Wakin (S’01–M’06–SM’13) received Anna C. Gilbert (M’05) received an S.B. degree
the B.S. degree in electrical engineering and the from the University of Chicago and a Ph.D. from
B.A. degree in mathematics in 2000 (summa cum Princeton University, both in mathematics. In 1997,
laude), the M.S. degree in electrical engineering in she was a postdoctoral fellow at Yale University and
2002, and the Ph.D. degree in electrical engineering AT&T Labs-Research. From 1998 to 2004, she was
in 2007, all from Rice University. He was an NSF a member of technical staff at AT&T Labs-Research
Mathematical Sciences Postdoctoral Research in Florham Park, NJ. Since then she has been with
Fellow at the California Institute of Technology from the Department of Mathematics at the University of
2006–2007 and an Assistant Professor at the Univer- Michigan, where she is now a Professor. She has
sity of Michigan in Ann Arbor from 2007–2008. He received several awards, including a Sloan Research
is now an Associate Professor in the Department of Fellowship (2006), an NSF CAREER award (2006),
Electrical Engineering and Computer Science at the Colorado School of Mines. the National Academy of Sciences Award for Initiatives in Research (2008),
His research interests include sparse, geometric, and manifold-based models the Association of Computing Machinery (ACM) Douglas Engelbart Best
for signal and image processing, approximation, compression, compressive Paper award (2008), the EURASIP Signal Processing Best Paper award (2010),
sensing, and dimensionality reduction. In 2007, Dr. Wakin shared the Hershel a National Academy of Sciences Kavli Fellow (2012), and the SIAM Ralph E.
M. Rich Invention Award from Rice University for the design of a single-pixel Kleinman Prize (2013).
camera based on compressive sensing; in 2008, Dr. Wakin received the DARPA Her research interests include analysis, probability, networking, and algo-
Young Faculty Award for his research in compressive multi-signal processing rithms. She is especially interested in randomized algorithms with applications
for environments such as sensor and camera networks; and in 2012, Dr. Wakin to harmonic analysis, signal and image processing, networking, and massive
received the NSF CAREER Award for research into dimensionality reduction datasets.
techniques for structured data sets. Dr. Wakin is an Associate Editor for IEEE
Signal Processing Letters.