47th AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Confere AIAA 2006-1955

1 - 4 May 2006, Newport, Rhode Island

Damage Detection of Low Energy Impacts

in a Composite UAV Wing

Stephen T. Trickey∗ , Jonathon M. Nichols∗

and Mark Seaver†
US Naval Research Laboratory, Washington, DC 20375, USA

The increased use of composite materials in aerospace platforms necessitates the under-
standing of the particular failure mechanisms of composites (fiber breakage, resin cracking,
delamination) and how they can be detected. Of particular interest in this study is the
effect of low energy impacts on the structural integrity of a composite wing which might
result from rough handling, misuse (such as tool drops) or simply adverse flight conditions.
The novel analysis method used in this work is based on the simple assumption that dam-
age will manifest itself as an increasing nonlinear response of the system. Thus the damage
detection problem becomes a nonlinearity detection problem and a baseline healthy system
is not required. With transfer entropy as our metric, we assess how information propagates
through the system and if the system responds nonlinearly at some level. Nonlinearity is
detected by comparing the transfer entropy of the original experimental data to sets of
surrogate data. The surrogates retain the statistical properties of the original data up to
the second moment (σ 2 ) but discard any higher order correlations (i.e. nonlinearities).
In this way the surrogates serve as the undamaged baseline. Thus differences in transfer
entropy between the original data and the surrogates indicate the presence of nonlinearity
and hence damage. For comparison a subspace system identification method is used to
identify modal frequencies at each damage level.

Nomenclature
p(xi ) probability density function
p(xi , xj ) joint probability density function
p(xi |xi , xj ) conditional probability density function
i, j vector component indices
n, m absolute time indices
Sxi xj cross spectral density
Xi Fourier transform of xi
γ normalized mean separation statistic
σ2 variance
x vector quantity
x(n) ≡ xi (n) multivariate time series

I. Introduction
A number of researchers have applied vibration based techniques to the problem of damage detection
or structural health monitoring (SHM). The typical procedure in vibration-based SHM is to extract certain
features from the time series and then track those features as damage is incurred. By continually comparing
newly acquired features to a baseline set, ideally extracted from a healthy structure, the practitioner seeks to
∗ Mechanical Engineer, Smart Structures Section, 4555 Overlook Ave SW, Washington, DC 20375.
† Research Physicist, Head: Smart Structures Section, 4555 Overlook Ave SW, Washington, DC 20375.

1 of 10

This material is declared a work of the U.S. Government and American Institute
is not subject of protection
to copyright Aeronauticsin theand Astronautics
United States.
make confidence-based judgments as to whether or not the structure is damaged. An effective assessment is
complicated by the feature’s sensitivity to additional phenomenon unrelated to damage such as temperature
or humidity. The primary goal of research in the SHM field is to combine advances in both sensing and data
analysis to produce an automated system capable of detecting damage – reducing or eliminating altogether
the need for visual inspection.
Damage in structures is frequently modeled as the introduction of a nonlinearity into a structure or
structural component that, in its healthy state, is otherwise accurately described by a linear model.4, 5 Such
mechanisms might include the presence of a crack (bi-linear stiffness) post-buckled behavior (Duffing nonlin-
earity) bolt rattling (impacting system with discontinuities) or more general nonlinear stiffening or damping
effects. Modal properties of a structure are often the first features considered for detecting these and many
other types of damage. While modal frequencies are usually not very sensitive to small amounts of damage,
techniques using derived modal properties such as a mode curvature metric have been demonstrated.1, 2 Time
series analysis techniques such as autoregressive methods have also proven effective for assessing damage in
structures.3 In this work we consider an additional time series analysis method that was chosen specifically
to focus on the nonlinear aspect of damage. The method is a measure of time delayed transfer entropy which
is more traditionally associated with research in information theory.
Information theoretics are based on a probabilistic description of the underlying dynamics and can be
used to explore both linear and nonlinear relationships in time series data. When used in conjunction
with surrogate data techniques (to be described), the metric is capable of providing an absolute measure of
nonlinearity. This has the extremely beneficial effect of removing the requirement of a reference or baseline
data set in the damage detection problem. This metric is similarly insensitive to many models of ambient
variability. Temperature fluctuations, for example, are often assumed to affect global stiffness properties
but not necessarily the form of the underlying structural model (linear or nonlinear). A numerical study
demonstrating this facet of the proposed approach has recently been described.6
In the following we present the time delayed transfer entropy metric. We describe the experimental setup
which consists of a composite UAV wing, a shaker and load cell, a pendulum type impactor, and a fiber optic
strain measurement system. We discuss the results of applying the transfer entropy metric to the dynamic
response of the UAV wing after various amounts of impact damage have induced a mild nonlinearity into
the system and we compare the transfer entropy results to identified modal frequency results.

II. Theory
Many of the commonly used approaches in system identification were designed for analyzing linear sys-
tem dynamics. In a statistical sense, these approaches make the assumption that the covariance matrix
captures the necessary dynamical relationships (correlations) among the data. The linear auto and cross-
correlation functions, (and by the Weiner-Khinchine relationship) the auto and cross-spectral densities, and
the frequency response function are defined by second-order statistics. These algorithms comprise traditional
modal analysis and are ideal if the system being studied is accurately described by a linear model. Indeed,
for linear systems, the auto and cross-spectral densities sufficiently describe the dynamical relationships
among the data. For nonlinear systems, where higher-order correlations become important, these tools are
not well-suited. Nonetheless, traditional modal analysis can be adapted to account for nonlinearity (see
Worden and Tomlinson7 for a thorough discussion). For example, if the form of the nonlinearity is known a
priori the practitioner might look for specific frequency domain distortions. Similarly, if baseline data have
been collected with the structure in a known or assumed linear state, subsequently collected data may be
analyzed for the appearance of additional poles in the frequency domain, the assumption being that the
changes are due to the presence of a nonlinearity. Perhaps the most straightforward approach is to apply
variable amplitude loading and check the frequency response for dependencies on the level of excitation. This
approach was employed by Neild et al. in looking for damage in concrete beams.8 For complex structures
an accurate model of the nonlinearity may be difficult to acquire (hence we do not know what nonlinear
feature to expect). Furthermore, many situations call for the practitioner to retro-fit an existing structure
(no baseline data present). Exciting a structure with variable amplitude inputs may pose further practical
challenges.
Instead it would seem advantageous to utilize techniques that are specifically designed to explore both
linear and nonlinear correlations in time series data. The higher-order correlations associated with nonlinear-
ity are accounted for naturally if one defines coupling in terms of the signal’s probability density functions.

2 of 10

American Institute of Aeronautics and Astronautics

The time-delayed transfer entropy employs a probabilistic definition of coupling. In this work we focus on
two time series although the framework presented allows for the inclusion of as many variables as deemed
necessary.
Assume we can monitor a structure by recording it’s response at K separate locations for N discrete
points in time. The result is the multivariate time series x(n) ≡ xi (n), i = 1 . . . K n = 1 . . . N . Each
observation in time series i has associated with it an underlying probability density function p(xi (n)) which
returns the probability of observing the value xi (n). The joint density for data recorded from spatial locations
i and j is similarly given by p(xi (n), xj (n)). For the remainder of this work we assume stationarity such
that p(xi (n)) = p(xi (n + T ). Unless we are able to obtain multiple realizations of the structures response
this assumption will become necessary when estimating probabilities later. The assumption of stationarity
allows us to drop the absolute time index n when writing the densities i.e. p(xi ) ≡ p(xi (n)).
The central concept underlying this approach to nonlinearity detection is to form what amounts to a null
hypothesis and then map differences between observed and hypothesized behavior to a scalar. Mathemati-
cally speaking one can formulate the Kullback-Liebler (K-L) distance
Z Z  
f (xi , xj )
KL(f, g) = f (xi , xj ) log2 dxi dxj (1)
g(xi , xj )
xi xj

between the distributions g(·) and f (·) (base two logarithms are used here to place K-L distance in units of
bits). These distributions may involve single, joint, or conditional probability densities and are tailored to
the specific question being asked. The K-L distance has been utilized often in the field of Ecology.9

A. Transfer Entropy
Recently Schreiber13 (see also Kaiser & Schreiber14 ) proposed the transfer entropy metric for quantifying
dynamical relationships among time series data. This quantity is designed to account for transition probabil-
ities (dynamics) directly and hence does not require the use of a time delay to detect asymmetric coupling.
Assuming the dynamics of xi (n) can be described by the Markov model and the process is stationary we can
define
(k)
p(xi (1)|xi ) ≡ p(xi (n + 1)|xi (n), xi (n − 1), xi (n − 2), . . . , xi (n − k + 1)) (2)
where the probability of the system arriving at the value xi (n + 1) is conditional on the previous k values
only. Note only lags/advances need be retained in the notation. If we want to understand the nature of the
relationship between the two measurements xi (n), xj (m) we might ask what information the values xj (m)
carry about the transition probabilities of xi (n). In other words, explore the possibility that the dynamics
xi (n + 1) follow the model

p(xi (n + 1)|xi (n), xi (n − 1), . . . , xi (n − k + 1), xj (m), xj (m − 1), . . . , xj (m − l + 1)) (3)

(k) (l)
or in shorthand p(xi (1)|xi , xj ). Here the dynamics of xj (m) are modeled as an lth order Markov process.
The probability of the process xi being in a given state at time n + 1 is dependent on past history and
values of the process xj at discrete times m, m − 1, . . . , m − l + 1. The degree to which the values xj (n)
carry information about the transitions (dynamics) of xi (n) reflects dynamical dependence (as opposed to
the statistical dependence captured by an alternative IT metric, mutual information). This definition of
interdependence can also be quantified using a form of the K-L distance as
(k) (l)
p(xi (1)|xi , xj )
Z Z Z
(k) (l) (k) (l) (k) (l)
T E(xi (1)|xi , xj ) = p(xi (1), xi , xj ) log2 (k)
dxi (1)dxi dxj . (4)
p(xi (1)|xi )
Here the distributions f (·) and g(·) of equation (1) involve conditional probabilities. Equation (4) measures
the deviation from the assumption that the dynamics of xi (n) are influenced only by it’s past history and
not the dynamics xj (n). For simplicity we restrict ourselves to the case of k = l = 1 (i.e. first order Markov
models). It is useful to consider the time delay m = n + T . We are therefore asking what information the
values of the time series xj (n + T ) carry about the dynamics of xi (n). Using
Z Z Z
p(ai , bi ) log2 (p(ai )) = p(ai ) log2 (p(ai )) (5)
a b a

3 of 10

American Institute of Aeronautics and Astronautics

equation (4) can be rewritten in entropy form as
Z Z Z
T E(xi (1)|xi , xj (T )) = p(xi (1), xi , xj (T )) log2 p(xi (1), xi , xj (T ))dxi (1)dxi dxj (T )
Z Z Z
+ p(xi ) log2 p(xi )dxi − p(xi , xj (T )) log2 p(xi , xj (T ))dxi dxj (T )
Z Z
− p(xi (1), xi ) log2 p(xi (1), xi )dxi (1)dxi . (6)

If the process xj (n + T ) does not carry added information about the dynamics xi (n) the transfer entropy
will be zero. If the two processes are coupled such that there is some interdependence the transfer entropy
will rise to some non-zero value. It is interesting to note, however, that if the processes are perfectly coupled,
i.e. xi (n) = xj (n), for T = 0 the transfer entropy will also be zero. For the purposes of simply detecting the
presence of a nonlinear relationship between the time series this does not pose a problem.

B. Estimation Approach
Computing information theoretics involves the estimation of the various probability densities that comprise
equation 6. Unless we assume stationarity the densities can only be estimated using ensemble averaging
rather than from a single realization (measured time series). However, in many engineering applications,
including the ones described here, the data can be treated as stationary over the time scales on which
the measurements are taking place. Assuming stationarity, the most straightforward method of density
estimation is to sub-divide the data into bins and estimate the local densities based on the number of points
in those bins relative to the total number of points in the time series. For the ith bin we have p̂(i) ≈ Mi /N
where Mi arePthe number of points found in that bin. The Shannon entropies are then given by summing
over all bins p̂(i) log2 (p̂(i)).
A more accurate approach is to use kernel density estimates about each point. Two such estimators are
the fixed-bandwidth and fixed-mass kernels.6 In essence, the local density estimate about a given point with
time index n is given by the ratio of number of points in the local region to the volume of the local region.
A thorough discussion of density estimation in the context of information-theoretics is provided by Kaiser
and Schreiber.14 Because kernel-based methods involve summing over each point in the time series (rather
than bins), the entropy for the process x(n) may be written15, 16
Z
1 X
p(x(n)) log2 (p(x(n)))dx ≈ log2 (p̂(x(n))) (7)
N n

whereˆdenotes the estimated density. Substituting equation 7 into equation 6 then gives an estimate of the
transfer entropy.

C. Hypothesis Testing Using Surrogate Data

Testing for the presence of nonlinearity requires an alternative or null hypothesis. A logical choice of null
hypothesis is to assume that a linear model is sufficient to capture the relationship between the time series of
interest. Statistically significant deviations from this hypothesis are considered suggestive that the underlying
dynamics are indeed nonlinear. In this work we utilize surrogate data sets, an approach advocated by Theiler
et al.,17 Paluš,18, 19 and Schreiber & Schmitz.20 Surrogate time series are designed to preserve certain
properties of the original data (the properties being tested against) but are random with respect to others
(the properties being tested for). In this case we wish to construct surrogates that preserve the linear auto
and cross-correlations in the data but that are random with respect to all other correlations. Applying the
proposed algorithms to the original data and surrogates will therefore produce no difference if a linear model
is indeed adequate. However, the presence of nonlinearity will produce higher-order correlations not present
in the surrogates; this will be reflected in the transfer entropy. If the data are Gaussian, linear surrogates
may be constructed for the multivariate time series xk (n) k = 1 . . . K using the approach described by
Pritchard and Theiler.21 If Xi (f ), Xj (f ) denote the discrete Fourier transforms of xi (n), xj (n) i, j ∈ K, the
cross-spectral density may be written in terms of magnitude and phase as

Sxi xj (f ) = Xi (f )∗ Xj (f ) = |Xi (f )||Xj (f )|ei(φi (f )−φj (f )) . (8)

4 of 10

American Institute of Aeronautics and Astronautics

By the Weiner-Khitchine relationship, the linear cross-correlation associated with xi (n), xj (n) is simply the
inverse Fourier transform of Sxi xj (f ). Thus the linear auto and cross-correlations are preserved if a random
phase ψ(f ) ∈ [0, 2π) is added to both φi (f ), φj (f ). The general algorithm for creating surrogate data sets
for K time series is therefore
x̂k (n) = F −1 {Xk (f )eiψ } k = 1 . . . K. (9)
The resulting time series will exactly match only the linear correlations in the original data. Randomizing
the phases will destroy any higher-order correlations.
For non-Gaussian data, such as we use later in this work, a more general algorithm must be applied in
order preserve the amplitude distributions of the underlying time series. To this end we employ the iterative
algorithm described by Schreiber and Schmitz20 where by the original time series are continually re-shuffled
(phase randomized) in such a way as to preserve the linear correlations in the data as before. However,
because the surrogates are simply shuffled versions of the original time series, the amplitude distributions
are automatically preserved.
Applying the transfer entropy algorithm to the data and the surrogates will produce statistically sim-
ilar results if the system dynamics xi (n) are linear. Differences between the metrics computed from the
xi (n), x̂i (n) are assumed to be indicative of nonlinearity. Depending on the level of confidence desired in the
result the practitioner can utilize as few or as many surrogate data sets as necessary.

III. Experimental Setup

Figure 1. Experimental UAV wing (left) and a close-up of the damage site (right)

The wing used in this experiment (see figure 1) is based on the wing of an existing unmanned aerial vehicle
(UAV) with the exception that the stiffness has been reduced to provide a more tractable experimental
system. The reduced stiffness lowers resonant frequencies and provides greater displacement at lower force
levels than would be the case for the actual wing design. The wing has a span of 1320 mm, a chordwise
dimension of 152.4 mm and a 15.9 mm maximum thickness. It is a sandwich composite with the top and
bottom skins made of +/- 45 C282 Carbon Fabric/L285/H2875, 0 3k T300 Uni Tape/C285/H2875, 0 3k
T300 Uni Tape/C285/H2875, +/- 45 C282 Carbon Fabric/L285/H2875 bonded to a Last-a-Foam core.
Damage is induced in a localized region on one half of the wing near the center line (figure 1 inset) using
a home built rotary arm impactor capable of up to 100 J impacts. For the measurements in this experiment,
damage is kept to a barely visible damage level with impacts never exceeding 10 J. Following each impact,
the wing is mounted on a 50 N shaker and excited with a variety of input signals. The two excitation signals
considered for this work are a 0-1000 Hz Gaussian noise band and a signal based on the properties of a gust
load. A load cell placed in line between the shaker and the wing records the input. The dynamic strain
responses were recorded using a fiber optic strain measuring system which includes surface mounted fiber
Bragg gratings (FBGs) and custom demodulation hardware designed at NRL.22

5 of 10

American Institute of Aeronautics and Astronautics

 Three fiber Bragg grating strain sensors were positioned on the wing in the locations depicted in figure 1.
Strain time series data records of length 32768 were collected at a sampling rate of 2kHz. In an effort to
test the generality of the proposed approach we excited the wing with a non-Gaussian excitation generated
from a power spectral density model24 of a gust load.

(1 + 3(L/U )2 ω 2 )
S(ω) = (AπU ) . (10)
((1 + (L/U )2 ω 2 )2 )

The parameters U and L are the free-stream velocity and length scale of the gust respectively and A is a
gust velocity. Here we take U = 44.7m/s, L = 1.0m, and A = 10m/s. While this load is not actually a
gust load in the sense that it was applied at a point on the wing via a stinger and not as a uniform fluid
loading, the spectrum content is that of the gust load model. Damage was induced via a pendulum-type
impact machine and quantified in terms of the impact energy. A close-up of the damage site is shown in
figure 1 and the impact energies are tabulated in table 1 along with the locations of the impacts. A total of
7 damage cases were considered. The first three impacts occurred at location “L1” in figure 1. On the third

Table 1. Impact energies associated with the different damage levels

Damage Level 0 1 2 3 4 5 6 7
Impact Energy 0 7.3J 9.2J 8.2J 7.8J 8.9J 8.1J 8.8J
Damage Location - L1 L1 L1 L2 L2 L3 L3

impact we observed a slight crack in the skin of the wing. At this point we moved the location of the damage
to L2 for damage levels four and five. After the second impact at location L2 we again observed a slight
cracking in the skin and therefore moved to location L3 for the final two damage scenarios. For this study
40 surrogate data sets were generated. Because we used ambient (as opposed to Gaussian) excitation the
surrogates were generated using the iterative approach of Schreiber and Schmitz20 described in section (C).
Rather than examining relationships between the three strain sensors, in this experiment we examined the
relationship between the input (as recorded by the load cell) and the output at the various sensor locations.
To estimate the transfer entropy we used the adaptive kernel (fixed-mass) method.6

IV. Results
As a comparison Matlab’s n4sid subspace system identification method was used to identify modal fre-
quencies for the wing in the undamaged state and after each impact. We considered 60 segments of data each
1024 points long at each damage level using strain sensors 1, 2 and 5. The data were demeaned, detrended,
filtered at 90% of the cutoff frequency and normalized by one standard deviation. The top of figure 2 shows
an estimated transfer function of the wing with identified frequencies for models of order 1 to 20. The eighth
order model is the smallest order to successfully capture the first four modes of the structure. A 10th order
model was chosen for robustness. The mean identified frequency and the 95% confidence interval for modes
1 through 4 are plotted in bottom half figure 2. As mentioned earlier modal frequencies are not typically
sensitive to small amounts of damage. The frequency change for all of the modes was less than 1%. Mode 1,
3 and 4 show trends in the mean values. Only mode 3 has separable confidence bands from the undamaged
case to the most damaged level 7 case.
For the transfer entropy results, first the transfer entropy was calculated between the load cell and the
first strain sensor for each damage case. Next 40 surrogates were generated for each original transfer entropy
calculation at the various damage levels. The mean value of the surrogates at delays ranging from -40 to
+40 time steps are plotted as the dark solid line in figure 3. The dashed lines indicates the 95% confidence
range and the dots indicate the transfer entropy of the experimental data. The results for the undamaged
case (damage level 0) are shown on the left of figure 3. On the right of figure 3 are the results for damage
level 7 the most damaged case. The experimental data clearly moves out of the 95% confidence interval. If
we focus on the time delay range of 15 to 35 time steps which corresponds to the frequency range of 55-130

6 of 10

American Institute of Aeronautics and Astronautics

Figure 2. (Top) Identified frequencies for increasing model order overlaying an estimated transfer function of
the system. (Bottom) Mean identified frequencies with 95% confidence limits for modes 1-4 (A-D).

Hz we can calculate a normalized mean separation statistic as

P35
1 i=15 |T Edata,i − T E surrogate,i |
γ= (11)
20 σT Esurrogate

where T Edata,i is the transfer entropy calculated from the experimental data, T E surrogate,i is the mean

7 of 10

American Institute of Aeronautics and Astronautics

Figure 3. Time delayed transfer entropy results for undamaged level 0 (left) and damage level 7 (right).
Points indicate transfer entropy calculated from experiment. Thick solid line indicates mean transfer entropy
of surrogate data and dot dashed line shows 95% confidence level of surrogate data.

Figure 4. Normalized mean separation of time delayed transfer entropy experiment to surrogate from undam-
aged (level 0) to most damaged (level 7). Error bars indicate 95% confidence level.

transfer entropy calculated from the 40 surrogate data sets and σT Es urrogate is the standard deviation of
the transfer entropy calculated from the 40 surrogate data sets all values at delay i. The transfer entropy
results are shown in figure 4. A number of interesting observations can be made about this result. First, the
confidence bands of the third and all higher damage levels do not overlap with the undamaged confidence
bands indicating a clear separation for these damage cases. It is also interesting to note that the normalized
mean separation metric (γ) has a distinct upward jump when the composite skin cracked after the third
impact (damage level three) although the same effect is not seen at the breaking of the skin at the second
damage location (damage level 5). Perhaps the most striking observation is that the undamaged result is
not needed to make a determination of the presence of damage. Recall deviation from the surrogate transfer
entropy is an indication of nonlinearity and we hypothesize damage. Thus one need only set a confidence
threshold for γ such as one standard deviation ( 66%) or a more conservative value such as the damage level

8 of 10

American Institute of Aeronautics and Astronautics

three 1.6σ ( 90%) to establish a damaged criteria. A final observation is that the excitation was based on a
non-Gaussian ambient gust type load. The implication is that this damage metric could be assessed in an
in situ mode of operation.

V. Conclusion
In this work we have described an approach for detecting damage-induced nonlinearities in structures.
An information-theoretic (IT) metric, the time delayed transfer entropy was described. It offers a probabilis-
tic description of the coupling between two or more dynamical processes. An algorithm was presented for
estimating the IT metrics from time series data. We demonstrated how the method may be used in conjunc-
tion with appropriately designed surrogate data to detect the presence of a nonlinearity in an experimental
system. Results were presented for a composite UAV wing subject to impact damage and excited with a
non-Gaussian load with the same power spectral density of an ambient gust load. The transfer entropy based
metric was able to identify the damage as soon as the composite skin cracked after three 9 Joule impacts.
A clear strength of this approach is that it obviates the need for explicitly recording a baseline data set.
Rather the surrogates (i.e. the hypothesis equating a healthy structure with a linear structure) effectively
serve as a baseline thus allowing for an absolute measure of nonlinearity as opposed to a relative one.
Consequently this approach also eliminates variability associated with experimental set-up, sensor drift,
and/or global temperature changes from being seen as damage, assuming such changes do not introduce a
nonlinearity into the system.

References
1 Ratcliffe, C. P. and Bagaria, W. J., “Vibration technique for locating delamination in a composite beam,” AIAA Journal,

Vol. 36, No. 6, 1993, pp. 1074–1077.

2 Yoon, M. K., Heider, D., Gillespie, J. W., Ratcliffe, C. P., and Crane, R. M., “Local damage detection using the two-

dimensional gaped smoothing method,” Journal of Sound and Vibration, Vol. 279, 2005, pp. 119–139.
3 Sohn, H. and Farrar, C. R., “Damage Diagnosis Using Time Series Ansalysis of Vibration Signals,” Smart Materials and

Structures, Vol. 10, No. 3, 2001, pp. 446–451.

4 Wong, L.-A. and Chen, J.-C., “Damage Identification of Nonlinear Structural Systems,” AIAA Journal, Vol. 38, No. 8,

2000, pp. 1444–1452.

5 Johnson, T. J. and Adams, D. E., “Transmissibility as a Differential Indicator of Structural Damage,” Journal of Vibration

and Acoustics, Vol. 124, 2002, pp. 634–641.

6 Nichols, J. M., Seaver, M., and Trickey, S. T., “Detecting Damage-induced Nonlinearities in Structures Using Information

Theory,” Journal of Sound and Vibration (conditionally accepted).

7 Worden, K. and Tomlinson, G. R., “Nonlinearity in Experimental Modal Analysis,” Philosophical Transactions of the

Royal Society of London - Series A, Vol. 359, 2001, pp. 113–130.

8 Neild, S. A., Williams, M. S., and McFadden, P. D., “Nonlinear Vibration Characteristics of Damaged Concrete Beams,”

Journal of Structural Engineering, Vol. 129, No. 2, 2003, pp. 260–268.

9 Burnham, K. P. and Anderson, D., Model selection and multimodel inference: A practical information-theoretic approach,

Springer-Verlag, New York, NY, 2002.

10 Destexhe, A., “Oscillations, complex spatiotemporal behavior, and information transport in networks of excitatory and

inhibitory neurons,” Physical Review E , Vol. 50, No. 2, 1994, pp. 1594–1606.
11 Vastano, J. A. and Swinney, H. L., “Information Transport in Spatiotemporal Systems,” Physical Review Letters, Vol. 60,

No. 18, 1988, pp. 1773–1776.

12 Ho, M. C. and Shin, F. C., “Information flow and nontrivial collective behavior in chaotic-coupled-map lattices,” Physical

Review E , Vol. 67, No. 5, 2003, pp. 056214.

13 Schreiber, T., “Measuring Information Transfer,” Physical Review Letters, Vol. 85, 2000, pp. 461.
14 Kaiser, A. and Schreiber, T., “Information Transfer in Continuous Processes,” Physica D, Vol. 166, 2002, pp. 43–62.
15 Liebert, W. and Schuster, H. G., “Proper choice of the time delay for the analysis of chaotic time series,” Physics Letters

A, Vol. 142, 1989, pp. 107–111.

16 Prichard, D. and Theiler, J., “Generalized Redundancies for Time Series Analysis,” Physica D, Vol. 84, 1995, pp. 476–493.
17 Theiler, J., Eubank, S., Longtin, A., Galdrikian, B., and Farmer, J. D., “Testing for nonlinearity in time seires: the

method of surrogate data,” Physica D, Vol. 58, 1992, pp. 77–94.

18 Paluš, M., “Testing for nonlinearity using redundancies: quantitative and qualitative aspects,” Physica D, Vol. 80, 1995,

pp. 186–205.
19 Paluš, M., “Detecting Nonlinearity in Multivariate Time Series,” Physics Letters A, Vol. 213, 1996, pp. 138–147.
20 Schreiber, T. and Schmitz, A., “Improved Surrogate Data for Nonlinearity Tests,” Physical Review Letters, Vol. 77,

No. 4, 1996, pp. 635–638.

21 Prichard, D. and Theiler, J., “Generating Surrogate Data for Time Series with Several Simultaneously Measured Vari-

ables,” Physical Review Letters, Vol. 73, No. 7, 1994, pp. 951–954.

9 of 10

American Institute of Aeronautics and Astronautics

 22 Todd, M. D., Johnson, G. A., and Althouse, B. L., “A Novel Bragg Grating Sensor Interrogation System Utilizing a

Scanning Filter, a Mach-Zehnder Interferometer, and a 3x3 Coupler,” Measurement Science and Technology, Vol. 12, 2001,
pp. 771–777.
23 Quirin, W. L., Probability and Statistics, Harper & Row, 1978.
24 Dowell, E. H., Crawley, E. F., Curtiss, H. C., Peters, D. A., Scanlan, R. H., and Sisto, F., A Modern Course in

Aeroelasticity, Kluwer Academic Publishers, 1995.

10 of 10

American Institute of Aeronautics and Astronautics