You are on page 1of 14

Original Article

Damage detection in composite


materials structures under variable
loads conditions by using fiber Bragg
gratings and principal component
analysis, involving new unfolding and
scaling methods

Journal of Intelligent Material Systems


and Structures
114
The Author(s) 2014
Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/1045389X14541493
jim.sagepub.com

Julian Sierra-Perez1, Alfredo Guemes2, Luis E Mujica3 and Magda Ruiz3

Abstract
An innovative methodology based on the use of fiber Bragg gratings as strain sensors and strain field pattern recognition
is proposed for damage detection in composite materials structures. The strain field pattern recognition technique is
based on principal component analysis. Damage indices (T2 and Q) and detection thresholds are presented. New techniques for unfolding and scaling tridimensional matrices arrays obtained from structures working under variable load
conditions are presented.
Keywords
Fiber Bragg gratings (FBGs), principal component analysis (PCA), structural health monitoring (SHM), damage detection.

Introduction
Nowadays, the aerospace and civil industries, among
others, are focusing on improving the reliability of their
structures through the development of new systems for
the monitoring and detection of damages. In recent
years, efforts have been concentrated on designing
smart structures that integrate materials, sensors, actuators and algorithms capable of diagnosing the structural
condition in real time. The detection of any anomaly in
its early stages is then possible. This philosophy has
been called structural health monitoring (SHM).
The conventional inspection techniques and nondestructive evaluations are used only for structures during maintenance procedures and are only able to detect
single local damage. On the contrary, smart systems are
designed to detect global damage in complex structures
and can operate autonomously during service. The main
difference between traditional Nondestructive Testing
(NDT) techniques and SHM lies in the integration of
the sensors and algorithms with the structure, so that,
during the whole operation life, a sensors network will
measure the structural response and the algorithms will
predict the remaining life or will give information about
the structural integrity.

The main objective of SHM is to identify the principal characteristics that are related to the physical state
of the system during its operation, without affecting its
integrity. Among these characteristics the following can
be found: the operational and environmental loads, the
load-induced damages, the damage growing as the system operates, and the performance of the system as
function of the accumulated damages (Adams, 2007).
When these SHM systems are fully developed, they will
reduce the time and costs associated with maintenance
by eliminating unnecessary inspections and replacements (Balageas et al., 2006).
There are several classifications of the different levels of SHM in the literature, however, perhaps the most
1

Aerospace Engineering Research Group, Universidad Pontificia


Bolivariana, Medelln, Colombia
2
Department of Aerospace Materials and Processes, Universidad
Politecnica de Madrid, Madrid, Spain
3
Department of Applied Mathematics III, EUETIB Universitat Politecnica
de Catalunya BarcelonaTech, Barcelona, Spain
Corresponding author:
Julian Sierra-Perez, Aerospace Engineering Research Group, Universidad
Pontificia Bolivariana, Circular 1 70-01, Medelln, Colombia.
Email: julian.sierra@upb.edu.co

Downloaded from jim.sagepub.com by guest on January 20, 2015

Journal of Intelligent Material Systems and Structures

complete classification includes six basic levels of SHM


(complementary) that can be interpreted as the objectives to be achieved by the use of SHM techniques
(Suleman, 2011):







Level 1: damage detection;


Level 2: damage location;
Level 3: identification of type of damage;
Level 4: quantification of damage severity;
Level 5: estimation of remaining life;
Level 6: development of self-diagnosis and selfhealing capabilities.

To date, according to Suleman (2011), there are partial solutions, at laboratory level, for the first three levels and solutions for the fourth level are under
development. One of the main reasons why the implementation of different techniques in real SHM applications has progressed so slowly is that there are
uncertainties in each level. Perhaps the only area where
SHM techniques are relatively mature and are being
applied industrially is in the field of rotary machinery.
This has happened because there is a lot of information
about the typical location and type of damage present
in this kind of machinery (Lopez and Sarigul-Klijn,
2010).
SHM systems currently use modern experimental
techniques that allow very precise measurement of different types of structural response to external loads.
The result of the measures can subsequently be analyzed by numerical and statistical techniques to identify
the onset of damage. Generally, current SHM systems
consist of sensors, data acquisition and preprocessing
blocks of data (software), data communication systems
and post-processing blocks of data (Holnicki-Szulc,
2008).
A very promising group of measurement techniques
is based on fiber optic sensors (FOS). FOS are very sensitive to strain and temperature (among other physical
variables for some specific kinds of sensors). Thanks to
their small size, low weight, non-electrical nature, multiplexing availability, high accuracy, etc., FOS are suitable for SHM applications. Among the different types
of FOS the wavelength-based sensors, also called fiber
Bragg gratings (FBGs), have recently attracted most of
the attention from researchers due to their high sensitivity to strain and temperature. The FBGs can measure
strains with similar or more accuracy than standard
electric strain gauges and have the main advantages
that they are more reliable for long-term measurements
because they do not drift by aging, and can be multiplexed since several FBGs can be engraved in the same
fiber (Ghatak and Thyagarajan, 1998).
The first works on damage detection used identification techniques based on physical models, for example,
determination of the stiffness matrix or modal parameters. In these approaches all parameters are

considered as measurable and different uncertainties


are not represented for the model directly. This causes
difficulties in assessing how reliable the damage estimates are (Doebling et al., 1998; Fritzen and Bohle,
2003; Fritzen et al., 1998; Li et al., 2012).
A more recent technique estimates the occurrence of
damage based on experimental data. These models take
experimental measurements for training or learning
and assessment of the current state of a component or
structure. These methods are a very robust way of indicating the presence of damage (Fritzen et al., 2009).
Recently, some researchers are trying to use both
methods (physical models and models based on experimental data) in mixed models, in order to maximize the
reliability and the obtained information from damage
diagnosis. Lopez and Sarigul-Klijn (2010) claim that
the increase in reliability is directly proportional to the
level of the considered uncertainty. In systems or structures diagnostic applications, uncertainty can take the
form of the variability of a system or component, environmental or operational conditions, data acquisition
errors, errors in data processing, etc.
With the advent of composite materials it became
possible to design and manufacture increasingly complex structures. As the complexity of the structures
increases, so do the physical/mathematical models that
can accurately describe relationships between external
loads and responses in the structure.
Pattern recognition techniques use experimental data
in order to avoid the explicit use of complex physical/
mathematical models while being efficient and robust
against uncertainties.
Another common SHM limitation lies in the amount
of information usually handled when dozens of sensors
measure different variables for long periods of time.
Often, much of this data must be rejected in order to
reduce noise or irrelevant data and additional data processing must be done in order to reduce the redundant
information and focus the study only on the correlatable information.
There are many preprocessing techniques that can
treat data acquisition errors, noise, transformation and
compression. Many signal-processing techniques
include Fourier analysis for stationary time-invariant
problems and time-varying methods such as wave
methods (wavelets), time-frequency methods and time
series analysis. The main methods for time-frequency
analysis are the short-time Fourier transform (STFT),
WignerVille distribution (WVD) and ChoiWilliams
distribution (CWD), among others. The main difference between the different time-frequency methods is
the treatment of uncertainty (Baseville et al., 2007;
Fritzen and Kraemer, 2009; Lopez and Sarigul-Klijn,
2010).
After data preprocessing, many techniques are used
for the discrimination of information, the general purpose of which is to find information and highlight

Downloaded from jim.sagepub.com by guest on January 20, 2015

Sierra-Perez et al.

hidden patterns in data. With this it is possible to manage the information for optimization purposes, decision
support and control processes, among others. The
strength of the different techniques used is that they are
able to transform highly correlated, redundant or noisy
data on a statistical or data-driven model whose elements provide an overview of occult phenomena and
correlations that determine the behavior of the system
(Lopez and Sarigul-Klijn, 2010; Mujica et al., 2008;
Westerhuis et al., 1999).
Some of the techniques currently available include, for
example, Markov models, linear regressions, reducedorder modeling (ROM), singular value decomposition
(SVD), independent component analysis (ICA), proper
orthogonal decomposition (POD), principal component
analysis (PCA), multiway PCA (MPCA) (also called
Tucker 1), parallel factor analysis (PARAFAC), Tucker
3, partial least squares (PLS), multiway partial least
squares (MPLS), etc. (Westerhuis et al., 1999).
The main purpose of this work is to develop and
experimentally validate a pattern recognition technique
for damage detection in structures under variable loads,
based on the local strain fields and global stiffness of
the structure, by means of strain measurements
acquired using FBGs attached to the structure. An
innovative adaptation of a very well-known multivariate technique, often used as a dimensional reduction
tool, called PCA, is proposed (Jollife, 2002), along with
the development of damage indices and damage detection thresholds. An innovative way to incorporate PCA
in operational variable conditions, in particular under
variable loads, is also presented.
In a previous work Sierra et al. (2013) developed a
pattern recognition technique for damage detection in
structures under stationary load conditions, based on
the local strain fields and global stiffness of the structure. However, a limitation was observed when several
load conditions were used. The results were automatically grouped according to the load cases and the reading of such results should be performed manually.
In this work, the authors present an extension of the
methodology previously developed. The extension includes
new unfolding and standardizing methodologies for the
setting where the structure is working under variable load
conditions. The main goal consisted in unify all the results
regardless of the load case these belongs. This is achieved
by obtaining damage indices with the same scale regardless
of the load condition. In this way it is possible to automate
the damage assessment process since no human interpretation is needed for each different load case.

FOS
Introduction
Optical fibers are dielectric waveguides for light propagation. Usually they are made from high-purity silica or

other transparent materials (like some polymers). The


optical fiber has a core with a refractive index slightly
higher than the surrounding material, called cladding,
due to the presence of dopants. Then, light is confined to
the core, since when the light arrives at the core/cladding
interface with an angle higher than the total reflectance
angle (as defined by the Snell law), it follows the total
reflection and remains confined into the core (Guemes
and Sierra, 2013; Hill and Meltz, 1997).

Fiber optic
Optical fibers can be classified into two main groups:
monomode and multimode optical fibers. The main difference between the two groups lies in the diameter of
the core. Since electromagnetic waves traveling trough
the core must satisfy Maxwells equations, the cylindrical contour conditions only allow for a discrete number
of solutions V, which it is dependent on the core diameter and the wavelength.
r
n 2n 2
V 2p2 a2 1 2 2 ;
l

where n1 represents the refractive index of the core, n2


the refractive index of the cladding, l the wavelength
and a the diameter of the core. When V is less than 2.4,
only one mode (two orthogonal polarizations) propagates through the core. This is the case for the monomode optical fiber (Guemes and Menendez, 2006).
From a sensing point of view monomode fiber optics
are preferred since they have smaller optical attenuation and do not suffer group dispersion due to different
modes traveling at different speeds. The preferred
wavelength window for sensing is around 1550 nm
since telecommunications industries have been developing optoelectronic components in the 1300 nm to 1550
nm window, so those components are more easily available. The main reason for that is that power loss is
minimized in such a window, where minimum Rayleigh
dispersion and minimum infrared absorption are found
(Culshaw and Kersey, 2008).
Fiber optics sensors (FOS) are able to measure
strain, temperature or other physical parameters (chemical, mechanical, electrical, etc.) either for distributed
sensing or discrete sensing (distributed sensing implies
the use of the whole fiber extension as a sensor whereas
discrete sensing implies the use of a small portion of the
fiber as a sensor). Unfortunately, distributed sensing
has some accuracy and spatial resolution limitations.
This has limited its use in some applications. For that
reason, local sensors (discrete) are more usual, either by
including an external device into the optical path
(extrinsic sensor), or by modifying the fiber for a short
length (intrinsic sensor) in order to give to the fiber sensibility to a physical parameter. For smart structures,
local intrinsic sensors are the most appreciated, because

Downloaded from jim.sagepub.com by guest on January 20, 2015

Journal of Intelligent Material Systems and Structures

of their simplicity, the minimal amount of perturbation


to the host structure and high sensitivity and accuracy
(Guemes and Sierra, 2013).
FOS have a wide variety of advantages such as small
size and weight, non-electrical nature (immunity to electromagnetic interference and to electrical noise), high
sensitivity, wide operating-temperature range, multiplexing ability, etc.
Besides the previous topological classification for
FOS (local or distributed and intrinsic or extrinsic), a
general classification can be done according to the optical parameter that defines the light as an electromagnetic wave: frequency, intensity, phase and
polarization. That is, phase-modulated optical fiber
sensors (or interferometers), intensity-based sensors
and wavelength-based sensors (or Bragg gratings).

FBGs
The wavelength-based sensors, developed at the beginning of the 90s, have recently attracted most of the
attention from researchers due to their high sensitivity
to strain and temperature. The basic idea consists in
engraving, at the core of a special photosensitive fiber
optic, a periodic modulation of its refractive index.
This can be accomplished thanks to dopants like germanium, present at the core, which, under certain conditions (typically exposing the core to an ultraviolet
light source), promote the changing of the refractive
index. The periodic modulation behaves as a set of
weak partial reflecting mirrors, which by accumulative
phenomenon (called diffraction) reflects back the optical wavelength that is exactly proportional to the spacing of the periodic modulation engraved.
Braggs diffraction law simplifies under normal incidence to the following expression
lb = 2ne L0

where lb is the Bragg wavelength, L0 is the pitch or


period of the modulation and ne is the average refractive index.
The FBGs behave then as a narrow optical fiber, promoting the reflection back of a very narrow-wavelength
band as shown in Figure 1. When the grating is subjected to an axial strain or a thermal gradient, the central wavelength of the spectrum reflected by the grating
is shifted due to changes in the pitch and the refractive
index. By means of an optical spectrum analyzer, this
shifting can be converted into readable information.
FBGs can be used, then, as strain and temperature
sensors, among others. To employ the FBGs as strain
sensors, these must be adhered to or embedded in the
material in which the strain measurement is desired. In
this way, the FBG is subjected to the same strain as the
structure to which it is attached. On the other hand, in
order to use the FBGs as temperature sensors, these

must be isolated from the mechanical effects on the


structure to which are attached, so that the modulation
period is only affected by thermal effects (changes in
refractive index plus thermal expansion or contraction).
Since in real operational scenarios an FBG integrated
with a structure works under mechanical and thermal
loads, it is necessary to make a thermal compensation
in order to discount the thermal effects from strain
measurements (Peters, 2009).
Other parameters that can be monitored with FOS
include the detection of impact damage and delamination, structural strains, crack propagation in bonded
repair systems, and temperature- and fatigue-induced
damage.
The impact detection in composite materials is performed by taking advantage of the sensitivity of the
FBGs to the transversal strains. The changes in the distortion of the strain field promoted by transversal
strains when an impact has occurred in a composite
material can be detected by means of the FBGs
(Menendez and Guemes, 2005).
Another promising application of FBGs is the measurement of Lamb waves. FBGs offer some advantages
over piezoelectricity for this purpose, for instance the
smaller size of the sensors and the directionality which
offers the possibility of locate the sensors in such way
that only waves in certain directions are detected.
(Takeda et al., 2005).

PCA
Introduction
When strain field pattern recognition techniques are
used, the main idea is to correlate all the strain measurements gathered from a network of sensors in a
complex structure and to discern if something has
changed, in particular, the global stiffness and the
strain field between different sensors, as a product of
damage appearance in the structure. If, for example,
the response of two sensors is simultaneously analyzed,
it is possible to detect small changes in the local stiffness or the strain field as result of the occurrence of a
defect in the structure. This technique has been called
differential strains (Fernandez et al., 2007).
The following example, which was built with the
information gathered during the experimental phase
(which will be explained later) shows such a concept. A
plot strain versus strain for two neighboring sensors
under several load cases was built for the healthy structure. After that, data induced damage cases were
plotted in same figure. As a result of damage, a change
in the slope is advisable under the same load conditions.
In Figure 2 can be seen such a difference in the slopes.
In SHM applications, all the measurements must be
studied together in order to increase the probability of
damage detection. Then, it is necessary to use any

Downloaded from jim.sagepub.com by guest on January 20, 2015

Sierra-Perez et al.

Figure 1. FBGs operating principle scheme.

Data organizing and unfolding

Figure 2. Differential strains for two neighboring sensors.

multivariate statistical tool in order to get valuable


information about the system behavior.
PCA is a statistical technique whose main purpose is
to identify the most important dynamics of a system,
determining which data are redundant and which are
just noise. By applying PCA to experimental samples it
is possible to reduce the complexity and size of the samples, revealing patterns and trends that may be hidden
under the data. This is achieved by determining a new
coordinate space based on the covariance of the original
dataset (Jollife, 2002; Sierra and Guemes, 2011; Sohn
et al., 2002).

Normally, the measurements implied in SHM techniques are continuous dynamic signals as functions of
time. Therefore, before applying PCA to experimental
data, it is necessary to discretize the signals and, in some
cases, rearrange the data in a proper way.
For example, if one SHM system gathers measurements of several variables (m) by means of different
kinds of sensors, during a specific time, and for a discrete number of experimental trials (n), the information
can be arranged in an X matrix, where each row vector
(xi) represents the measurements from all sensors in a
specific experimental trial and each column vector represents the measurements from one sensor (variable) in
the whole set of experiment trials.
However, if each variable is a set of dynamic measurements, all the gathered information could take the shape
of a tridimensional matrix composed of J sensors or variables, K samples or time intervals and I experiments or
trials, as can be seen in the bottom of Figure 3.
In order to apply PCA or any other multivariate
analysis technique, the tridimensional matrix (X3D) must
be rearranged in a proper way. This process is often
called unfolding (Kourti and MacGregor, 1995).
In the literature, several ways of unfolding tridimensional data arrays are reported, each one allowing the
study of different kinds of variability. Westerhuis et al.
(1999) studied the six possible ways of unfolding a tridimensional matrix, resulting in six different bidimensional
matrices (see Figure 3): matrix A (KI 3 J), matrix B

Downloaded from jim.sagepub.com by guest on January 20, 2015

Journal of Intelligent Material Systems and Structures

S1

EI

EI

S1
S2
SJ

Ei

Ei

Sj

E2

E2

E1

E1

X3D

S2

Sj

SJ

T1

T2

Tk

Tk

T1

T2

Tk

Tk

Figure 3. Different ways of unfolding tridimensional data arrays.

(JI 3 K), matrix C (IJ 3 K), matrix D (I 3 KJ),


matrix E (I 3 JK) and finally matrix F (J 3 IK).
If a PCA study is performed, matrices B and C are
equivalent (rows in a different order). In the same way,
matrices D and E are also equivalent (columns in a

different order). The matrix F corresponds to the transposed matrix of A. If matrix F is used to perform a
PCA study, only the T-scores will be changed (in comparison with the T-scores obtained by using the matrix
A) if no technique of centering and scaling (often called

Downloaded from jim.sagepub.com by guest on January 20, 2015

Sierra-Perez et al.

standardization) is used before performing the PCA


study.
Several authors have reported the strengths and
weaknesses associated with using each of the forms
shown. In particular, Nomikos and MacGregor (1994),
who introduced the method of monitoring batches,
reported that unfolding the initial data array (X3D) into
the form of matrix D (I 3 KJ) is the most significant
for analysis and monitoring of batch data. Thanks to
unfolding data in this manner, the entire batch of data
(all the trajectories of process variables, obtained in
each test group) is considered as a single object.
Through this type of unfolding it is possible to study
the differences between batches of data. This allows us
to compare each batch of data with a group of batches
that are known to be good (for example, in the case of
interest, a structure operating in an undamaged condition) and in this way classify whether a batch is good
or not.
Other authors have proposed the use of the unfolded
type A. With this type of unfolding, each point of the
trajectory of each batch of data is considered an object.
When centering is performed on the unfolded matrix
A, a constant is subtracted from the trajectory of each
variable in each batch of data (for example, the overall
average of each variable for each batch of data and for
all instants of time). This preserves the nonlinear trajectories in the time-dependent data array. Such unfolding
and centering in the average is useful when a wide variety of trajectories for some of the variables has been
tested. However, this technique is not very useful in the
case in which the data batches have been obtained consistently, that is, if only one or few trajectories have
been tested (Wold et al., 1998).

Scaling
Regardless of which type of unfolding has been used,
there is another additional consideration which must
be taken into account: the scale effect. Since physical
variables do not have the same magnitudes and scales,
it is necessary to process the experimental data before
any analysis of statistical nature. In the literature there
are a variety of techniques to rescale the experimental
data, expressed in unfolded matrices. Among the most
used techniques are continuous scaling (CS), variable
scaling (VS), group scaling (GS), autoscaling (AS), etc.
(Gurden et al., 2001; Kourti and MacGregor, 1995;
Mujica et al., 2010, 2011; Nomikos and MacGregor,
1994; Sierra et al., 2013; Villez et al., 2009; Westerhuis
et al., 1999; Wold et al., 1998).
When prior knowledge is available, an operator or
expert may define the scaling approach. For instance,
Westerhuis et al. (1999) argued that AS is appropriate
in systems where the variables have different units (e.g.
temperature, pressure and concentration). A variation
in temperature could be much greater than a variation

in concentration; however, the effect on the system may


be similar. Mean and standard deviation by each time
instant by each sensor are denoted by
I
1X
xijk
I i=1

I
1 X
(xijk  mjk )2
IK i = 1

mjk =

s2jk =

Therefore, the new scaled sample is given by


xijk =

xijk  mjk
q
s2jk

In CS the data are treated as if all samples are drawn


from the same distribution. Thus, for each sensor, one
mean and one standard deviation are calculated on the
basis of all data (all experiments and all time instants).
By doing so, J means, mj, J standard deviations, sj and
the new scaled sample are obtained by means of the following equations
I X
K
1 X
xijk
IK i = 1 k = 1

I X
K
1 X
(xijk  mj )2
IK i = 1 k = 1

xijk  mj
q
s2j

mj =

s2j =

xijk =

All the measurements by sensor (throughout the


whole set of experiments and time instants) are then
scaled with the same mean and standard deviation. In
general, however, it is not expected that samples gathered by a sensor in different experiments are drawn
from the same distribution and this possibly leads to
bad performance of the PCA model.
GS partly addresses this problem by removing the
trajectory from the time instants throughout the sensor.
This is done by defining a separate mean for each sensor at each time instant in the experiment thereby
obtaining a mean trajectory for each sensor. After scaling with the mean trajectories, one standard deviation
is calculated per sensor, similar to CS. Now, one
obtains J 3 K means, mjk, and J standard deviations,
sj, as follows

Downloaded from jim.sagepub.com by guest on January 20, 2015

I
1X
xijk
I i=1

I X
K
1 X
xijk
IK i = 1 k = 1

mjk =

mj =

Journal of Intelligent Material Systems and Structures

Pr

Figure 4. Scheme of PCA methodology for SHM.

s2j =

I X
K
1 X
(xijk  mj )2
IK i = 1 k = 1

xij  mjk
xij = q
s2j

10
11

Even if the trajectory of the data is removed from


the dataset by GS, the standard deviation of the sensors is assumed not to change along the experimental
trials.
It should be noted that these approaches can be only
applied in cases where the data trajectories of variables
have the same length.

Pattern recognition based on PCA


Once the X3D matrix has been unfolded, centered and
scaled (if necessary), an X matrix is obtained. After this,
a PCA study can be performed based on the covariance
matrix which quantifies the degree of linearity between
all possible pairs of variables. Ordering the highest eigenvalues (in descending order) associated with the covariance matrix, and retaining only a few first components
(eigenvectors associated to the highest eigenvalues), it is
possible to obtain an important dimensional reduction
(Manson et al., 2001; Sierra et al., 2013). For a further
discussion about PCA fundamentals see Jollife (2002).
The principal purpose is to transform the X matrix
by means of a linear transformation in order to achieve
minimal redundancy between the new data as follows
T = X Pr

12

where Pr is a matrix containing the eigenvectors


associated to the first r highest eigenvalues (sorted in
descending order).
The T matrix (called score matrix) has uncorrelated row vectors and its column vectors are the projection of the original data over the direction of the
principal components. These column vectors are called
T-scores. In this work, PCA is used to construct a

baseline statistical model using data gathered from a


healthy structure. This means calculating matrix Pr for
the healthy structure. Later, results for an unknown
structure condition (current structure) (X ) are projected
into the baseline model (Pr) and the score matrix is
analyzed. The methodology is outlined in Figure 4
(Mujica et al., 2010, 2011).
From the obtained projections it is possible to calculate different kinds of damage indices and establish
detection thresholds in order to classify the structural
condition, infer the presence of a defect in the structure
(Huedo et al., 2006; Mujica et al., 2010, 2011).
If the structure remains healthy during operation (so
that the measured variables are influenced only by common causes variations), the values of the r first principal components and therefore their associated indices
should be inside an interval close to the index obtained
during baseline building. Conversely, if the principal
components remain stable over time or experimental
trials, the common effects that influence the process are
likely to remain constant.
There are statistical tools that, together with PCA
analysis, allow us to detect anomalous behavior in the
system. The two most commonly used indices (also called
statistics) are the Q index and the T2 index. The Q index
is a measure of how well each sample fits the PCA model.
Basically, this index indicates the difference between a
particular sample and its projection into the main principal components retained by the PCA model. In other
words, it is a measure of the error given by the data which
was not projected into the new subspace (principal components). The Q index for the ith experiment is given by
Qi = ~xi~xTi = xi (I  Pr PTr )xTi

13

where xi is the row vector of the original matrix (X ),


and ~xi is the projection into the residual subspace.
On the other hand, the T2 index is a measurement of
the variation of each sample inside the PCA model and
is based on the analysis of the score matrix (T). This
matrix allows us to study the variability of the projected

Downloaded from jim.sagepub.com by guest on January 20, 2015

Sierra-Perez et al.

data in the new principal component space (Jackson


and Mudholkar, 1979; Sierra et al., 2013).
The T2 index for the ith experiment is given by
Ti2

r t2
X
sij
j=1

lj

tsi tsiT
L

xi Pr PTr xTi
L

17

xijkl  mjk
q
s2j

18

14

New unfolding and scaling methods


Once the data have been grouped according to the load
case, a new unfolding and centering methodology is
proposed. The methodology consists of unfolding the
data so that groups related to different load cases and
data from each sensor are grouped in columns. A new
fourth dimension (L) must be added (representing
load case). Therefore, the original data matrix X 2 R4 ,
where each data point xijkl represents the ith experiment, the jth sensor, the kth instant time and the lth
load. A general Scheme of the proposed methodology
can be seen in Figure 5.
According to the unfolding method used, a new methodology for centering and scaling is proposed. In this
case, the scaling was performed using the mean of each
sensor at each instant of time for the total number of
experiments, and the standard deviation by sensor for all
measurements, and all load cases simultaneously, thus

mj =

I X
K X
L
1 X
(xijkl  mj )2
IKL i = 1 k = 1 l = 1

xijkl =

where tsi is the row vector (row i within matrix T),


which is the projection of the experiment xi into the
new space. Further, tsi and xi are related by tsi = xiPr
as indicated in equation (12), lj is the jth eigenvalue of
the covariance matrix of (X ) and L = diag(l1, l1, .,
lm) (Mujica et al., 2010, 2011).
In real operation, most structures operate under
varying load conditions over time. For this reason, it is
necessary to develop a methodology that allows us
firstly to classify data groups according to the specific
load case in question, and secondly to group and unfold
all the previously classified data in a single data matrix
that can be projected onto the PCA model.
There are several methods of data classification by
using different automatic and semi-automatic techniques (mostly based on neural networks), however, in
order to reduce the variability and uncertainty in the
experiment presented in this article, no automatic or
semi-automatic classification techniques were used.
When data was acquired, the applied load was known
and it was possible to classify a priori data groups. For
a further discussion about classification techniques see
Lopez and Sarigul-Klijn (2010).

I X
L
1X
xijkl
IL i = 1 l = 1

15

I X
K X
L
1 X
xijkl
IKL i = 1 k = 1 l = 1

16

mjk =

s2j =

Damage produces a variation in the different strain


field responses (patterns). However, the variation
sensed by one particular sensor could be much greater
than the variation sensed by other sensor. In the same
way, the variation of the patterns under one load condition could be much greater than the variation for
other load conditions.
By means of this new unfolding methodology, it is
possible to treat a group of data associated to different
load cases as a new single variable instead several isolated variables. In this way, all the information gathered
for each sensor is treated together with all the other sensors, and all the possible correlations between sensors
under all the load cases are analyzed simultaneously.

Experimental setup
A wing section of an unmanned air vehicle (UAV) 1.5
m long and fully made of composite materials was used.
The used section is schematized in Figure 6.
Four fiber optics, each one containing eight FBGs,
were bonded to the structure at the surface (wing skin),
two at the intrados (one fiber optic close to the main
spar and the other close to the secondary web) and two
at the extrados (in the same locations as the fiber optics
bonded at the intrados). In total, 32 FBGs were used.
The wing was fixed to a testing bench emulating the
way in which it is fixed to the aircraft fuselage (see
Figure 7). The wing structure is manufactured of carbon fiber, glass fiber, PVC foam and balsa. Figure 8
shows a schematic representation of the internal structure of the tested wing section.
Once the wing was fixed, the testing phase was performed. The first step consisted of gathering the strain
(response) at sensor locations for the healthy state
under all the load cases (baseline), in order to build the
PCA model. Each experiment consisted in loading the
structure in bending mode, increasing the load progressively. Four different loads were used in the experiments: 3.25 kg, 4.75 kg, 6.25 kg and 7.25 kg. Each load
case (for the healthy structure) was repeated 20 times,
of which 50% were used for building the baseline and
the rest were used to validate the methodology. When
these experimental data are projected into the PCA
model, representative changes should not be seen in the
different studied statistical indices when these are compared with the baseline model ones.
After building the baseline statistical model for the
healthy structure, two kinds of accumulative artificial
damage were induced in the structure. Figure 9 shows

Downloaded from jim.sagepub.com by guest on January 20, 2015

10

Journal of Intelligent Material Systems and Structures

X3D

X3D
Load case1

S1

S2

X3D

Sj

SJ

Figure 5. Proposed unfolding methodology.

Figure 7. Experimental setup.

Glass fiber
Carbon fiber
carbon UD

carbon fiber

Figure 6. Tested wing section.

Figure 8. Scheme of wing structure.

the sensors and induced damage locations. The first


kind of induced damage consisted in a longitudinal cutting close to the spar (see Damage 1 in Figure 9).

The second kind of damage consisted in a transversal


cutting close to the first kind of damage (see Damage
2 in Figure 9). Five damage cases were evaluated. The

Downloaded from jim.sagepub.com by guest on January 20, 2015

Sierra-Perez et al.

11

Figure 9. Sensors and induced damage locations.

first case consisted in the initial longitudinal damage


with initial length of 1 cm. For the second case, the first
damage was propagated 2 cm. The third damage case
consisted in the initial transversal cutting (plus the previous longitudinal cuttings) with an initial length of 1
cm. From the fourth to the fifth case, the transverse
cutting was increased by 1 cm each time. In the fifth
case, the spar was also superficially cut. For each damage case, the four load magnitudes (load cases) were
used and each experiment was performed 10 times. For
each experiment, strain was taken 400 times under stationary conditions. Later, a mean was calculated for
each measurement. Since all experiments were performed in a controlled-temperature room, and each
one took no more than 45 s, the temperature changes
can be disregarded and therefore thermal effects also
can be disregarded.
A Micron Optics Si 425 swept laser interrogator was
used for data acquisition. This equipment gathers the
wavelength reflected by each one of the FBGs. The
strains for each sensor can be calculated using the following expression
e=

DlB
= (1  ra )De + (1 + j)DT = ke De + kT DT
lB
19

where lB is the Bragg wavelength, ra is the photoelastic coefficient of the fiber optics and j is the thermooptic coefficient for the fiber optics. For the used fiber
optics ke = (0.7991 6 0.0055)me21 and kT = (6.334 6
0.074) 3 1026K21 (Garc a, 2010).
Finally, after performing all the experiments for the
healthy structure, PCA is applied to get the baseline
model, and later, the experimental data for the different
damage cases are projected into the PCA model. Since
for this setup the first five principal components
explained more than 90% of the variability, only these
principal components were retained in the PCA model.

Analysis of results
After projecting the experimental data for different
damage cases and for all different load cases into the
PCA model, the projections of the two principal components are depicted in Figure 10. Additionally,

Figure 10. Projections of experiments into PCA model (first


two principal components). Confidence ellipses for 50% (dotted
line), 95% (dashed line) and 99% confidence (solid line).

confidence intervals are calculated for the projections.


The T-scores are linear combinations of the measured
variables and according to the central limit theorem, it
should be considered normally distributed. Assuming
an approximate multivariate normality, it is possible to
define control limits (confidence intervals) for the Tscores. The variance of the first principal component
^1
vector, named T^1 , is given by the largest eigenvalue l
and the variance of the second principal component
^2 .
vector, named T^2 , is the second-largest eigenvalue l
Because both sample components are uncorrelated, a
confidence ellipse is equal to the possible pairs of values
(T^1 , T^2 ) (Johnson and Wichern, 2007)
T^12
T^ 2
+ 2  x22 (a)
^1
^2
l
l

20

where x 22 (a) is the upper (100a)th percentile of a chisquare distribution with two degrees of freedom at significance level a.
Looking at Figure 10, it is possible to distinguish
how different points lie far away from the main cluster
restrained by the elliptical contours defined as the confidence ellipses. This is a clear fault indication. The area
included in the confidence ellipses can be defined as the
normal operational region in the reduced space. The
closer to the origin of the reduced space the points are,
the more similar to the healthy structure (baseline
model) the structure is. In this way, all the data out of
the ellipses could be classified as abnormal (or defects).
However, the unfolding method used can hide some
features when only the projections into the principal
components are studied. For example, almost all the
points associated with the fifth damage case (the
most severe induced damage) lies into the 95% confidence ellipse. Then, studying only the projections into
the principal components may not be sufficient in order
to detect anomalous behavior in the system.

Downloaded from jim.sagepub.com by guest on January 20, 2015

12

Journal of Intelligent Material Systems and Structures

Q index

T 2 index

19,000
18,000

Figure 11. T2 index. Control limits (confidence intervals) for


95% (dashed line) and 99% confidence (solid line).

Nevertheless, it is possible to appreciate some differences between the projections (T-scores) of the data corresponding to damaged cases and the baseline. Due to
this kind of behavior with the T-scores in this particular
study it is necessary to integrate the results by means of
quantitative indices which allows us to re-express the
results in same metrics.
The T2 indices, which include the eigenvalues and
eigenvectors of the correlation matrix for all the experiments, are depicted in Figure 11. Besides that, a confidence region for 95% (dashed line) and 99% (solid line)
is also presented. The upper control limit is defined by
(Johnson and Wichern, 2007)
UCLT 2 = c2 = x 2r2 (a)

21

where r is the number of retained principal components in the PCA model.


Looking at Figure 11, it is possible to appreciate
that almost all the indices associated to damage cases
are far away from the baseline ones, and the indices
associated to undamaged cases (not used for building
the PCA model) fit very well with the baseline ones.
Again, some of the points corresponding to certain
damage cases lie under the 95% confidence region. One
should remember that the T2 index measures the variation of a particular sample within the PCA model. This
index only detects variations in the subspace defined by
the first five principal components, which are greater
than the variations explained by common causes. Then,
larger T2 values mean bigger unusual variations inside
the PCA model.
In the same way, the Q index for all the experimental trials is depicted in Figure 12. The Q index, as mentioned before, quantifies the change of the variations
that are not explained by the PCA model. Usually, the
T2 index is bigger than the Q index but the Q index is
more sensitive because it explains how well each sample
fits into the PCA model. Large Q index values mean
big unusual variations outside the PCA model.

Figure 12. Q index. Control limits (confidence intervals) for


95% (dashed line) and 99% confidence (solid line).

The main importance of the Q index lies in that it


provides additional information that is not included in
the T-scores or the T2 index (based on the T-scores).
For the Q index, which is a quadratic form of the
errors, and since these errors are well approximated by
a multi-normal distribution, the control limits (based
on Boxs equation which is an approximate distribution
for such quadratic forms; see Box, 1954, for a further
description) can be defined by (Nomikos and
MacGregor, 1995)
UCLQ =

y 
x 2 2 (a)
2m 2m =y

22

where x22m2 =y (a) is the upper (100a)th percentile of a


chi-square distribution with 2m2/y degrees of freedom
at significance level a, with m and y equal to the mean
and the variance of the Q index sample respectively.
From Figure 12 it is possible to appreciate that all
indices for different damage cases lie outside of the
defined control limits and are located far away from
the baseline and undamaged indices. At the same time
(like in the T2 or the T-scores cases), the undamaged
indices fit the baseline ones very well. From this index,
it would be possible to discern if an anomalous phenomenon (like damage) has occurred in the structure.
An important remark must be made at this point. As
shown before in previous figures, different damage
cases seem to exhibit nonlinear behavior, since the magnitude of their respective indices do not increase in
direct proportion to damage severity. This is because
under the high load magnitudes used in the experimentation, with some of the induced damage, the structure
shows local buckling. Then, the behavior exhibited
between some couples of sensors (like the example presented in Figure 2) is highly nonlinear under certain
conditions. Besides that, the multivariate unfolding
method used in this work combines all the experiments
performed for each damage case (i.e. all the load cases

Downloaded from jim.sagepub.com by guest on January 20, 2015

Sierra-Perez et al.

13

Figure 13. T2 index vs Q index. Control limits (condence


intervals) for 95% (dashed line) and 99% confidence (solid line).

for the j th sensor) and treats them as one variable.


Then, it is possible to mix linear cases with nonlinear
cases in one variable. Because of this, it is not surprising that the different indices present apparent nonlinear behavior.
In order to see the relationships between T2 indices
and Q indices, a plot of both indices against each other
is presented in Figure 13.
From this figure, it is clear that indices for all damage cases are far away from the baseline and the undamaged ones. It is possible to observe independent
clusters for some damage cases, for instance, damage 2.
This figure is very useful if neither the T2 index nor the
Q index are able by themselves to give information to
discern damage.

Concluding remarks
In order to achieve the first level of SHM by using
strain measurements in a structure with different load
cases, gathered from FBGs and by means of the application of PCA and different damage indices, several
scenarios were experimentally analyzed. The effectiveness of the proposed technique was tested using a complex wing structure fully made of composite materials.
A PCA baseline model was built using the readings
for the healthy structure. In subsequent steps, several
experiments were performed inducing five different
artificial kinds of damage to the structure. All these
experimental data were projected into the PCA model,
for which a selected number of principal components
were retained (T-scores). Finally, the T2 index and Q
index were calculated.
In order to deal with different load cases, a new
unfolding technique was developed. This technique
treats all the measurements of one sensor (regardless
of the load magnitude) as one variable (new sensor). Besides that, a new scaling (standardization)

methodology for these kinds of new variables was


proposed.
Since a small amount of damage slightly affects the
global stiffness and the local strain field of the structure, very high sensitivity techniques for strain measurements must be used in order to maximize the resolution
of the technique. FBGs have shown to be very sensitive
to small strain changes in the structure, which makes
them suitable for the proposed technique.
In all the cases (T-scores, T2 indices, Q indices and
2
T vs Q) it was possible to detect some deviations
between the baseline (and the undamaged case) and the
different damage cases. The Q index has shown more
sensitivity in this study.
A general methodology for fault detection could
consist of the following steps: the first step consists in
testing the Q index. If Q index is significant, a fault condition may be assumed and the procedure is completed.
On the other hand, if the Q index is not significant, this
is an indication that the PCA model holds the same
characteristics as the actual structure and the T2 index
must then be tested. If T2 index is significant, a fault
condition can be assumed and the procedure is complete. If the T2 index is not significant, then the third
step must be performed. In the third step, the individual
principal components should be tested in order to determine the nature of the irregularity. However, it is advisable to study and analyze all the indices and the Tscores simultaneously. This could increase the accuracy
of the damage detection.
Funding
This work has been supported by the Ministerio de Ciencia e
Innovacion in Spain (coordinated research project DPI201128033-C03-03).

Declaration of Conflicting Interests


The authors declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.

References
Adams D (2007) Health Monitoring of Structural Materials
and Components. New York, NY: John Wiley and Sons.
Balageas D, Fritzen C and Guemes A (2006) Structural
Health Monitoring. London: ISTEHermes Science.
Baseville M, Benveniste A, Goursat M, et al. (2007) In-flight
vibration monitoring of aeronautical structures. IEEE
Control Systems Magazine (27): 2742.
Box G (1954) Some theorems on quadratic forms applied in
the study of analysis of variance problems: Effect of
inequality of variance in one-way classification. The
Annals of Mathematical Statistics (25): 290302.
Culshaw B and Kersey A (2008) Fiber optic sensing: A historical perspective. IEEE/OSA Journal of Lightwave Technology 26: 10641078.

Downloaded from jim.sagepub.com by guest on January 20, 2015

14

Journal of Intelligent Material Systems and Structures

Doebling S, Farrar C and Prime M (1998) A summary review


of vibration-based damage identification methods. The
Shock and Vibration Digest (30): 91105.
Fernandez A, Menendez J and Guemes A (2007) Damage
detection in a stiffened curved plate by measuring differential strains. In: International conference on composite
materials.
Fritzen CP and Bohle K (2003) Global damage identification
of the steelquake structure using modal data. Mechanical
Systems and Signal Processing (17): 111117.
Fritzen C-P and Kraemer P (2009) Self-diagnosis of smart
structures based on dynamical properties. Mechanical Systems and Signal Processing (23): 18301845.
Fritzen CP, Jennewein D and Kiefer T (1998) Damage detection based on model updating methods. Mechanical Systems and Signal Processing (12): 163186.
Fritzen CP, Li HN and Li DS (2009) A note on fast computation of effective independence through QR downdating for
sensor placement. Mechanical Systems and Signal Processing (23): 11601168.
Garc a C (2010) Caracterizacion de coeficientes de expansion
termica. (In Spanish.) Technical report, Universidad Politecnica de Madrid, Spain.
Ghatak A and Thyagarajan K (1998) Introduction to Fiber
Optics. Cambridge: Cambridge University Press.
Guemes A and Menendez J (2006) Fiber optics sensors. In: D
Balageas, C Fritzen and A Guemes (eds) Structural Health
Monitoring. New York, NY: Wiley.
Guemes A and Sierra J (2013) Fiber optic sensors. In: W
Ostachowicz and A Guemes (eds) New Trends in Structural Health Monitoring. New York, NY: Springer.
Gurden S, Westerhuis J, Bro R, et al. (2001) A comparison of
multi-way regression and scaling methods. Chemometrics
and Intelligent Laboratory Systems (59): 121136.
Hill K and Meltz G (1997) Fiber Bragg grating technology
fundamentals and overview. IEEE/OSA Journal of Lightwave Technology 15: 12631276.
Holnicki-Szulc J (2008) Smart Technologies for Safety Engineering. New York, NY: John Wiley & Sons.
Huedo T, Sanchez J, Mar n F, et al. (2006) Assessing heterogeneity in meta-analysis: Q static or I2 index. Psychological
Methods (11): 193206.
Jackson E and Mudholkar G (1979) Control procedures for
residual associates with PCA. Technometrics (21): 341349.
Johnson R and Wichern D (2007) Applied Multivariate Statistical Analysis. London: Pearson.
Jollife I (2002) Principal Component Analysis. New York, NY:
Springer.
Kourti T and MacGregor J (1995) Process analysis, monitoring and diagnosis, using multivariate projection methods.
Chemometrics and Intelligent Laboratory Systems (28):
321.
Li DS, Li HN and Fritzen CP (2012) Load dependent sensor
placement method: Theory and experimentation validation. Mechanical Systems and Signal Processing (31):
217227.
Lopez I and Sarigul-Klijn N (2010) A review of uncertainty in
flight vehicle structural damage monitoring, diagnosis and

control: Challenges and opportunities. Progress in Aerospace Sciences 46: 247273.


Manson G, Worden K, Holford K, et al. (2001) Visualization
and dimension reduction of acoustic emission data for
damage detection. Journal of Intelligent Material Systems
and Structures 12(8): 529536.
Menendez J and Guemes A (2005) Embedded fibre Bragg
gratings for design and manufacturing optimisation of
advanced composites. In: Proceedings of the 4th IWSHM.
Mujica L, Rodellar A, Fernandez A, et al. (2011) Q-statistic
and T2-statistic PCA-based measures for damage assessment in structures. Structural Health Monitoring 10:
539553.
Mujica L, Tibaduiza D and Rodellar J (2010) Data driven
multiactuator piezoelectric system for structural damage
localization. In: Fifth world conference on structural control
and monitoring, pp. 113.
Mujica L, Veh J, Ruiz M, et al. (2008) Multivariate statistical
process control for dimensionality reduction in structural
assessment. Mechanical Systems and Signal Processing
(22): 155171.
Nomikos P and MacGregor J (1994) Monitoring batch processes using multi-way principal component analysis.
AIChE Journal (40): 13611375.
Nomikos P and MacGregor J (1995) Multivariate SPC charts
for monitoring batch processes. Technometrics (37): 4159.
Peters K (2009) Fiber Bragg grating sensors. In: Boller C,
Chang FK and Fujino Y (eds) Encyclopedia of Structural
Health Monitoring. New York, NY: Wiley.
Sierra J and Guemes A (2011) Deteccion de dano en materiales compuestos mediante fibra optica. (In Spanish.) In:
Actas del IX congreso nacional de materiales compuestos,
pp. 631636.
Sierra J, Guemes A and Mujica L (2013) Damage detection
by using FBGs and strain field pattern recognition techniques. Smart Materials and Structures (22): 110.
Sohn H, Worden K and Farrar C (2002) Statistical damage
classification under changing environmental and operational conditions. Journal of Intelligent Material Systems
and Structures 13: 561574.
Suleman A (2011) An SHM overview. In: New trends in structural health monitoring, pp. 156.
Takeda N, Okabe Y, Kuwahara J, et al. (2005) Lamb wave
sensing using fiber Bragg grating sensors for delamination
detection in composite laminates. In: Proceedings of the
SPIE.
Villez K, Steppe K and De Pauw D (2009) Use of unfold PCA
for on-line plant stress monitoring and sensor failure detection. Biosystems Engineering (103): 2334.
Westerhuis J, Kourti T and MacGregor J (1999) Comparing
alternative approaches for multivariate statistical analysis
of batch process data. Journal of Chemometrics (13): 397
413.
Wold S, Kettaneh N, Friden H, et al. (1998) Modelling and
diagnostics of batch processes and analogous kinetic
experiments. Chemometrics and Intelligent Laboratory Systems (44): 331340.

Downloaded from jim.sagepub.com by guest on January 20, 2015

You might also like