See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/228907097

Sensor validation using principal component

analysis

Article in Smart Materials and Structures February 2005

DOI: 10.1088/0964-1726/14/1/004

CITATIONS READS

78 446

4 authors, including:

Gatan Kerschen J.-C. Golinval

University of Lige University of Lige
255 PUBLICATIONS 4,581 CITATIONS 224 PUBLICATIONS 3,866 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

OUFTI-1 CubeSat View project

ERC Starting Grant View project

All content following this page was uploaded by Gatan Kerschen on 19 July 2017.

The user has requested enhancement of the downloaded file.

INSTITUTE OF PHYSICS PUBLISHING SMART MATERIALS AND STRUCTURES
Smart Mater. Struct. 14 (2005) 3642 doi:10.1088/0964-1726/14/1/004

Sensor validation using principal

component analysis
Gaetan Kerschen1,3, Pascal De Boe1 , Jean-Claude Golinval1 and
Keith Worden2
1
Department of Materials, Mechanical and Aerospace Engineering, University of Liege,
Chemin des Chevreuils 1 (B52), B-4000 Liege, Belgium
2
Dynamics Research Group, Department of Mechanical Engineering, University of
Sheffield, Sheffield S1 3JD, UK

E-mail: g.kerschen@ulg.ac.be

Received 25 August 2003, in final form 28 July 2004

Published 26 November 2004
Online at stacks.iop.org/SMS/14/36
Abstract
For a reliable on-line vibration monitoring of structures, it is necessary to
have accurate sensor information. However, sensors may sometimes be
faulty or may even become unavailable due to failure or maintenance
activities. The problem of sensor validation is therefore a critical part of
structural health monitoring. The objective of the present study is to present
a procedure based on principal component analysis which is able to perform
detection, isolation and reconstruction of a faulty sensor. Its efficiency is
assessed using an experimental application.

1. Introduction output of a given sensor to look for indications of faults. This

It is widely recognized that the capability of quickly assessing
structural health is important economically. To this end,
the last few years have witnessed the appearance of smart
structures. Such structures are instrumented with a large
number of embedded sensors which continuously measure the
structural response (see e.g. [1, 2]).
For a reliable on-line vibration monitoring of structures,
it is necessary to have accurate sensor information. However,
sensors may sometimes be faulty (e.g. gain fault or damaged
sensor) or may even become unavailable due to failure or
maintenance activities. The problem of sensor validation is
therefore a critical part of structural health monitoring. As
pointed out in [3], this field of research has received little
attention in the structural dynamics community compared to
the process control or chemical engineering communities (see
e.g. [47]).
There are essentially three approaches to sensor
validation. The first is commonly adopted in the process
control community and can be found described in [4, 5, 8];
it relies on redundancy of information between different
sensors. These sensors are of a standard simple design,
i.e. accelerometers, and it is actually the sensor network which
is self-validating. The second approach is to examine the
3 Author to whom any correspondence should be addressed.
is the basis of the measurement aberration detection (MAD)
scheme described in [9]. The third approach is to have
a complex sensor which is itself self-validating. This is
the approach adopted by Henry and Clarke in the SEVA
programme [10, 11]; a concrete example of this approach is
provided by the Coriolis flowmeter described in [12].
In this study, sensor validation refers to the capability
of detecting, isolating and reconstructing a faulty sensor.
The requirement of reconstructability has implications for the
design of sensor networks which should be constructed to be
fail-safe as far as possible [13, 14]. The present paper aims
at presenting a procedure based upon principal component
analysis which is able to tackle all these objectives. The
proposed methodology is data driven, i.e., it does not require
a structural model, which is an important feature for practical
applications. It only assumes that the number of sensors is
large enough to guarantee redundancy in the measured data
set.
2. Principal component analysis
Principal component analysis (PCA) is a multivariate analysis
technique which was first introduced by Pearson [15] in 1901
and developed independently by Hotelling [16] in 1933. It is
also closely related to proper orthogonal decomposition, also
known as the KarhunenLoeve transform.

0964-1726/05/010036+07$30.00 2005 IOP Publishing Ltd Printed in the UK 36


Given an observed n S -dimensional response vector x, the
goal of PCA is to reduce the dimensionality of x. This is
realized by finding r principal axes pi with i = 1, . . . , r
onto which the retained variance under projection is maximal.
These axes, denoted as principal directions or PCA modes,
are given by the eigenvectors associated with the r largest
eigenvalues of the covariance matrix :
= E[(x )(x )T ] (1)
where E[] is the expectation and = E[x] is the mean of the
data.
If the principal directions are collected in a matrix P =
[p1 pr ], then z = PT (x ) is a reduced r -dimensional
representation of the observed vector x. Among all linear
techniques, PCA provides the optimal reconstruction x =
+ Pz of x in terms of the quadratic reconstruction error

x x
2 .
It is worth pointing out that PCA is also closely related
to singular value decomposition. If the mean is subtracted
from the data and if m observations are collected in a matrix
X (n S rows and m columns), then the left singular vectors of
X, as eigenvectors of XXT , are the principal directions. The
singular values indicate how the corresponding left singular
vectors participate and the right singular vectors are the time
coefficient vectors of the principal directions.
3. Sensor fault detection and isolation
The purpose of this section is to investigate the problem of
sensor fault detection and isolation in structural dynamics.
Friswell and Inman [3] proposed two approaches based on
the comparison between the subspace of the response and
the subspace generated by the lower modes of a structural
model. In [8], the detection of sensor failures relies on an
auto-associative neural network which is known to implement
PCA. The inputs and the targets of the neural network are
merely the measured signals, and the network is trained using
data which are known to be healthy. During the detection
stage, the residual vector between the inputs and the outputs
of the network is computed. If the residual of a sensor exceeds
a certain threshold whilst the residuals of the other sensors
remain relatively low, then that sensor is declared as faulty.
In this paper, an approach similar to that adopted
in [17] for damage detection using a piezo-sensor array is
considered. From the mode acceleration method [18], the
system response x(t) to external loading f(t) with a limited
frequency bandwidth (  k+1 . . .) can be expressed through
mode superposition:
 t

k
i Ti
x(t) = f( ) sin(i (t )) d
i=1
i i 0
 

n
i Ti
+ f(t)
2
i=k+1 i i


k century [19]. The principal angles between two subspaces
= i i (t) + rres (t) (2)
i=1
where i , i and i represent the mode shapes, natural
frequencies and generalized masses, respectively. The vector
r is the residue associated with the modes located outside the
Sensor validation using principal component analysis
Figure 1. Principal angles between two planes.
frequency band, and the variables are the modal amplitudes.
This shows that the system response contains the contribution
of the modes which respond dynamically and the contribution
of the residual modes which respond statically. Equation (2)
suggests that the subspace spanned by the system response
does not depend on the excitation signal f(t), provided that the
modes in the frequency bandwidth participate in the system
response. It should be noted that the system is also assumed
to be linear and time invariant.
Our interest in PCA lies in the fact that it offers a
means of computing the subspace spanned by the data without
the need to compute the mode shapes. The singular value
decomposition of the response matrix X is computed:
a1 (t1 ) a1 (tm )
. ..
X = USVT = [u1 un S ] .. . (3)
an S (t1 ) an S (tm )
where U is an (n S n S ) orthonormal matrix containing the
left singular vectors; S is an (n S m) pseudo-diagonal and
semi-positive definite matrix with diagonal entries containing
the singular values, and V is an (m m) orthonormal matrix
containing the right singular vectors. The foregoing section
indicates that the columns ui of matrix U are the principal
directions and that ai (t) are the corresponding time coefficient
vectors scaled by the singular values. Although the principal
directions do not generally coincide with the mode shapes,
equations (2) and (3) show that the subspace spanned by
the active principal directions, i.e. those with non-zero
time vectors or, equivalently, those associated with non-zero
singular values, is identical to the subspace spanned by the k
mode shapes and by the residual vector r. Obviously, this is
true provided that the number of excited mode shapes is lower
than the number of sensors n S .
The sensor fault detection methodology consists in
comparing the subspace spanned by the reference data
which are assumed to be healthy to the subspace spanned
by the current data. A very efficient way of performing
the comparison is to use the concept of principal angles
between subspaces introduced by Jordan in the nineteenth
are a generalization of an angle between two vectors, and their
number is equal to the dimension of the smallest subspace. An
illustration of this concept for two planes is given in figure 1.
There are two principal angles 1 and 2 , one of which is zero
because the two planes intersect.
37
G Kerschen et al

Figure 2. Comparison of the principal angles of the current and reference data.

A numerical algorithm for the computation of the reference data set contains enough information to cover normal
angles involving a QR factorization and the singular value process operation. The most likely value for the errant
decomposition was proposed by Bjorck and Golub [20] and sensor is defined as the value which minimizes the magnitude
can also be found in [21]. It should be noted that the of the deviation between the response vector x(t) and its
angles have already found several applications in electrical reconstruction x(t). In the approach pioneered by Kramer [4]
engineering [22, 23] and in structural dynamics [3, 17, 24, 25]. and in [28], the reconstruction is performed using non-linear
In this study, our interest lies in the largest angle which allows PCA, but in this study, we restrict ourselves to the linear PCA
one to quantify how the subspaces are globally different. described in section 2. Replacement of a sensor value thus
In practice, due to the noise inherent in a measurement involves finding the value of x j (t) such that
process, the largest angle between the subspaces spanned by
the reference and current data will not be exactly zero even if all min J =
x(t) x(t)
2 . (4)
xj
sensors are functioning correctly. Before applying the sensor
validation process, the reference data are thus partitioned into Equation (4) is a univariate optimization problem. If
several sets. The principal angle between the subspace spanned there is more than one missing sensor at a time, it requires
by each of these sets and the subspace spanned by the whole a multivariate approach. It is however interesting to note that
data set is computed, which gives us a collection of different there is an analytical solution to this problem [8]. Suppose
angle values. When dealing with the current data set, an again that the j th sensor fails and that the value to be optimized
alarm is issued when the monitored angle exceeds the upper is labelled x . The objective function is
control limit (UCL) defined as the mean angle plus three
times its standard deviation (see outlier statistics [26, 27]). J = (x1 x1 )2 + + (x j x )2 + + (xn S xn S )2 (5)
This corresponds to a 99.7% confidence interval for a normal
distribution. This is illustrated in figure 2. with
When an alert is given, the faulty sensor is then isolated by x = PPT x = Ax (6)
removing one by one the sensors from both the reference and for zero-mean data (we recall that matrix P contains the
current data sets. The angle should then be minimum when principal directions). Now, defining A = A I or ai j =
the faulty sensor is discarded. ai j i j where I is the identity matrix, the objective function
A final remark concerns the normalization of the data. becomes
When using PCA, it is often advised to deal with equally


nS

2

important variables (e.g. with unit variance). In some cases J = ak1 x1 + + ak j x + + akn

x
S nS
. (7)
(e.g. sensor gain fault), it should be kept in mind that the k=1
normalization might mask the fault.
The minimum with respect to x is obtained by setting
J /x = 0. This gives
4. Sensor correction


nS


Let us suppose that n S sensors are available giving a response ak j ak1 x1 + + ak j x + + akn

x = 0.
S nS
(8)
vector x(t) and that for some reason the j th sensor fails. Let k=1
us also assume that the errant sensor has been identified. If the
After a little rearrangement this gives
response given by this sensor contains important information,
correction is then necessary. It is first assumed that the x1 {a j }T {a1 } + + x {a j }T {a j } + + xn S {a j }T {an S } = 0 (9)

38

Figure 3. Geometric interpretation of the sensor correction procedure.
Figure 4. The experimental set-up.
(This figure is in colour only in the electronic version)
where {ak } is the kth column of A . It follows that
S Length (m) Width (m) Thickness (m) Material
{a j }T nk= 
j =1 x k {ak }
x = . (10)
{a j }T {a j }
From a geometric point of view, this procedure amounts to
finding the intersection between a straight line and the subspace
spanned by the PCA modes. The straight line is parallel to the
axis of the faulty sensor and is defined by the coordinates of
the remaining sensors. This is pictured in figure 3 where the
first sensor is unavailable, i.e., x1 = 0.
Finally, it should be noted that, as in the previous section,
the number of sensors must be greater than the number
of structural modes involved in order to guarantee enough
redundancy in the data.
5. Experimental application
5.1. Description of the experimental structure
The benchmark is similar to the one proposed by the Ecole
Centrale de Lyon (France) in the framework of COST
Action F3 working group on Identification of non-linear
systems [29]. This experimental application involves a
clamped beam with a thin beam part at the end of the main
beam (cf figure 4). The geometrical and mechanical properties
of the set-up are listed in table 1.
Seven accelerometers which regularly span the beam are
used to measure the response. The excitation force provided
by an electrodynamic shaker is a white-noise sequence band-
limited in the 0500 Hz range. Due to the thin beam part,
the effect of gravity is not negligible. In order to reduce its
influence, a set-up in which the thin beam is vertical and the
shaker, located at position 3, excites the structure in a horizontal
plane is considered.
Sensor validation using principal component analysis
Table 1. Geometrical and mechanical properties of the set-up.
Main beam 0.7 0.014 0.014 Steel
Thin beam part 0.04 0.014 0.0005 Steel
An interesting feature is that, if the structure may be
assumed to be linear for low excitation levels, this is no longer
the case for higher levels. Indeed, if the excitation level is
increased, the thin part is excited in large deflection and a
geometrical non-linearity is activated.
Two different kinds of sensor fault are simulated. Firstly,
the acceleration measured at the third sensor is multiplied by
1.2 (gain fault). Secondly, it is replaced by a white-noise
sequence of the same variance (sensor failure).
5.2. Sensor fault detection and isolation
5.2.1. Reference data. The reference data set contains 70 000
points from each of the seven channels. It corresponds to
an excitation level equal to 1.4 N for which the structural
behaviour is linear. An important thing to check is that there
are enough sensors in relation to the number of excited mode
shapes. Table 2 lists the normalized singular values and their
cumulative sum. Two singular values are zero which ensures
enough redundancy in the data set. It should be noted that
a modal analysis has revealed the presence of three natural
frequencies in the 0500 Hz range. This is relatively well
reflected by the cumulative sum of the singular values as
the first three singular values already account for 98.66% of
the total energy. The remaining 1.34% may be attributed
to the (small) participation of the modes outside the excited
frequency range but also to the experimental noise.
The next step is to estimate the UCL. To this end, the data
are partitioned into 28 different sets containing 2500 points
39
G Kerschen et al

than (or equal to) the true dimensionality of the system. It

is therefore decided to keep three PCA modes in the analysis;
they capture almost 99% of the energy, and the mean angle
is almost equal to zero. For a three-dimensional subspace, an
alarm will thus be issued when the monitored angle will exceed
the following angle (in degrees):

UCL3 = 0.65 + 3 0.46 = 2.03. (11)

5.2.2. Current data. The current data set contains 40 000

Figure 5. Mean principal angle versus number of retained PCA
modes.
Table 2. Singular values and their cumulative sum.
Singular values (%) Cumulative sum (%)
1 68.59 68.59
2 22.07 90.66
3 8.00 98.66
4 1.30 99.96
5 0.04 100
6 0 100
7 0 100
each which gives us a collection of 28 different angle values.
Figure 5 shows the average value as a function of the subspace
dimension. This graph is used for choosing the appropriate
subspace dimension. For seven PCA modes, the angle is
exactly zero, which is expected because the data are seven
dimensional. Obviously, it is not recommended to choose this
number of modes. It is also interesting to notice that the angle
suddenly increases from a dimension equal to 5 to a dimension
equal to 6. There is a simple geometric explanation for this
phenomenon. Suppose that we have a line and that we want
to embed this one-dimensional object into a two-dimensional
plane. The problem is ill-posed, and actually there are an
infinite number of planes which fulfil this condition. As a
result, the principal angle between two such planes can take
an arbitrary value from 0 to 90 . This clearly demonstrates
that the subspace dimension must always be chosen smaller
points whose last 20 000 correspond to a sensor fault. The
current data set is partitioned into 16 different sets containing
2500 points each. Figure 6 displays the principal angle
between the subspace spanned by each of these sets and the
subspace spanned by the whole reference set for both sensor
faults. In both cases, it can be observed that the first 20 000
points are well below the UCL, whereas this is no longer the
case for the remainder of the data set. Note that the sensor
failure is much easier to identify than the gain fault.
Now that an alarm has been issued, the faulty sensor needs
to be identified using the methodology described previously.
Figure 7 presents all the angles obtained when the sensors are
removed one by one. For both faults, it clearly appears that the
angle approaches zero when the third sensor is removed.
It is worth pointing out that there is a even simpler way to
identify the faulty sensor in the case of a gain fault. The strategy
is based on the observation that the intersection between the
reference subspace and the current subspace contains data
points which are not affected by the fault. In other words,
data points at the intersection must have a zero component at
the faulty sensor. In addition, since we know the angle between
the subspaces, it is possible to obtain an estimation of the gain
error using the cosine rule. Because a gain fault is certainly
not the most common sensor failure, this is not investigated
further.
5.2.3. Non-linear behaviour. The whole procedure has also
been tested for when the reference data set corresponds to an
excitation level for which the system is non-linear (22 N).
Surprisingly enough, the results obtained were satisfactory
then as well. This is expected as long as the excitation level
remains the same for both the reference and current data.
For instance, the results for the detection stage are shown in
figure 8.

Figure 6. The monitored angle. (a) Gain fault; (b) sensor failure.

40
 Sensor validation using principal component analysis

Figure 7. Results of the isolation stage. (a) Gain fault; (b) sensor failure.

Figure 8. The monitored angle for a non-linear structural behaviour. (a) Gain fault; (b) sensor failure.

Table 3. Optimum number of PCA modes.

Number of modes MSE (%)
1 71.11
2 70.93
3 5.60
4 0.99
5 0.13
6 0.14

5.3. Sensor correction

5.3.1. Linear behaviour. Now that the third sensor has
been declared as faulty, the next step is to retrieve its original
response. As underlined in the previous section, we have at our
Figure 9. Sensor correction (linear behaviour). Solid curve:
disposal 70 000 reference data points. First, the optimum num-
original response; dotted curve: reconstructed response.
ber of retained PCA modes is determined using a sort of cross-
validation. More precisely, the response of the third sensor for
the last 20 000 points is predicted using equation (10) and the 6. Concluding remarks
modes of the first 50 000 points. The number of modes which
gives the minimum mean square error (MSE) is then selected. The object of the present paper was to explore the use of PCA
The results are given in table 3 and suggest five PCA modes.
for sensor fault detection, isolation and correction. For each of
The reconstructed response of the faulty sensor is now
these three objectives, reasonable consistency and accuracy has
compared to its original response in figure 9. 20 000 points
been observed in an experimental application. Furthermore,
were faulty, but for the sake of clarity only 500 points are
represented. It can be observed that the two curves agree to the procedure also seems to work well when dealing with non-
the point where the difference between them is not visible linear structural behaviour provided that the excitation levels
(MSE = 0.04%). are approximately the same for the reference and current data.
In order to integrate this approach in the framework
5.3.2. Non-linear behaviour. Sensor correction has also been of structural health monitoring, one has still to be able to
carried out for non-linear behaviour. Again, the results are discriminate between a sensor fault and structural damage.
almost as accurate as for the linear behaviour (MSE = 0.18%). This issue will be carefully investigated in future work.

41
 G Kerschen et al
Acknowledgment
The author G Kerschen was supported by a grant from the
Belgian National Fund for Scientific Research (FNRS) which
is gratefully acknowledged.
References
[1] Guran A and Inman D J 1995 Smart Structures, Nonlinear
Dynamics, and Control (Englewood Cliffs, NJ:
Prentice-Hall)
[2] Worden K, Bullough W A and Haywood J 2003 Smart
Technologies (Singapore: World Scientific)
[3] Friswell M and Inman D J 1999 Sensor validation for smart
structures J. Intell. Mater. Syst. Struct. 10 97382
[4] Kramer M A 1992 Autoassociative neural networks Comput.
Chem. Eng. 16 31328
[5] Dong D and McAvoy T J 1996 Nonlinear principal component
analysisbased on principal curves and neural networks
Comput. Chem. Eng. 20 6578
[6] Qin S J, Yue H and Dunia R 1997 Self-validating inferential
sensors with application to air emission monitoring Ind.
Eng. Chem. Res. 36 167585
[7] Guo T H, Wrest D J and Uhrig R E 1996 Plant wide sensor
calibration monitoring Proc. 1996 IEEE Int. Symp. on
Intelligent Control (Dearborn, USA) (Piscataway, NJ:
IEEE) pp 37883
[8] Worden K 2003 Sensor validation and correction using
auto-associative neural networks and principal component
analysis Proc. IMAC XXI (Orlando, USA)
[9] Yung S K and Clarke D W 1989 Sensor validation Meas.
Control 22 13250
[10] Henry M and Clarke D W 1993 The self-validating sensor:
rationale, definitions and examples Control Eng. Pract. 1
585610
[11] Henry M 1994 Validating data from smart sensors Control
Eng. 41 (9) 636
[12] Henry M 1995 Self-validation improves Coriolis flowmeter
Control Eng. 42 (6) 816
[13] Side S, Staszewski W J, Wardle R and Worden K 1997
Fail-safe sensor distributions for damage detection
DAMAS 97: Proc. 2nd Int. Workshop on Damage
42
View publication stats
Assessment using Advanced Signal Processing Procedures
(Sheffield, UK) pp 13546
[14] Staszewski W J, Wardle R, Worden K and Tomlinson G R
2000 Fail-safe sensor distributions for impact detection in
composite structures Smart Mater. Struct. 9 298303
[15] Pearson K 1901 On lines and planes of closest fit to systems of
points in space Phil. Mag. 2 55972
[16] Hotelling H 1933 Analysis of a complex of statistical variables
into principal components J. Educ. Psychol. 24 41741
Hotelling H 1933 J. Educ. Psychol. 24 498520
[17] De Boe P and Golinval J C 2003 Principal component analysis
of a piezo-sensor array for damage localization Struct.
Health Monit. 2 13744
[18] Geradin M and Rixen D 1994 Mechanical Vibrations, Theory
and Application to Structural Dynamics (Chichester: Wiley)
[19] Jordan C 1875 Essai sur la geometrie a n dimensions Bull. Soc.
Math. 3 10374
[20] Bjorck A and Golub G H 1973 Numerical methods for
computing angles between linear subspaces Math. Comput.
27 57994
[21] Golub G H and Van Loan C F 1996 Matrix Computations
(Baltimore, MD: Johns Hopkins University Press)
[22] De Cock K and De Moor B 2002 Subspace angles for fault
detection ISMA 2002: Proc. Int. Conf. on Noise and
Vibration Engineering (Leuven, Belgium)
[23] De Cock K 2002 Principal angles in system theory,
information theory and signal processing PhD Thesis
Katholieke Universiteir Leuven
[24] Yang M and Brown D 1996 Model updating techniques using
perturbed boundary conditions testing data Proc. IMAC XIV
(Detroit, USA)
[25] Garvey S D, Penny J E T and Friswell M I 1996 Quantifying
the correlation between measured and computed mode
shapes J. Vib. Control 2 12344
[26] Worden K, Manson G and Fieller N R J 2000 Damage
detection using outlier analysis J. Sound Vib. 229 64767
[27] Sohn H, Czarnecki J A and Farrar C R 2000 Structural health
monitoring using statistical process control J. Struct. Eng.
126 135663
[28] Kerschen G and Golinval J C 2003 Auto-associative neural
networks: Part I. Replacement of missing sensor values
Proc. 8th Int. Conf. on Recent Advances in Structural
Dynamics (Southampton, UK)
[29] http://www.ulg.ac.be/ltas-vis/costf3/costf3.html