Engineering Structures 25 (2003) 1785–1802

www.elsevier.com/locate/engstruct

Quantification and localisation of damage in beam-like structures

by using artificial neural networks with experimental validation
M. Sahin, R.A. Shenoi ∗
School of Engineering Sciences, Ship Science, University of Southampton, Highfield, Southampton SO17 1BJ, UK

Received 19 May 2003; received in revised form 1 August 2003; accepted 4 August 2003

Abstract

This paper presents a damage detection algorithm using a combination of global (changes in natural frequencies) and local
(curvature mode shapes) vibration-based analysis data as input in artificial neural networks (ANNs) for location and severity predic-
tion of damage in beam-like structures. A finite element analysis tool has been used to obtain the dynamic characteristics of intact
and damaged cantilever steel beams for the first three natural modes. Different damage scenarios have been introduced by reducing
the local thickness of the selected elements at different locations along finite element model (FEM) of the beam structure. The
necessary features for damage detection have been selected by performing sensitivity analyses and different input–output sets have
been introduced to various ANNs. In order to check the robustness of the input used in the analysis and to simulate the experimental
uncertainties, artificial random noise has been generated numerically and added to noise-free data during the training of the ANNs.
In the experimental analysis, two steel beams with eight distributed surface-bonded electrical strain gauges and an accelerometer
mounted at the tip have been used to obtain modal parameters such as resonant frequencies and strain mode shapes. Finally, trained
feed-forward backpropagation ANNs have been tested using the data obtained from the experimental damage case for quantification
and localisation of the damage.
 2003 Elsevier Ltd. All rights reserved.

Keywords: Damage identification; Vibration-based analysis; Strain mode shape; Finite element analysis; Artificial neural networks

1. Introduction changes in natural frequencies, mode shapes and struc-

Non-destructive inspection techniques are generally
used to investigate the critical changes in the structural
parameters so that an unexpected failure can be pre-
vented. These methods concentrate on a part of the struc-
ture and in order to perform the inspection, the structure
needs to be taken out of service. Since these damage
identification techniques require a large amount of
human intervention, they are passive and costly methods.
On the other hand, damage can be detected, quantified
and localised by on-line damage assessment techniques
using vibration-based analysis data in the service life of
a structure.
The effects of common damage on a structure are
∗
Corresponding author. Tel.: +44-2380-592356; fax: +44-2380-
593299.
E-mail address: r.a.shenoi@ship.soton.ac.uk (R.A. Shenoi).
tural damping. Since the measurement of natural fre-
quencies is easier than that of changes in structural
damping, damage can be detected from dynamic analysis
using natural frequencies and mode shapes. Damage
detection techniques based on changes in the structural
dynamic characteristics have been reviewed [1–3] and
studied [4–7]. Although changes in natural frequency
give a useful indication of the existence of damage, it
can be concluded that a damage identification method
using natural frequency change is not sufficient to locate
the damage. Therefore, mode shape information is
required to determine the location of the damage. Since
displacement [8,9] and curvature [9–13] mode shapes
can be obtained by both experimental as well as numeri-
cal means, several researchers [14–17] have focused on
damage detection techniques which use this fact.
It is seen from previous studies that changes in natural
frequencies can be used to characterise the severity of
the damage. However, locating the damage from the

0141-0296/$ - see front matter  2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2003.08.001
1786 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Nomenclature
E elastic modulus
h distance between two collocation points
i number of data points
N number of random entries
vi normalised displacement
v⬙i curvature
g gravitational acceleration

Greek symbols

e strain
r density
n Poisson’s ratio
⌬ noise percentage

Abbreviations

ANN artificial neural network

DL location of the damage
DS severity of the damage
FEA finite element analysis
FEM finite element model
LOC location where the maximum absolute difference in curvature mode shape occurs
MADC maximum absolute difference in curvature mode shape
RNF reduction in natural frequency

changes in natural frequencies alone is difficult as modal
frequencies are global properties of the structure and
hence cannot provide spatial information about structural
changes. Therefore, displacement or curvature mode
shapes that provide spatial information about the damage
are needed in order to overcome this drawback. In
addition to this, multiple modes need to be considered
for better estimation of severity and location of the dam-
age since each natural frequency and corresponding
mode shape are affected to different extents depending
on the location of the damage. As damage detection is
an inverse, non-linear and non-unique problem, that gen-
erally does not have a feasible algorithmic solution or
for which an algorithmic solution is too complicated to
be found and handled, a different approach is required.
Therefore, artificial neural networks (ANNs) [18–22]
offer capabilities such as self-adaptiveness, generalis-
ation, abstraction and suitability for real-time appli-
cations. They have been widely used to estimate the
damage severity and its location in monotonic [23–31]
and composite [32–37] structures.
Different vibration-based analysis data have been used
either directly or as an input feature for ANN appli-
cations for different levels of damage assessment.
Experimental validation however is quite limited. The
technique used here aims to test the efficacy of global
or local vibration parameters and their different combi-
nations as input to feed-forward backpropagation ANNs
for accurate prediction of damage location and severity.
The global parameters are changes in natural fre-
quencies, and the local parameters are changes in curva-
ture mode shapes of the structure in the first three natural
modes. To generate these parameters, free vibration
dynamic behaviour of intact and damaged beams has
been obtained from FEA, and an experimental procedure
has been applied to cantilever steel beams to validate the
proposed method.
2. Adopted approach
The adopted vibration-based damage detection pro-
cedure can be divided into three main parts: numerical
analysis based on finite element model (FEM), experi-
mental analysis, and application of the designed ANNs
in the prediction of damage severity and location.
The numerical analysis concentrates on extraction and
selection of features from normal mode dynamic analy-
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802
sis performed using the FEM for cantilever steel beams.
Different damage scenarios are obtained by changing the
thickness of the selected elements by different extents at
different locations. The outputs from FEA are the natural
frequencies and corresponding normalised displacement
mode shapes in the first three vibration modes. After per-
forming sensitivity analysis, the normalised natural fre-
quencies of the damaged beam with respect to the intact
one, and the maximum absolute difference in curvature
mode shapes between the intact and the damaged beams,
are selected as input features. Such features [38] are
good indicators for damage assessment.
The experimental component of this work examines
intact and damaged cantilever steel beam specimens,
with eight surface-bonded electrical strain gauges and
an accelerometer attached at the free end. Experimental
specimens have the same geometrical and material
properties with the ones used in the numerical analysis.
The experimental study also aims to use damage sensi-
tive signal features obtained from the sensory structures
(i.e. measured modal data such as resonant frequencies
and normalised strain mode shapes in the first three
modes) as an input to selected ANNs for the experi-
mental verification of the proposed method.
The ANNs used in the analysis (Table 5) are selected
from the previous numerical work [38] performed on
composite beam models; and chosen on the basis of
minimum mean square error and better convergence with
an increase in the number of epochs (i.e. the number of
presentations of the training input set to the ANN).
These ANNs are also trained by using numerically gen-
erated input data from FEM of the intact and damaged
steel beams. After validation of neural networks, exper-
imentally obtained input data are introduced to the selec-
ted ANNs for severity and location prediction of the real
damage case.
The outline of the damage assessment method is sche-
matically shown in Fig. 1.
Fig. 1. Damage assessment method.
1787
3. Finite element modelling
3.1. Background
In model-dependent vibration-based analyses, it is
important to have an accurate numerical model before
performing experiments. In numerical analysis, the FEM
of intact and damaged steel beams are created by using
a commercial software, ANSYS 6.1 [39]. The structure
considered in the analysis is a 450 mm long, 40 mm
wide and 3 mm thick steel cantilever beam.
A four-node quadrilateral two-dimensional linear
structural shell element, SHELL63 [40], is selected to
model the beam. Since shell elements have six degrees
of freedom (three translational and three rotational) at
each node, cantilever boundary conditions can also be
modelled more realistically by constraining all degrees
of freedoms of the nodes located at the root of the beam.
Fig. 2 shows the finite element mesh of the steel beam
having 360 shell elements with 455 nodes; and the
location of the damage.
The first three bending natural frequencies of the
intact cantilever beam, for different mesh densities, are
also obtained in order to check the mesh independency.
The variation in natural frequencies is less than 1%
Fig. 2. FEM of cantilever steel beam.
1788 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Table 1 3.3. Normal mode dynamic analysis

First three natural frequencies of intact beam (Hz)
A normal mode dynamic analysis is performed to find
Mode no. Closed-form solution Shell elements
the first three undamped natural frequencies and their
4 × 90 16 × 360 corresponding displacement mode shapes for the cantil-
ever beam. Table 3 shows the first three natural fre-
1st bending 12.2893 12.3859 12.3818 quencies of the intact and the damaged beam. Ninety-
(0.7860) (0.7527) one points along the centre line of the beam are selected
2nd bending 77.0156 77.6044 77.5748 to obtain one-dimensional displacement mode shapes.
(0.7645) (0.7261) During the analysis, absolute differences in curvature
3rd bending 215.6461 217.5093 217.3731
(0.8640) (0.8008) mode shapes, functions of the second derivative of the
deflected shapes, are used as a damage index. Curvature
mode shapes [10] can be obtained from the normalised
displacement mode shapes with a central difference
between 4 × 90, i.e. 4 and 90 element divisions along
approximation by using the following equation:
the width and the length of the beam, respectively, and
16 × 360 mesh densities. The former, having 360 (vi+1⫺2vi ⫹ vi⫺1)
elements, is used for all later simulations due to the com- v⬙i ⫽ (1)
h2
putational efficiency. After performing modal analysis
for the first three bending natural frequencies, results Since the length of each element in the FEM is constant
obtained from different models are compared to the along the beam, and the absolute differences between
closed-form solution in Ref. [41] with results shown in the curvature mode shapes of the intact and the damaged
Table 1. The number in parenthesis shows the percent- structures are of primary interest, the h2 term appears as
age difference between theoretical and FEA results. a constant scaling factor and has been taken as unity.
Here, all the three bending modes refer to the vertical
bending of the beam. Horizontal bending, torsional and 3.4. Feature extraction
bending–torsion coupling modes are not considered in
the analysis. In the first part of this analysis, the effect of the sever-
ity and the location of the damage on the first three natu-
3.2. Damage scenarios ral frequencies of the structure are investigated. Fig. 3a–
h shows the percentage reduction in natural frequencies
Local damage is introduced with 32 damage severities with 32 different severities at eight different locations.
at 15 different spatial locations along the span of the It can be seen from Fig. 3a–h that different vibration
beam, i.e. 480 different damage scenarios. Structural modes are affected to different extents due to the
damage is modelled as a local reduction in thickness of reduction in thickness, depending on the location of the
the selected elements. The incremental reduction in damage. This illustrates the need to consider more than
thickness is chosen as 2.5% up to 80% local damage. one vibration mode for improved damage severity pre-
The damage is located at the fractions of the beam length diction. In the second part, the variation of the magnitude
from the cantilever end, shown in Table 2. The width of the absolute differences in curvature mode shapes
(10 mm) and the length (40 mm) of the damage are kept between the intact and the damaged beam is considered
constant during the analysis corresponding to two at the same locations in the first three modes of vibration.
elements in span wise and four elements in the width Figs. 4–6 show the variations in modes 1, 2 and 3,
direction in FEM. In Fig. 2, the damaged beam with a respectively. Each collocation point corresponds to
2 mm slot corresponding to 66.67% reduction in thick- location of the strain gauges used in the experiment and
ness at the location of 0.4556L is shown. they are numbered accordingly.
One aim of this study is to keep the vibration-based
Table 2
Different damage locations as a fraction of beam length Table 3
First three natural frequencies of intact and damaged beam (Hz)
0.044 0.44
0.10 0.50 Finite element analysis results
0.14 0.54
0.20 0.60 Mode no. Intact beam Damaged beam fdamaged/fintact
0.24 0.64
0.30 0.70 1st bending 12.3859 10.7415 0.8672
0.34 0.74 2nd bending 77.6044 60.3581 0.7778
0.40 3rd bending 217.5093 212.4555 0.9768
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802 1789

Fig. 3. Percentage reduction in natural frequencies for different damage locations: Damage located at (a) 0.24L (b) 0.3L (c) 0.34L (d) 0.44L
(e) 0.54L (f) 0.6L (g) 0.64L (h) 0.74L.

analysis inputs (i.e. the parameters introduced to ANNs) 4. Experimental study

as low as possible and yet obtain a high accuracy in the
location and severity predictions. Therefore, reduction in
natural frequencies, the maximum value of the absolute
difference in curvature and its corresponding location
along the beam in the first three flexural modes of
vibration are used as input to the ANN.
To validate the numerical method proposed for quanti-
fication and localisation of the damage, an experimental
study was carried out on steel beam specimens. Experi-
mental procedure, beam specimens and modal analysis
results are described in detail in the following sections.
1790 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Fig. 4. Variation of the magnitude of the absolute differences in curvature along the beam with different severities in Mode 1: Damage at
(a) 0.24L (b) 0.3L (c) 0.34L (d) 0.44L (e) 0.54L (f) 0.6L (g) 0.64L (h) 0.74L.

4.1. Test procedure and specimens
Two steel beams having same geometrical (450 ×
40 × 3 mm) and material properties
(E = 207.109 (Pa), r = 7850 (kg / m3), n = 0.33) are
used during the experimental modal analysis. One of the
beams is selected as a control specimen (intact structure)
and the other is used as a damaged beam. The local dam-
age is introduced to the steel beam as a 2 mm deep
(66.67% reduction in thickness) and 10 mm wide slot,
which is created by removing the material from the sur-
face of the specimen along the width (40 mm) at 205
mm (0.4556L) away from the cantilever end.
The dynamic characteristics of the structure are inves-
tigated by performing two different tests. The first used
a digital vibration controller to measure and record the
acceleration response of the structure under different
excitations, covering a certain range of frequencies,
which can then be used to obtain resonant frequencies
of the beam. In the second experiment, a signal-con-
ditioning amplifier is used to measure the dynamic
strains at eight different locations, from which the strain
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802 1791

Fig. 5. Variation of the magnitude of the absolute differences in curvature along the beam with different severities in Mode 2: Damage at
(a) 0.24L (b) 0.3L (c) 0.34L (d) 0.44L (e) 0.54L (f) 0.6L (g) 0.64L (h) 0.74L.

mode shapes of the structure can be obtained by using
surface-bonded electrical resistance strain gauges. Fig. 7
shows the schematic view of the experimental set-up for
vibration testing, used to determine frequency response
functions, resonant frequencies, dynamic strain histories
and strain mode shapes of the cantilever intact and dam-
aged beams.
4.2. Frequency measurement
A four-channel digital vibration controller with a data
acquisition unit is used to measure acceleration ampli-
tude histories from two channels including both input
(control excitation) and output (dynamic behaviour of
the specimen). The software is used to post-process the
data, plot the input and output frequency spectra and
1792 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Fig. 6. Variation of the magnitude of the absolute differences in curvature along the beam with different severities in Mode 3: Damage at
(a) 0.24L (b) 0.3L (c) 0.34L (d) 0.44L (e) 0.54L (f) 0.6L (g) 0.64L (h) 0.74L.

zoom to peak values for more accurate prediction of the
resonant frequencies. One of the two accelerometers is
used to control the input excitation amplitude supplied
to electro-dynamic vibration generator and the other is
used to monitor the output acceleration amplitude of the
specimen. The accelerometer monitoring the response of
the beam is attached on the centre line at the free end
of the beam. This location gives sufficient responses for
all three modes.
4.3. Strain measurement
A 10-channel Measurements Group Model 2000 sig-
nal-conditioning amplifier is used to amplify the signals
from the strain gauges, which are recorded using a
National Instruments VXI A/D acquisition system. The
excitation level is set to 2 V that is equivalent to
1000me and the bridge is balanced. In order to find cali-
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802
Fig. 7. Experimental set-up for vibration testing.
brated and measured strain values, the following equa-
tions [42] are used:
Rg·106
mecal ⫽ where
K·(Rc ⫹ (Rg / 2))
Rc ⫽ 59.94 ⍀, Rg ⫽ 120 ⍀, K ⫽ 2.11
Vloaded
memeasured ⫽ me where
(Vcal⫺Vunloaded) cal
Vunloaded ⫽ 0, Vcal ⫽ 2 V
In these equations, K, Rg, Rc, Vcal, Vloaded, ecal and emeasured
are the gauge factor of strain gauge, strain gauge resist-
ance, calibration resistor, calibration voltage, voltage
reading from loaded strain gauge, calibrated strain and
measured strain, respectively. Eight single element elec-
trical resistance gauges, with gauge length of 3 mm,
nominal resistance of 120⍀ ± 0.3 and a gauge factor of
2.11, are used to measure the direct strain on the surface
of the beams. The distributed surface-bonded strain
gauges, which are numbered from the free end of the
beam, can be seen in Fig. 8. Detailed information about
Fig. 8. Steel beam specimen with distributed strain gauges.
1793
the spatial location of the gauges along the beam is given
schematically in Fig. 9.
4.4. Excitation
(2)
Several experimental techniques are available to mea-
sure the dynamic characteristics of the structures. These
techniques generally excite the structure externally to
(3) acquire a dynamic response. During the experimental
analyses, two different types of excitation are applied
using an electro-dynamic vibration generator controlled
by a digital vibration control system.
4.4.1. Frequency sweep excitation
A swept sine excitation is used to identify the
response of a test piece to vibration across a range of
frequencies. In this way, resonant frequencies of the test
specimen are investigated. A uni-directional type swept
sine, with linear incremental sweep rate, excites the sys-
tem from 3 to 253 Hz, at increments of 0.5 Hz for the
experimental modal analysis used to find the first three
resonant frequencies of the intact and the damaged can-
tilever beams. The range is estimated from theoretical
Fig. 9. Strain gauge locations on damaged beam (not in scale, dimen-
sions in meter).
1794 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Table 5
ANNs used in the analyses

Input Output Architecture

(1) RNF DS 3:6:1

(2) MADC DL 3:6:1
(3) MADC&LOC DL 6:9:1
(4) MADC&LOC DS&DL 6:12:2
(5) RNF&MADC&LOC DS&DL 9:18:2

normalised resonant frequencies of the damaged beam

with respect to the intact beam are slightly lower than
the values obtained from FEA (Table 3). The possible
sources for this slight variation between numerical pre-
dictions and experimental results are material properties
(e.g. E and r), geometrical properties (e.g. variation in
length, width and thickness), boundary conditions,
measurement sensitivities (weight of the accelerometer
and the cables of the strain gauges), etc.

4.4.2. Constant frequency excitation

Fig. 10. Frequency response of cantilever beams. (a) Intact beam,
(b) damaged beam.
and finite element modal analysis results. The excitation
amplitude is set to 0.5g throughout the sweep and con-
trolled by an accelerometer mounted to the electro-
dynamic vibration generator. The output signal coming
from the test specimen is also monitored with another
accelerometer attached at the free end of the beam. Fig.
10a,b shows half of the peak-to-peak values of the accel-
eration amplitude of the intact and damaged specimens
with increasing frequency, respectively. The resonant
frequencies for intact and damaged beams are also tabu-
lated in Table 4. It can be seen from Table 4 that the
This type of excitation can be applied with constant
acceleration input operating at a single frequency. After
obtaining resonant frequencies of the test specimen by
using a swept sine, corresponding strain mode shapes
of intact and damaged beams are extracted by applying
constant frequency excitation at each resonant frequency
of interest. The sampling rate and sample length are of
800 Hz and 4000 points, respectively, corresponding to
5 s strain measurements. The output of eight strain
gauges in the first vibration mode of the intact and dam-
aged beams is shown in Figs. 11 and 12, respectively.
Normalisation is performed with respect to strain gauge
at location 8, which provides a reasonable finite amount
of strain for all three modes, in order to investigate the
relative change in curvature at eight different locations.
Figs. 13 and 14 show normalised strain mode shapes of
intact and damaged beams in the first three natural
modes, respectively. Since the additional features
extracted from vibration-based analysis are the absolute
differences in strain mode shapes between the intact and
damaged beams, Fig. 15 is obtained from Figs. 13 and
14.

Table 4
5. Application of ANN
First three resonant frequencies of intact and damaged beams (Hz)
5.1. Design of ANNs
Mode no. Experimental results
In order to find the most effective ANN that uses
Intact beam Damaged beam fdamaged/fintact vibration-based analysis data for the severity and
location predictions, different neural networks (Table 5)
1st bending 11.8 9.2 0.7796 with one hidden layer are designed by using MATLAB
2nd bending 73.1 50.5 0.6908
3rd bending 205.6 200.8 0.9766
neural network toolbox [43]. Different combinations of
input and output pairs are then introduced to these super-
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802 1795

Fig. 11. Strain output of eight strain gauges in the first mode (intact beam).

vised feed-forward backpropagation ANNs for training
and validation. Values (separated with semicolon) used
in the architecture column of Table 5 show the total
number of neurons in the input, hidden and output lay-
ers, respectively.
5.2. Training and validation
5.2.1. Noise-free case
In this analysis, 480 different damage scenarios are
generated by using 32 different reductions in thickness
at 15 different locations throughout the beam. Four hun-
1796 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Fig. 12. Strain output of eight strain gauges in the first mode (damaged beam).

dred and fifty input–output pairs are given to the ANN
for training and the rest of the input–output pairs are
used to check the generalisation of the learning during
the validation process.
5.2.2. Addition of artificial noise case
Artificial random noise (Eq. (4)) is added to normal-
ised frequency and maximum differences in curvature
mode shape data that were used in the noise-free case,
in order to achieve a better generalisation during the
training of ANNs and to simulate the experimental
uncertainties. A random number generator, randn in
MATLAB, generates random numbers with a normal
distribution having zero mean and a variance and stan-
dard deviation of one.
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802 1797

Fig. 13. Normalised strain mode shapes of intact beam. (a) Mode 1, Fig. 14. Normalised strain mode shapes of damaged beam.
(b) Mode 2, (c) Mode 3. (a) Mode 1, (b) Mode 2, (c) Mode 3.
1798 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Datawith noise ⫽ Datanoise-free· 1 ⫹冋 ⌬

100
randn(N) 册 (4)

Fifty copies of the normalised natural frequencies (when

RNF is an input) and 25 copies of normalised natural
frequencies, maximum absolute differences in curvature
mode shapes and the location corresponding to
maximum absolute difference in curvature mode shape
(when MADC&LOC and RNF&MADC&LOC are
inputs), are obtained and different artificial noise histor-
ies are added. In the case of RNF, 22 500 and 1500
entries are used for training and validation, respectively.
Similarly, 11 250 and 750 entries are used for training
and validation of MADC&LOC and RNF&MADC&
LOC. Table 6 shows the different percentages of noise
combinations added to normalised frequency and
maximum absolute differences in curvature mode
shape data.

5.3. Predictions of ANNs

In this section, the severity and location predictions

for the noise-free case (Figs. 16–20) are shown in detail
by considering the training, validation sets and the vari-
ation in mean square error with increasing number of
epochs. Here, the dotted lines lying on both sides of the
centre line indicate a 5% deviation from the target valu-
es.
The training performance of the first ANN (RNF-DS)
can be seen in Fig. 16a considering the trend in mean
square error values during training. Fig. 16b shows the
prediction of this ANN regarding the severity of the
damage. In the second and third ANNs, the location of
the damage is predicted by using two different sets of
input data, which are MADC (Fig. 17) and MADC&
LOC (Fig. 18), respectively. It can be seen from Figs.
17 and 18 that better generalisation (Fig. 18a) and pre-
diction (Fig. 18b) are achieved by providing the
additional location information. Since the input features,
the maximum absolute differences in strain mode shapes
and their locations, are good indicators for location pre-
diction of the damage, the next test run is performed by
using these features as an input for both severity (Fig.
19a) and location (Fig. 19b) predictions. In the final test

Table 6
Addition of different percentage of noise to input data

Input data Noise percentage (⌬)

Fig. 15. Absolute differences in strain mode shapes between the RNF MADC
intact and damaged beams. (a) Mode 1, (b) Mode 2, (c) Mode 3.
RNF 0.5, 1, 2 –
MADC&LOC – 1, 3, 5
RNF&MADC&LOC 0.5 1
RNF&MADC&LOC 1 3
RNF&MADC&LOC 2 5
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802 1799

Fig. 16. ANN results for severity prediction (input: RNF and output: Fig. 17. ANN results for location prediction (input: MADC and out-
DS). (a) Mean square error with number of epochs, (b) severity predic- put: DL). (a) Mean square error with number of epochs, (b) location
tion. prediction.

run, the severity (Fig. 20a) and location (Fig. 20b) of
the damage are predicted by introducing inputs, namely,
reduction in natural frequencies, maximum absolute dif-
ferences in curvature mode shape and their correspond-
ing locations along the beam from the first three natural
modes. The results for noise-free (Table 7) and for
additional artificial noise (Table 8) cases are tabulated
for performance comparison. It can be concluded from
Table 8 that more accurate results are obtained in
location predictions (MADC&LOC-DL) for the damage
than severity predictions (RNF-DS), although the noise
levels are comparably higher than the ones used on RNF
data. When the gradually increased noise combinations
applied to the normalised frequency and maximum
absolute differences in curvature mode shape are used
in the prediction of both severity and location of the

Table 7
Predictions of ANNs (noise-free case)

Input Output Target Prediction Difference

RNF DS 0.6667 0.6653 0.0014

MADC DL 0.4556 0.4866 0.0310
MADC&LOC DL 0.4556 0.4656 0.0100
MADC&LOC DS&DL 0.6667 0.4556 0.6291 0.4610 0.0376 0.0054
RNF&MADC&LOC DS&DL 0.6667 0.4556 0.7123 0.4375 0.0456 0.0181
1800 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802

Fig. 18. ANN results for location prediction (input: MADC&LOC,

output: DL). (a) Mean square error with number of epochs, Fig. 19. ANN results for severity and location prediction (input:
(b) location prediction. MADC&LOC). (a) Severity prediction (output: DS), (b) location pre-
diction (output: DL).

Table 8
Predictions of ANNs (additional artificial noise case)

Input (noise %) Output Target Prediction Difference

RNF (0.5) DS 0.6667 0.6541 0.0126

RNF (1) DS 0.6667 0.6908 0.0241
RNF (2) DS 0.6667 0.7157 0.0490

MADC&LOC (1) DL 0.4556 0.4613 0.0057

MADC&LOC (3) DL 0.4556 0.4471 0.0085
MADC&LOC (5) DL 0.4556 0.4142 0.0414

RNF&MADC&LOC (0.5 and 1) DS&DL 0.6667 0.4556 0.7162 0.4319 0.0495 0.0237
RNF&MADC&LOC (1 and 3) DS&DL 0.6667 0.4556 0.7836 0.4712 0.1169 0.0156
RNF&MADC&LOC (2 and 5) DS&DL 0.6667 0.4556 0.8057 0.4507 0.1390 0.0049
 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802
Fig. 20. ANN results for severity and location prediction (input:
RNF&MADC&LOC). (a) Severity prediction (output: DS),
(b) location prediction (output: DL).
damage, better estimations are achieved in location pre-
dictions compared to severity at each level of additional
artificial noise.
6. Conclusions
An experimentally validated damage detection algor-
ithm using features extracted from vibration-based
analysis data as input for ANNs, for location and sever-
ity prediction of damage in beam-like structures was
presented. Different damage scenarios have been created
by reducing the local thickness of the selected elements
at different locations, and simulated vibration responses
have been introduced to ANNs with and without arti-
ficial noise during training. Sensitivity analysis has also
been performed on extracted features by using different
vibration modes considering the effect of damage
location and severity before introducing them to ANNs.
1801
In this study, the test data (i.e. resonant frequencies and
strain mode shapes) has been obtained from a steel beam
with a local damage of 2 mm deep (66.67% reduction
in thickness) and 10 mm wide slot at 205 mm (0.4556L)
away from the cantilever end by using eight distributed
surface-bonded electrical strain gauges and an acceler-
ometer mounted at the tip. It can be concluded from the
ANN predictions that the reduction in natural frequency
provides necessary information for the existence and
severity of the damage. On the other hand, additional
location information for the maximum absolute differ-
ence in curvature mode shape served as better indicator
in the location predictions. The ANN results, regarding
severity and location of the damage, are not as promising
as the ones obtained form individual inputs in both
noise-free and with artificial noise cases. Considering the
robustness of the ANNs with increased noise level,
localisation of the damage is more accurate than the
quantification.
As a conclusion, the features extracted from vibration-
based analysis and used as an input for the ANNs, the
level of the noise on these features, the experimental pro-
cedure, measuring devices and their effectiveness at dif-
ferent modes on vibration, play an important role in the
severity and location prediction of the damage in beam-
like structures.
Acknowledgements
The financial assistance of EPSRC, contract number
GR/N3458101, is gratefully acknowledged since part of
the experiment has been carried out using dynamic test
equipment purchased from EPSRC grant.
References
[1] Salawu OS. Detection of structural damage through changes in
frequency: a review. Eng Struct 1997;19(9):718–23.
[2] Zou Y, Tong L, Steven GP. Vibration-based model-dependent
damage (delamination) identification and health monitoring for
composite structures—a review. J Sound Vib 2000;230(2):357–
78.
[3] Doebling SW, Farrar CR, Prime MB, Shevitz DW. Damage
identification and health monitoring of structural and mechanical
systems from changes in their vibration characteristics: a litera-
ture review. Technical Report No.: LA-13070-MS, 1996.
[4] Cawley P, Adams RD. The location of defects in structures from
measurements of natural frequencies. J Strain Anal
1979;14(2):49–57.
[5] Yuen MMF. A numerical study of the eigenparameters of a dam-
aged cantilever. J Sound Vib 1985;103(3):301–10.
[6] Inada T, Shimamura Y, Todoroki A, Kobayashi H, Nakamura
H. Damage identification method for smart composite cantilever
beams with piezoelectric materials. In: Proceeding of the 2nd
International Workshop on Structural Health Monitoring, Stan-
ford, 1999. Technomic Publishing Co; 2000. p. 986–94.
[7] Fox CHJ. The location of defects in structures: a comparison of
1802 M. Sahin, R.A. Shenoi / Engineering Structures 25 (2003) 1785–1802
the use of natural frequency and mode shape data. In: Proceeding
of the International Modal Analysis Conference. 1992. p. 522–8.
[8] Hu N, Wang X, Fukunaga H, Yao ZH, Zhang HX, Wu ZS. Dam-
age assessment of structures using modal test data. Int J Solids
Struct 2001;38(18):3111–26.
[9] Cornwell P, Doebling SW, Farrar CR. Application of the strain
energy damage detection method to plate-like structures. J Sound
Vib 1999;224(2):359–74.
[10] Pandey AK, Biswas M, Samman MM. Damage detection from
changes in curvature mode shapes. J Sound Vib
1991;145(2):321–32.
[11] Ratcliffe CP. Damage detection using a modified Laplacian oper-
ator on mode shape data. J Sound Vib 1997;204(3):505–17.
[12] Ho YK, Ewins DJ. Numerical evaluation of the damage index.
In: Proceeding of the 2nd International Workshop on Structural
Health Monitoring, Stanford, 1999. Technomic Publishing Co;
2000. p. 995–1011.
[13] Wahab MMA, Roeck DD. Damage detection in bridges using
modal curvatures: application to a real damage scenario. J Sound
Vib 1999;226(2):217–35.
[14] Lakshmi K, Jebaraj C. Sensitivity analysis of local/global modal
parameters for identification of a crack in a beam. J Sound Vib
1999;228(5):977–94.
[15] Pai PF, Young LG. Damage detection of beams using operational
deflection shapes. Int J Solids Struct 2001;38(18):3161–92.
[16] Pai PF, Jin S. Locating structural damage by detecting boundary
effects. J Sound Vib 2000;231(4):1079–110.
[17] Waldron K, Ghoshal A, Schulz MJ, Sundaresan MJ, Ferguson F,
Pai PF et al. Damage detection using finite element and laser
operational deflection shapes. Finite Elements Anal Des
2002;38(3):193–226.
[18] Beale R, Jackson T. Neural computing: an introduction. Bristol:
The Institute of Physics, 1990.
[19] Hassoun MH. Fundamentals of artificial neural networks. The
M.I.T. Press, 1995.
[20] Mehrotra K. Elements of artificial neural networks. The M.I.T
Press, 1997.
[21] Patterson DW. Artificial neural networks: theory and appli-
cations. Prentice Hall, 1996.
[22] Bishop CM. Neural networks for pattern recognition. Clarendon
Press, 1995.
[23] Worden K, Tomlinson GR. Damage location and quantification
using neural networks. In: Edwards JH, Kerr J, Stanley P, editors.
Engineering integrity assessment. Sheffield: Engineering Publi-
cations; 1994. p. 11–31.
[24] Worden K, Burrows AP. Optimal sensor placement for fault
detection. Eng Struct 2001;23(8):885–901.
[25] Worden K, Ball AD, Tomlinson GR. Fault location in structures
using neural networks. In: Proceeding of 75th Meeting of the
AGARD Structures and Materials Panel on Smart Structures for
Aircraft and Spacecraft. 1992. p. 1-9.
[26] Yun CB, Bahng EY. Substructural identification using neural net-
works. Comput Struct 2000;77(1):41–52.
[27] Mukherjee A, Ravindra V. A fuzzy-neuro health monitoring tool
transient vibration response. Adv Vib Eng 2002;1(1):100–16.
[28] Worden K, Ball AD. Modal analysis-based condition assessment
of structures using neural networks. In: Proceedings of 6th Inter-
national Congress on Condition Monitoring and Diagnostic
Engineering Management. 1994. p. 139–46.
[29] Barai SV, Pandey PC. Vibration signature analysis using artificial
neural networks. J Comput Civil Eng ASCE 1995;9(4):259–65.
[30] Elkordy MF, Chang KC, Lee GC. Neural networks trained by
analytically simulated damage states. J Comput Civil Eng ASCE
1993;7(2):130–45.
[31] Kudva JN, Munir N, Tan PW. Damage detection in smart struc-
tures using neural networks and finite-element analyses. Smart
Mater Struct 1992;1(2):108–12.
[32] Islam AS, Craig KC. Damage detection in composite structures
using piezoelectric materials. Smart Mater Struct
1994;3(3):318–28.
[33] Okafor AC, Chandrashekhara K, Jiang YP. Delamination predic-
tion in composite beams with built-in piezoelectric devices using
modal analysis and neural network. Smart Mater Struct
1996;5(3):338–47.
[34] Staszewski WJ, Worden K, Wardle R, Tomlinson GR. Fail-safe
sensor distributions for impact detection in composite material.
Smart Mater Struct 2000;9(3):298–303.
[35] Zapico JL, Worden K, Molina FJ. Vibration-based damage
assessment in steel frames using neural networks. Smart Mater
Struct 2001;10(3):553–9.
[36] Xu YG, Liu GR, Wu ZP, Huang XM. Adaptive multilayer per-
ceptron networks for detection of cracks in anisotropic laminated
plates. Int J Solids Struct 2001;38(32-33):5625–45.
[37] Sensburg O, Worden K, Tomlinson GR. Detection and location
of structural damage using dynamic test data. In: Proceedings of
the 19th ICAS Congress/AIAA Aircraft Systems Conference.
1994. p. 2834–44.
[38] Sahin M, Shenoi RA. Vibration-based damage identification in
beam-like composite laminates by using artificial neural net-
works. Proc Inst Mech Eng, Part C, J Mech Eng Sci
2003;217(6):661–76.
[39] ANSYS. User Manual. ANSYS, Inc, 2002.
[40] ANSYS Help System. Elements Manual. ANSYS, Inc, 2002.
[41] Blevins RD. Formulas for natural frequency and mode shape.
Robert E. Krieger Publishing Co., Inc, 1979.
[42] Cunningham PR, White RG. A new measurement technique for
the estimation of core shear strain in closed sandwich structures.
Composite Struct 2001;51(3):319–34.
[43] Demuth H, Beale M. Neural network toolbox, user’s guide, ver-
sion 4. The Math Works, Inc, 2000.