Modal Based Structural Damage Detection using Artificial Neural Networks

ABSTRACT

The growing awareness on identification of damage in structures is gaining importance in the recent past due to
safety and economical aspects. Studies have been carried out by several researchers to detect and identify the
damage directly through the changes in measured vibration responses of the structure. The recent studies
showed that the neural networks and genetic algorithms are the two most promising methods in the area of
damage identification and localization. In the present paper, damage identification of structures has been carried
out using Artificial Neural Networks (ANN) and vibration data. The natural frequencies obtained from the finite
element analysis for the first six modes have been considered for generating the training data for ANN. It was
observed that the neural network is capable of predicting the location and magnitude of the damage with better
accuracy when fractional difference of frequencies of undamaged and damaged structures is used as the input.

INTRODUCTION

Structural identification is an objective tool for condition assessment of structures. By using

structural identification and parameter estimation as a means of determining the actual state
properties, performance, and limit states of a structure, it is possible to gain an improved
understanding of a structure's capacity and typical performance during its service. Thus, at any
point of time, it would be possible to assess the reliability of the structure using the objective
results obtained through structural identification. The main objective of health monitoring is to
detect, locate and assess the level of structural damage to the civil infrastructure.

The objective of the detection and identification of structural damage is to construct the qualitative
or quantitative description of the deterioration in a physical structural system from the measured
loads and the measured responses of the structure (Chou and Ghaboussi, 2001). One class of
damage location and assessment procedures which has attracted considerable interest recently is
that based on the alteration of the vibrational characteristics of structures when damage occurs.
These methods make use of the concept that the modal vibration test data i.e., the natural
frequencies and mode shapes can be considered to represent the state of the structure (Ananda Rao
et.al., 2004). Methods based on modal analysis have several advantages over alternative
techniques due to the fact that the modal parameters depend only on the mechanical characteristics
of the structure and not on the excitation applied, and furthermore that theoretically the structure
can be represented by measurements taken at a single location (Mares and Surace, 1996).

Various researchers have begun to experiment with artificial neural networks (ANNs) for damage
identification purposes during the last decade (Rytter and Kirkegaard, 1997) as an alternative to the
updating methods (Friswell and Mottershead, 1999). ANNs have some advantages that make them
very attractive: ability to treat damage mechanisms implicitly, capacity to generalise their
responses and robustness in the presence of noise. The strategy of these approaches is to train a
ANN to recognise different damage scenarios from the measured response of the system. In these
approaches, the selection of damage parameters, damage scenarios and the adjustment of the
numerical model to the physical system are prerequisites for success (Maia and Silva, 1997).

The present paper explores the possibility of application of artificial neural networks for
identification and localization of damage in structures. The studies carried out by various
researchers using the method for the detection of damage has been presented. Some case studies
have been carried out using neural networks for damage detection for a cantilever plate to
demonstrate the validity of the method. The demonstrative application of damage detection will
allow researchers in structural health monitoring to exercise the developed framework for real-life
applications such as bridges, and contribute to the toolkit of methodologies and algorithms.
DAMAGE DETECTION PROCESS

Every structure has its own natural vibration (or resonance) frequencies, which are related to the
materials, structural geometry, and integrity of the bridge. When the structure is under distress,
either in local or global, there would be reduction in stiffness and a decrease in the free energy
stored in the body (Salawu, 1997; Maeck and Roeck, 1999; Brownjohn, 2001). Since the dynamic
response is governed by system parameters viz., stiffness, mass, and damping; changes in these
parameters would lead to changes in the vibrational response as characterized by the modal
parameters viz. natural frequencies, mode shapes, and modal damping values. Due to the fact that
each vibration mode has a different energy distribution, any localized damage will affect each
mode differently depending on the location and severity of the damage. Modal parameters are also
sensitive to boundary conditions, i.e., physical constraints of the structure. The dynamic testing
allows the engineer to use the structural responses in conjunction with appropriate analysis and
modelling techniques to assess the condition of a structure.

Damage will soften the structure and thus modify its dynamic characteristics such as the frequency.
The relationship between the degree of damage and the change in frequency is fundamental in
order to diagnose any damage occurring in a structure. Thus, the presence of even a small damage
in a structure affects its dynamic behavior. Most of the quantitative global damage detection
methods that can be applied to complex structures examine the changes in the vibration
characteristics of a structure. Artificial neural network (ANN) has been used in the present study to
detect damage in the structure by the use of changes in natural frequencies between damaged and
undamaged structures. ANN was trained using the vibration data obtained by simulating the
different degrees of damage in finite element model. The trained neural network has been tested for
the unforeseen data for validation and further could be used for identification of damage based on
the frequency data obtained from the field tests.

DAMAGE DETECTION OF A CANTILEVER PLATE USING ANN

To demonstrate the capability of ANN in condition assessment/structural health monitoring,

detection of damage and its location in a cantilever plate has been considered. A cantilever steel
plate shown in Fig. 1 having a cross section of 50.75 mm x 6.0 mm and the mass density of 7670
kg/m3 has been considered. The first six natural frequencies for intact and damaged configurations
were obtained using finite element software. Damage was introduced by cuts at the clamped end
(Cut-1) and the mid-span (Cut-2) of the plate (Fig.1a). The depth of the two cuts are d=20mm. The
analysis procedure was repeated on simulated finite element model to obtain the first 6 natural
frequencies of the damaged plate. Fig.1b shows the cantilever plate with the damage locations.

To obtain the training data for the neural network, the cantilever plate was modeled in finite
element analysis program. The plate was divided into 9 divisions each of which represents one
damage location as shown in Fig.1b. The plate was equally subdivided into 36 elements, i.e., 4
elements for each division for accurate modeling of the resonant frequencies.

The Neural Network Architecture

A feed-forward back-propagation neural network has been used in the present study for the damage
detection in the cantilever plate. The architecture of the network consists of a vector of resonant
frequency information, and that the desired output would correspond to the damage magnitude and
the most likely location where that damage may occur. Hence, it is essential to decide on the format
of the input and output of the network, which in essence define the number of processing elements
(PEs) to be placed in the input and output layers. Once these layers were defined, the number of
hidden layers and the number of PEs for each of them were determined by the Root Mean Square
(RMS) error associated with the output layers as a decision quantity.

As the test data available for this particular problem consists of natural frequencies of the first 6
modes, it was decided to use 6 processing elements in the input layer representing a vector of the
first 6 natural frequencies of the beam. After going through the various literature studies, it was
decided to use 10 PEs for the output layer, in that, the first node consist of damage magnitude and
the remaining 9 nodes represent the damage indicators at each of 9 locations. The damage
indicators were considered as either ‘0’ or ‘1’, depending on whether that particular location is
undamaged or damaged respectively. For choosing the hidden layer, several options were tried with
different number of PEs in a single hidden layer. Finally, it was observed that 18 PEs in single
hidden layer network results in less RMS error. Fig. 2 shows the average RMS error versus number
of PEs in single hidden layer network. Hence, a three layer neural network of 6 nodes in input
layer, 18 nodes in hidden layer and 10 nodes in output layer is selected in the present study. The
final architecture of the ANN is shown in Fig. 3.

Training Data

To obtain the training data for the network, a finite element model has been generated for the
cantilever plate with 36 elements. The resonant frequencies in the undamaged state for the FE
model is matched with those of the test data results. This FE model has been used to generate the
training data containing a database of resonant frequencies as functions of the damage magnitude
and the location for the different damage levels. The damage was defined as the fractional loss of
the second moment of area over one location. The data was generated for a wide range of damage
magnitudes of 1%, 5%, 10%, 15%, 20%, 25% and 30%. Hence, the database consists of training
sets corresponding to the 7 damage states in each of 9 locations of the cantilever beam. This results
in 63 training sets.

In the multiple damage scenarios, viz., damage in two locations at a time, training data has been
generated for 1%, 10% and 20% damage magnitudes in each of two damage locations
simultaneously which results in 144 training sets.

Test/Validation Data

For validating the network, a test data set has been generated for 3 damage magnitudes of 3%, 13%
and 23%. This test database consist of test set for 3 damage levels for each of 9 damage locations
which resulted in 27 test sets in single damage case. For validating multiple damage state, a test
data set for the above three damage magnitudes in 3 sets of damage locations i.e., at (1,3), (1,5) and
(1,7) has been generated which resulted in 9 test sets.

Neural Network Training

The neural network selected as above is trained with the training data generated from the FE
model. The learning rate parameter () for the network is chosen as 0.01 and the error tolerance is
chosen as 0.01, after careful study from several trial runs. Several alternatives were tried with
different input formats to the network. Initially, the network was trained with the actual resonant
frequencies obtained from FE analysis. This resulted in less accurate prediction, especially in the
lower damage magnitudes.

After careful considerations, another training data set was generated in the form of fractional
changes of frequencies between the undamaged and damaged beam due to a given magnitude of
damage for a single damage as well as for multiple damage locations. The fractional change
information makes the database applicable to any cantilever structure with a uniform section. In
this scheme, the difference of resonant frequency values corresponding to a given damage situation
and the undamaged values were divided by the undamaged values, thus, resulting in a vector of six
values containing the fractional difference of the undamaged and damaged resonant frequencies as
shown below.
f ui  f di
zi 
f ui
where, zi is the fractional change in the i-th mode, and fu and fd are the undamaged and damaged
frequencies. Another neural network of 6-18-10 architecture was trained with this new training data
which resulted in less RMS error of 0.15% compared to 0.29% of first data set. Hence, it was
decided to use the fractional difference of frequencies as the training as well as test data set.

Damage detection using trained ANN

The trained neural networks were used to evaluate the performance of the networks in predicting
the location and magnitude of the damage. The neural networks were evaluated using several sets
of data that were analytically generated. For the evaluation process, three magnitudes of damage
that were different from the magnitudes used in the training process were selected such that three
deterministic data sets were generated. Each deterministic set contained 9 cases, one damage case
at each of nine locations. The damage magnitudes for the evaluation phase were 3%, 13% and
23%. These levels were artificially introduced into the finite element model, in each of 9 locations,
to generate a total of 27 deterministic cases. For each case, an input vector of fractional changes of
the first six resonant frequencies was also generated. Also, data sets were generated for the multiple
damage cases for three damage magnitudes in three locations. This test data were utilized to
evaluate the above two networks in terms of their ability: (a) to predict the magnitude of the
damage, and (b) to detect the correct location.

The damage identification has been carried out for different damage states for the test data set using
the ANN trained with actual frequency values. It was observed that, the network is able to predict
reasonably for higher damage states but is less efficient in predicting the location in lower damage
states. This type of nature is observed for all the locations.

To overcome this, another neural network having the same architecture as of the first test case, but
trained using the fractional changes of the frequencies in undamaged and damaged case is used for
the evaluation. The test data same as used above was used for this case also, by converting them
into fractional difference of frequencies. The results for this case which were obtained from FE
analysis and from ANN predicted values are shown in Figs.4(a) to 4(d). In the figures, the
normalized values of the frequencies obtained from the finite element analysis results are shown in
hatched bars and those of ANN predicted values are shown in shaded bars for three levels of
damage magnitudes i.e., 3%, 13% and 23% for all nine locations. From the figures, it can be
observed that the normalized values of the frequencies which were predicted by ANN are in close
proximity to the values obtained from FE analysis.

Multiple Damage Detection

For detecting damage in multiple damage scenarios, the neural network was trained with the data
generated by FE analysis by artificially introducing the damage into finite element model at two
locations simultaneously. For this case also, the fractional difference of frequencies have been used
for generating the training data. This trained neural network has been evaluated by test case data
generated for three levels of damage magnitudes. The multiple damage locations were chosen as at
(1,3), (1,5) and (1,7). The ANN prediction results have been shown against the target values in
Figs. 4(e) to 4(g). The results indicate that the ANN is efficient in predicting the damage locations
and also the magnitude of the damage.
SUMMARY AND CONCLUSIONS

The detection and identification of structural damage is a vital part of the monitoring and servicing
of structural systems during their lifetime. The effects of common damages on a structure are
changes in natural frequencies, mode shapes and structural damping. Since the measurement of
natural frequencies is easier than that of changes in structural damping, damage can be detected
from dynamic analysis using natural frequencies. Recently, artificial intelligence techniques such
as neural networks (NN) and genetic algorithms (GA) are being used in structural damage detection
and identification because of their high adaptivity, effectiveness and robustness in handling
uncertainty, insufficient information, and noise. In the present work, the applicability of ANN in
identifying the location and evaluation of magnitude of the damage has been demonstrated
considering a cantilever plate. Single and multiple damage location scenarios have been
considered. The natural frequencies of the first six modes obtained from the finite element analysis
have been considered for the training of ANN. It was observed that prediction capability of ANN is
better by using fractional difference of frequencies instead of using the actual frequencies. The
results show that the ANN method is capable of predicting the location and magnitude of damage
with good accuracy for single and multiple damage scenarios.

ACKNOWLEDGEMENTS

This paper is being published with the kind permission of the Director, Structural Engineering
Research Centre, Chennai.

REFERENCES

1. Chou, J.H., Ghaboussi, J., (2001), “Genetic algorithm in structural damage detection”,
Computers and Structures, 79, pp.1335-1353.
2. Ananda Rao, M., Srinivas, J., Murthy, B.S.N., (2004), “Damage detection in vibrating bodies
using genetic algorithms” Computers and Structures, 82, pp. 963–968.
3. Mares, C., and Surace, C., (1996), “An application of genetic algorithms to identify damage in
elastic structures”, Journal of Sound and Vibration, 195(2), pp. 195-215.
4. Rytter, A. and Kirkegaard, P. H., (1997), “Vibration based inspection using neural networks”,
Proceedings of DAMAS’97, University of Sheffield, UK., pp. 97–108.
5. Friswell, M. I. and Mottershead, J. E., (1999), “Finite Element Updating in Structural
Dynamics”, Dordrecht, The Netherlands: Kluwer Academic Publishers,.
6. Maia, N. M. M. and Silva, J. M. M., (1997), “Theoretical and Experimental Modal Analysis”,
Tauton, U.K.: Research Studies Press Ltd..
7. Salawu, O.S., Detection of structural damage through changes in frequency: a review,
Engineering Structures, Vol. 19, No. 9, 1997, pp. 918-723.
8. Maeck J., and De Roeck, G., Damage detection on a prestressed concrete bridge and RC
beams using dynamic system identification, Proc. of DAMAS '99, Dublin, Ireland, June 1999,
pp. 320-327.
9. Brownjohn, J.M.W., Hong, H. and Tsochien, P., (2001), “Assessment of structural condition
of bridges by dynamic measurements”, Applied Research Report RG 5/97, Nanyang
Technological University, Singapore.
(a) Damage introduced by cuts at Locations 1 and 5 for the cantilever
plate

6
900 mm 50.75
Cantilever plate

1 2 3 4 5 6 7 8 9

9 x 100 = 900 mm
Cantilever plate with damage locations

(b) Finite element model for simulation (9x4 = 36 elements)

Fig.1 Cantilever plate for simulation

 1.6
1.4

Ave. RMS error %

1.2
1
0.8
0.6
0.4
0.2
0
4 6 8 10 12 14 16 18 20 22
No. of Processing Elements
Fig. 2 Variation of RMS error with Processing Elements

18 Nodes
(1 Hidden layer)

10 Output
Patterns
O1
6 Input
Patterns O2

I1 O3

I2 O4

I3 O5

I4 O6

I5 O7

I6 O8
O9
O10

Fig. 3 ANN Architecture selected for damage detection of a cantilever plate (6-18-10
network)
 1.00
1.00
LOCATION 3 - Testing Patterns
LOCATION - 1 testing patterns 0.90 3% damage - Target
3% damage- Target 13% damage - ANN
0.90
3% damage- ANN 13% damage - Target
0.80
13% damage- Target 13% damage - ANN
0.80
13% damage- ANN 23% damage - Target
23% damage- Target 0.70
0.70
23% damage - ANN
23% damage- ANN

Normalized values
0.60
Normalized values

0.60

0.50
0.50

0.40
0.40

0.30 0.30

0.20 0.20

0.10 0.10

0.00 0.00
Damage
1 21 23 34 4 5 5 6 6 7 7 8 8 9 9 10 Damage
1 21 23 34 45 56 67 7 8 8 9 9 10
Magnitude |__________________________________________Damage Location__________________________________________________| |__________________________________________Damage Location__________________________________________________|
Magnitude

(a) (b)
1.00 1.00
LOCATION 5 - Testing Patterns LOCATION 6 - Testing Patterns
3% damage - Target 3% damage - Target
0.90 0.90
3% damage - ANN 3% damage - ANN
13% damage - Target 13% damage - Target
0.80 0.80 13% damage - ANN
13% damage - ANN
23% damage - Target 23% damage - Target
0.70 23% damage - ANN 0.70 23% damage - ANN

Normalized values
Normalized values

0.60 0.60

0.50 0.50

0.40 0.40

0.30 0.30

0.20
0.20

0.10
0.10

0.00
0.00 Damage 1 2 2 3 3 4 4 5 6 6 7 7 8 9
Damage 12 23 34 45 56 67 7 8 8 9 10 1 5 8 9 10
1 9 Magnitude |__________________________________________Damage Location__________________________________________________|
|__________________________________________Damage Location__________________________________________________|

(c) (d)
Magnitude

1.2000 1.2000
Output Patterns for damage at locations 1 & 3 Output Patterns for damage at locations 1 & 5
3% Damage - Target 3% Damage - Target
3% Damage - ANN 3% Damage - ANN
1.0000 1.0000 13% Damage - Target
13% Damage - Target
13% Damage - ANN 13% Damage - ANN
23% Damage - Target 23% Damage - Target
23% Damage - ANN 23% Damage - ANN
0.8000
0.8000
Normalized values

Normalized values

0.6000
0.6000

0.4000
0.4000

0.2000
0.2000

0.0000
Damage
1 12 23 34 45 5 6 6 7 7 8 8 9 9 10 0.0000
Magnitude |_______________________________________Damage Location_______________________________________________| Damage 12 23 34 45 5 6 6 7 7 8 8 9 9 10
1
Magnitude |_______________________________________Damage Location_______________________________________________|

(e) (f)
1.2000
Output Patterns for damage at locations 1 & 7 3% Damage - Target
3% Damage - ANN
13% Damage - Target
1.0000 13% Damage - ANN
23% Damage - Target
Fig.43 Comparison of Actual and ANN
23% Damage - ANN
predicted damage magnitude at different
0.8000
locations (ANN trained by fractional
Normalized values

0.6000
difference of undamaged and damaged
frequencies)
0.4000

(a,b,c,d) single damage cases

0.2000
(e,f,g) multiple damage cases
0.0000
Damage
1 12 23 34 45 5 6 6 7 7 8 8 9 9 10
Magnitude |_______________________________________Damage Location_______________________________________________|

(g)