CHAPTER 2 LITERATURE REVIEW ON

DAMAGE DETECTION AND HEALTH MONITORING

C HAPTER 2

LITERATURE REVIEW ON DAMAGE

2.1 Introduction

Numerous investigations in the damage detection or health monitoring of

structures have been vigorously carried out in the past decay. Some representative

works, including theoretical developments, experimental validations, and practical

applications, will be briefly reviewed in this chapter. Since the health monitoring

approaches in this work are based on the modal parameters obtained from the

structural vibrations, most reviews are focused on the modal-based (i.e. based on

structural modal parameters) methods. Moreover, by the efforts of some researchers

during these years, neural- network-based methods have become a major branch of

studying structural health monitoring. Therefore, the health monitoring and damage

detection related studies which were based on neural networks will also be reviewed

herein.

2.2 Modal-Based Damage Detection And Health Monitoring

The modal-based methods utilize the information from modal parameters to

detection and assess structural damage. The majority of this group of methods uses the

15
CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

lower modal frequencies and can best describe the global behavior of the structure.

Therefore, they hold promise for global non-destructive inspection of a variety of

structures, because surface measurements of a vibrating structure can provide

information about the health of the internal members without costly dismantling of the

structure. Also, because of their global nature, these techniques allow the

customization of measurement points. Another major advantage is that the modal

information is easy to extract from the measurements obtained through free, ambient,

and forced vibration tests. A number of damage assessment techniques based on

changes in structural modal parameters have been proposed. The concept underlying

such an approach is that damage to a structure reduces its natural frequencies, increases

the modal damping, and changes the modal shapes.

Various damage identification algorithms have been developed for dealing with

three key problems, i.e., detection of the presence of damages, detection of the

structural damage locations, and estimation of the damage extents. The first one

problem is always involved in the second and third problems.

In early research, structural damage detection methods use natural frequencies as

damage indicator. Salawu [14] made a comprehensive review on the detection of

structural damage trough changes in frequency. However, from dynamic tests on

bridges, Alampalli and Fu [15] and Salawu and Williams [16] concluded that the

change in natural frequencies is not spatially specific and not sufficiently sensitive to

detect local damage in the structure so that its application is limited. The results of

their work indicate that the modal assurance criterion (MAC) and the coordinate modal

assurance criterion (CMAC), which are based on mode shape data, are useful in

detecting local structural change. Since mode shapes can provide much more

16
 CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

information than natural frequencies, many studies have concentrated their efforts on

damage detection with mode shapes information [17-19].

2.2.1 Methods for Detection of Structural Damage Location

For the damage localization problem, Cawley and Adams [20] proposed the first

model by employing the changes in the natural frequencies, together with a finite

element model (FEM), to locate the damage site of a given structure. Following their

works, some researches [21, 22] have found this method susceptible to measurement

errors, and ways of improving the localization have been introduced. Hearn and Testa

[23] have illustrated that the ratio of the elemental strain energy to the total kinetic

energy of the whole system is a fraction of the eigenvalue, and the ratio of this fraction

for two different modes is dependent only on the location of the damage. Shi et al. [24,

25] presented a method based on modal strain energy for locating damage in a

structure. Their method makes use of the change of modal strain energy in each

structural element before and after the occurrence of damage. Some properties of the

modal strain energy change are given to illustrate its sensitivity in locating damage.

By employing control-based eigenstructure assignment techniques, a subspace

rotation algorithm was proposed by Zimmerman and Kaouk [26], in which the damage

vector and relative rotation angle are used to identify the DOFs affected by damage.

Lim and Kashangaki [27] put forward a similar method in concept using

best-achievable eigenvectors, however, to identify the damaged structural members

directly.

Yao et al. [28] presented a structural diagnosis technique using vibratory signature

17
CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

analysis and the concept of strain mode shape. When a structure experiences a damage

or change, a new state of force equilibrium is realized. Since force distribution is, in

general, greatest near the damaged area, the location of damage is implicitly identified

by the severity of the strain mode shape change. Due to its sensitivity to local damage,

strain mode shape change seems to be a suitable damage indicator for locating

structural damage.

Stubbs and Kim [29] presented a methodology to localize and estimate the

severity of damage in structures for which only postdamage modal parameters are

available for a few vibratory modes. First, a theory of damage localization and severity

estimation that utilizes only changes in mode shapes of the structures is outlined. Next,

a system identification method that combines the experimental modal data and the

modal parameters of a finite element model of the struc ture is developed to yield

estimates of the baseline modal parameters for the structure. This method is attractive

when the baseline modal information for the structure is unavailable.

Topole and Stubbs [30] used natural frequencies with mode shapes and showed

the importance of introducing mode shape orthogonality to identify the location and

extent of damage on a structure. Messina et al. [31] developed an assurance criterion

for detecting single damage site of structures. And this method was extended to

identify the relative amount of damage at multiple sites [22, 32]. Recently, Shi et al.

[33] proposed a sensitivity-and statistical-based method to localize structural damage

by direct use of incomplete mode shapes. This method is an extension of the work by

Messina et al.[32]. The damage detection strategy is to localize the damage sites first

by using incomplete measured mode shapes, and then to detect the damage site and

extent again by using the more accurate measured natural frequency information.

18
 CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

Another important and interesting category uses the characteristics of the

flexibility matrix. Unlike the stiffness matrix, the flexibility matrix can be formed more

accurately through the usage of first several order experimental modal data. Lin [34]

used this flexibility matrix to multiply the pre-damaged FEM stiffness matrix to

determine the damage locations. Pandey and Biswas [35] used the change in flexibility

matrix before and after the occurrence of damage in the structure as damage index to

identify the location and amount of damage. An important advantage in this category is

that the usage of the analytical model can be avoided. Also, some researchers used

some special information such as curvature modes [36-39] and strain data [7] to search

for the damage locations.

2.2.2 Methods for Estimation of Damage Extent

For the estimation of damage extent, one important class of methods for

correlating measured modal data with analytical finite element models is the

minimization or elimination of modal force error. This error is that resulting from the

substitution of the analytical FEM and the measured modal data into the structural

eigenproblem. Various approaches have been presented to minimize some measure of

the error in the eigenproblem by perturbing the baseline values in the analytical model,

such as the components of the stiffness, damping, and mass matrices. One type of

method, known as sensitivity-based model update, uses the sensitivities of the modal

response parameters of the FEM to the structural design parameters (such as Young’s

modulus, density, etc.) to iteratively minimize the modal force error [40, 41]. Another

type of method, known as eigenstructure assignment, designs a controller that

minimizes the modal force error [26, 27]. Further, another type of method, known as

19
CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

optimal matrix update, solves a closed- form equation for the matrix perturbations that

minimize the modal force error or constrain the solution to satisfy it [42, 43].

Using modal parameters, Koh et al. [44] proposed an improved-condensation

method to estimate the stiffness matrix that corresponds to observed degrees of

freedom. Then, these authors detected local structural changes by quantifying changes

in stiffness. Based on their own previous work [25], Shi et al. [45] further proposed an

improved structural damage quantification algorithm. The algorithm includes the

analytical stiffness and mass matrices of the system in the damage quantification. It

reduces significantly the modal truncation error ad the FE modeling error from higher

analytical modes in the computation.

2.3 Neural-Networks-Based Approaches

Over the last two decades, artificial neural networks have gradually been

established as a powerful tool in the fields of prediction and estination, pattern

recognition, and optimization [46-50]. Due to the features of robustness, fault tolerance,

and powerful computing ability, the model of artificial neural networks becomes a

promising tool in solving civil engineering problems, such as linear/nonlinear system

identification, structural control and health monitoring.

2.3.1 Neural Networks for System Identification

There have been a lot of conventional identification methods for a linear dynamic

system. For a nonlinear dynamic system, in contrast, few effective methods are

20
 CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

available. A main difficulty in dealing with a nonlinear system lies in finding a reliable

mathematical model for it. Neural network, however, does not rely on a preconceived

mathematical model or even the number of parameters. The operation of neural

networks in system identification for either a linear system or a nonlinear system is the

same. Owing to its brilliant learning capability and its nonlinear nature, neural network

is well suited for nonlinear system identification and was studied and applied to

real-world problems [51]. Masri et al. [10, 52] has demonstrated in their studies that

neural networks are a powerful tool for the identification of systems typically

encountered in the structural dynamics field.

In the study of ANN-based system identification, Chen et al. [53] used ANNs to

identify the structural dynamic system by using the structural responses recorded in a

real apartment building during earthquake. The simulation results for a real multistory

building subjected to earthquake ground motions had sho wed that the structural

dynamic behavior can be well modeled by the trained neural networks. Based on the

results the authors indicate great promise in structural dynamic model identification by

using neural networks. The idea of this work was further applied to active control of

structures [54]. Chassiakos and Masri [55] also explored the potential of using neural

networks to identify the internal forces of typical systems encountered in the field of

earthquake engineering and structural dynamics. After fo rmulating the identification

task as a neural network learning procedure, the method is applied to a representative

chain- like system under deterministic and stochastic excitations. The neural network

based identification method provides very good results for general classes of

multi-degree-of- freedom structural systems.

Huang and Loh [56] proposed a network-based method for the modeling and

21
CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

identification of a discrete-time nonlinear hysteretic system during strong earthquake

motion. They use two-dimensio nal models of a three-story frame and a real bridge in

Taiwan subjected to several earthquake accelerograms to validate the feasibility and

reliability of the method for estimating the changes in structural response under

different earthquake events. It is shown that a multilayer neural network with nonlinear

transfer function is a general type of NARMAX model, and is suitable for the extreme

nonlinear input-output mapping problems.

Most ANN-based system identification models are non-parametric due to the

feature of ‘black box’ in ANN model. Huang et al., [13] developed a novel procedure

for identifying the dynamic characteristics of structures. The dynamic characteristics

are directly extracted from the weighting matrices of the neural network trained by

observed structural responses and input base excitations. This work has firstly given a

meaningful explanation on the black box form the view of system identification.

Other neural-based system identification related articles can be found in the

references [57-59].

2.3.2 Neural Networks for Structural Health Monitoring

Recently, some researches have investigated the suitability and capabilities of

ANNs for damage detection purposes. Ghaboussi et al. [59] and Wu et al. [60]

discussed use of neural networks for detection of structural damage in a three-story

frame with rigid floors. They trained neural networks to recognize the frequency

response characteristics of undamaged and damaged structures. The varying damage

levels were simulated by adjusting the properties of individual members. Elkordy et al.

22
 CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

[61] question the reliability of the traditional methods for structural damage diagnosis

and monitoring that rely primarily on the visual inspection and simple on-site tests.

They proposed a structural damage monitoring system for identifying the damage

associated with changes in structural signatures using neural networks. For training,

they used an FEM to develop failure patterns that were used to train a neural network

so that it can later diagnose damage in the reference structure. Szewezyk and Hajela

[62] presented a neural network approach based on mapping the static equilibrium

requirement for a structure in a finite element formulation with the assumption that

structural damage is reflected in terms of stiffness reduction. The results showed that

even with input noise and incomplete measured data, neural networks can still obtain

satisfying diagnosis.

Pandey and Barai [63, 64] trained a multilayer perceptron and a time-delay neural

network respectively for the detection of steel-truss bridge structures. Zhao et al. [65]

used a counter-propagation neural network to locate structural damage for a beam, a

frame, and support movements of a beam. The required data such as natural

frequencies, mode shapes and their other derivatives are obtained through the use of

FEM.

The studies presented by Masri et al. [11], Nakamura et al. [12], and Masri et al.

[66] used neural network-based approaches for the detection of changes in the

characteristics of structure- unknown systems. Their approaches rely on the use of

vibration measurements from a ‘healthy’ system to train a neural network for

identification purposes. Subsequently, the trained network is fed comparable vibration

measurements from the same structure under different episodes of response in order to

monitor the health of the structure. Differ from other approaches, the attractive

23
CHAPTER 2 LITERATURE REVIEW ON
DAMAGE DETECTION AND HEALTH MONITORING

advantages of these approaches are that they do not require the analytical model for a

real structure and only vibratory responses are used.

Zapico et al. [67] proposed a multi- layer-perceptron-based procedure for damage

assessment in a two-storey steel frame and steel-concrete composite floors structure. A

simplified finite element model is used to generate the training data. Sahin and Shenoi

[68] presented 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 ANNs for location and severity prediction of damage in beam- like structures.

A FE model is used to introduce the damage scenarios to generate the training data.

Other recent published works can be referred to the references [69-74].

All of the aforementioned investigations indicated that neural networks provide a

powerful tool for linear/nonlinear system identification and structural health

monitoring purposes.

24