STRUCTURAL DAMAGE IDENTIFICATION

USING HIGH DIMENSIONAL MODEL

REPRESENTATION

Thesis
Submitted in partial fulfilment of the requirements for the degree of
DOCTOR OF PHILOSOPHY

by

NAVEEN B.O.

DEPARTMENT OF CIVIL ENGINEERING

NATIONAL INSTITUTE OF TECHNOLOGY KARNATAKA
SURATHKAL, MANGALORE – 575 025

March, 2018
 DECLARATION
by the Ph.D. Research Scholar

I hereby declare that the Research Thesis entitled “Structural Damage Identification
Using High Dimensional Model Representation” which is being submitted to the
National Institute of Technology Karnataka, Surathkal in partial fulfilment of the
requirements for the award of the Degree of Doctor of Philosophy in Department of
Civil Engineering, is a bonafide report of the research work carried out by me. The
material contained in this Research Thesis has not been submitted to any University or
Institution for the award of any degree.

(NAVEEN B.O.)
Register No. 145060CV14F04
Department of Civil Engineering
Place: NITK, Surathkal
Date: 02-03-2018
 CERTIFICATE

This is to certify that the Research Thesis entitled “Structural Damage Identification
Using High Dimensional Model Representation” submitted by Mr. NAVEEN B.O.
(Register Number: 145060CV14F04) as the record of the research work carried out by
him, is accepted as the Research Thesis submission in partial fulfilment of the
requirements for the award of degree of Doctor of Philosophy.

Dr. A.S. Balu

Research Guide
(Signature with date and seal)

Prof. Varghese George

Chairman - DRPC
(Signature with date and seal)
 Dedicated to

My Family and Teachers

 ACKNOWLEDGEMENT

I would like to express my sincere gratitude to my research supervisor Dr. A.S. Balu,
Department of Civil Engineering, National Institute of Technology Karnataka,
Surathkal for providing continuous support and encouragement. I wish to acknowledge
the precious time spent with my supervisor for technical discussions, and writings
undertaken, which gave me confidence in perusing the research work.
I would like to express my sincere thanks to Head of the Department Dr. Varghese
George, former Heads Prof. D. V. Reddy and Prof. Katta Venkataramana and all faculty
members of the Department of Civil Engineering for their constant encouragement and
for providing necessary facilities during the period of my research work.
I am thankful to the members of RPAC, Prof. K. Swaminathan, Department of
Civil Engineering and Dr. Satyanarayana Engu, Department of Mathematical and
Computational Sciences for their continuous evaluation of my research work and for
their suggestions, and constant encouragement.
I would like to express my sincere thanks to all the research scholars of civil
engineering department especially Basavana Gowda S.N., Punith B. Kotagi, Darshan
C Sekhar, Renuka Prasad M. S., Parameshwar, Kesava Rao B., Spoorthi S. K., Sharan
Kumar, and Vinod N. Tamburi, for their constant encouragement throughout my
research work. Also, I extend my special thanks to my senior colleagues Naveen Kumar
D.T., Sangeetha D.M., and K. Rajendra Prabhu for their suggestions.
My special thanks to my family members especially my parents, father in-law,
mother in-law, my wife Mrs. Deeksha S.U., daughter Vaishnavi, brother Mr. Gururaja,
Mrs. Sridevi Gururaja, Mr. Dhyan, Mr. Atharv, Ms. Aadhya, Mr. Suresh Babu, Mrs.
Pushpa Suresh Babu and my friends Abhilash T.S., Prashanth S. K., Sridhar B.S., Vinay
and Shruthi Badami, who always supported and encouraged me.

Naveen B.O.
 ABSTRACT

Any engineering structure is subject to various internal and external factors which
may cause wear or malfunction due to deterioration, an incorrect construction process,
lack of quality control or environmental effects. To be able to observe these changes
in the material and to react in a proper way before serious damage is caused, the
implementation of a damage identification system is crucial. In the past, many
methods have attempted to identify damage by solving an inverse problem, which
inevitably needs an analytical model. However, often the construction of these
analytical model requires considerable effort in building a mathematical framework
with acceptable level of accuracy and reliability which makes these approaches less
attractive. In order to circumvent this complexity, this work presents a
computationally efficient approach in structural damage identification (SDI) using
high dimensional model representation (HDMR).
In general, most of the structural systems are simulated with the help of finite
element (FE) models to predict static as well as dynamic behaviour of the systems
with different boundary conditions. Therefore the FE models have to be in tune with
the experimental observation to facilitate any modifications in the systems so that the
future responses can be accurately predicted, and subsequently utilized in design
optimization. Thus, finite element model updating (FEMU) is effective in improving
the correlation between predicted and observed ones by correcting the inaccurate
modelling assumptions.
The proposed methodology involves an integrated finite element modeling,
development of response surface model using HDMR, establishment of objective
function, and minimization of the function using genetic algorithm. An attempt has
been made to reduce the computational effort with increase in the accuracy of updated
parameters.
The proposed methodology is applied in model updating of a simulated beam
and an existing reinforce cement concrete (RCC) box culvert structure. The results
have demonstrated that the HDMR based FEMU is a good candidate featuring
computational efficiency. Further to validate the proposed methodology in SDI, three
case-studies (an experimental beam, a frame structure and a bridge structure) have
been considered. The damage patterns, locations and severity obtained using the
proposed methodology are compared with the experimental results available in
literature, and are found to be in good agreement. Based on the study conducted, it
can be concluded that the HDMR based FEMU in SDI is computationally efficient.

Keywords: Finite element analysis; High dimensional model representation; Model

update; Response surface method; Structural damage identification.
 CONTENTS

Page No.
TABLE OF CONTENTS i
LIST OF FIGURES iii
LIST OF TABLES v
LIST OF ABBREVIATIONS AND SYMBOLS vii
1. INTRODUCTION 1
1.1 Structural Health Monitoring 2
1.2 Model Update 4
1.3 Structural Damage Identification 5
1.4 Need for the Present Work 8
1.5 Thesis Organisation 9
2. LITERATURE REVIEW 11
2.1 Finite Element Model Updating 11
2.2 Structural Damage Identification 19
2.3 High Dimensional Model Representation 22
2.4 Summary of Literature Review 27
2.5 Objectives of Research Work 28
3. MODEL UPDATING USING HDMR 29
3.1 HDMR 29
3.2 HDMR Expansions 30
3.3 Model Updating using HDMR 34
3.4 Numerical Examples 35
3.4.1 Example 1: Simply Supported Beam 35
3.4.1.1 Response Surface Generation 36
3.4.1.2 Optimization using GA 38
3.4.2 Example 2: Reinforced Concrete Box Culvert 43
3.5 Summary 47

i
4. STRUCTURAL DAMAGE IDENTIFICATION USING HDMR 49
4.1 Simulation Study : Simply Supported Beam 49
4.2 Case Study 1: Experimental Beam 54
4.3 Case Study 2: Reinforced Concrete Frame 66
4.4 Case Study 3: Bridge Structure 73
4.5 Summary 78
5. CONCLUSIONS 81
5.1 Summary and Research Findings 81
5.2 Suggestions for Future Work 83

APPENDIX 85
REFERENCES 91
PUBLICATIONS 100

ii
 LIST OF FIGURES

Page No.
1.1 Principle and organization of a SHM system 2
1.2 Applications of SHM 3
2.1 Summary of articles 12
3.1 Sampling scheme: First-order HDMR 33
3.2 Sampling scheme: Second-order HDMR 34
3.3 Flow chart of HDMR based model updating and 34
damage identification
3.4 Simply supported concrete beam 35
3.5 Parameter screening results 41
3.6 Finite element model of RC box culvert 44
4.1 A simulated simply supported beam 49
4.2 Difference in frequencies after model updating (n=3) 51
4.3 Identified damage locations and stiffness distribution (n=3) 52
4.4 Difference in frequencies after model updating (n=5) 52
4.5 Identified damage locations and stiffness distribution (n=5) 53
4.6 Beam cross section and with two point loading 54
4.7 Crack pattern at each load steps 55
4.8 Stiffness distribution: Reference state model (After updating: n=3) 56
4.9 Stiffness distribution: Reference state model (After updating: n=5) 58
4.10 Stiffness distribution: Ref.-damage state model
(After updating: n=5) 60
4.11 Stiffness distribution: Damage state model (After updating: n=5) 62
4.12 Variation of Young’s modulus at different updating states 64
4.13 3D Numerical model of experimental beam 65
4.14 Reinforcement bars (Linear line elements of type T3D2) 65
4.15 Comparison of cracks 66
4.16 RC frame: Geometric dimensions and accelerometer arrangement 66
4.17 Modal test of an RC frame 67

iii
4.18 Parameter screening results (RC frame) 67
4.19 Comparison of absolute errors (%) of initial and
reference state model (n=5) 70
4.20 Comparison of absolute errors (%) of reference and
damage state model (n=5) 72
4.21 Variation of E at different updating states 72
4.22 General view of I – 40 Bridge 73
4.23 Damage scenarios: First and Second stage damage 73
4.24 Damage scenarios: Third and Fourth stage damage 74
4.25 Simplified I-40 bridge model with accelerator layout 75
4.26 Two dimensional beam model of bridge 75
4.27 Identified damage location at substructure S5 (n=3) 76
4.28 Identified damage location at substructure S5 (n=5) 77
4.29 Stiffness distribution at S5 substructure (n=5) 77

iv
 LIST OF TABLES

Page No.
3.1 Identified values of parameter using GA 40
3.2 Responses for development of HDMR expression (Beam example) 42
3.3 Variation of values (Error percentage) for Beam 43
3.4 Computational effort (Beam Example) 43
3.5 Variation of values (Error percentage) for Box-culvert 46
4.1 Initial frequencies of FE model: Simulated beam 51
4.2 Updated frequencies of simulated beam (n=5) 53
4.3 Six step static load magnitude 55
4.4 Frequencies and their differences: Experimental beam  n  5  59

4.5 Frequencies and its difference of Ref.-damage state model:

Experimental beam  n  5  60

4.6 Frequencies and its difference of damage state model:

Experimental beam  n  5  61

4.7 Differences in frequencies of Modal flexibility residual and

HDMR before updating: Experimental beam (Reference state) 63
4.8 Differences in frequencies of Modal flexibility residual and
HDMR before updating: Experimental beam (Damage state) 63
4.9 Updated values of reference state model
Experimental RC frame 68
4.10 Modal frequencies of reference state model:
Experimental RC frame  n  5  69

4.11 Updated values of E1 substructure in damage state:

Experimental RC frame  n  5  70

4.12 Values of E at different updating states:

Experimental RC frame  n  5  70

4.13 Modal frequencies of damage state model:

v
 Experimental RC frame  n  5  71

4.14 Frequency errors before updating: Bridge example 76

4.15 Frequency errors after updating:
Bridge model in damage state  n  5  78

vi
 LIST OF ABBREVIATIONS AND SYMBOLS

Abbreviations
ABC : Artificial Bee Colony
ANN : Artificial Neural Network
ANNOVA : Analysis of Variance
C3D8R : Continuum 3D 8-Node Reduced Integration
CCD : Central Composite Design
DOE : Design of Experiment
FE : Finite Element
FEA : Finite Element Analysis
FEM : Finite Element Method
FEMU : Finite Element Model Updating
FFT : Fast Fourier Transform
FHDMR : Factorized High Dimensional Model Representation
FRF : Frequency Response Function
FRP : Fibre Reinforced Polymer
GA : Genetic Algorithms
GHDMR : General High Dimensional Model Representation
GRSMU : Generalized Response Surface Model Updating
HDMR : High Dimensional Model Representation
HHDMR : Hybrid High Dimensional Model Representation
IESM : Inverse Eigen Sensitivity Method
ILs : Influence Lines
MLS : Moving Least Square
NDE : Non-Destructive Evaluation
NN : Neural Network
PSO : Particle Swarm Optimization
PZT : Piezoelectric ceramic material
RCC : Reinforced Cement Concrete
RFM : Response Function Method

vii
RS : Response Surface
RS-HDMR : Random Sampling High Dimensional Model Representation
RSM : Response Surface Method
SA : Simulated Annealing
SDI : Structural Damage Identification
SHM : Structural Health Monitoring
T3D2 : 2-noded linear 3-D Stress/Displacement Truss Element

Symbols
A : Area of cross section of girder
 : Density of concrete
Y : HDMR response equation
x : Input parameters/variables
 : Lagrange interpolation function

i : Mean

N : Number of parameters/variables
Fobj : Objective function

 : Poisson’s ratio
c : Reference point
n : Sample points
I : Second moment of area
i : Standard deviation
E : Substructure
E : Young’s modulus

viii
 CHAPTER 1

INTRODUCTION

In order to improve the safety and serviceability of any civil engineering structures,
the present condition of the structures should be known. A periodical inspection
facilitates to understand the condition of the structures which involves structural
modifications, repairs and/or reconstructions. In economical point of view,
maintenance and repair cost will be less than the reconstruction of major structural
components which are distressed/damaged or the entire structure. The most
commonly adopted method to detect damages is carried out by visual inspection using
non-destructive methods. The process of implementing a damage detection strategy
for aerospace, civil and mechanical engineering structures is referred to as structural
health monitoring (SHM). Monitoring and early damage detection in a structure
depends on the ability of an SHM technique implemented into the engineering
system, where it aims to give a diagnosis of the state of the constituent materials of
different parts, and of the full assembly of these parts constituting the structure as a
whole at every moment during the life of a structure. Damage is defined as the
changes in the physical properties introduced into a system that adversely affect
current or future performance of the system, which occurs due to natural or man-made
cause. SHM is considered as an efficient approach of non-destructive evaluation
(NDE), which involves the utilization of smart materials, sophisticated sensors, data
transmission mechanisms, and usage of advanced processing techniques having more
computational power for monitoring a structure of interest. The implementation of
SHM results in reduction of inspection costs, possibility to better understand the
behaviour of structures under dynamic loads, seismic protection, observation in real
or near real-time of the structural response and evolution of damage, so that it is
possible to produce post-disaster scenarios and support rescue operations. Thus, SHM

1
is a multidisciplinary field, where a number of different skills and institutions can
work together in order to increase the performance and reliability of structural
systems. Figure 1.1 presents the organisation of a typical a SHM system in detail
(Balageas et. al. 2006).

Monitored physical
phenomenon, depending
on the damage



Integrity monitoring system, 
 defined by a sensed physical 
Diagnosis
phenomenon and an adapted  
Structural Health
Monitoring
data reduction 
Sensors 
Usage
Sensors
Damage multiplexing
and networking
fusion

Monitoring


Monitored 
Prognosis
structure Monitoring of usage conditions Health and Usage

Monitoring

Data cumulative Maintenance organization
recording 
 Health Management of the

structure
Damage and 
behaviour laws Health Management
of the full system (fleet, plant...)


Fig. 1.1 Principle and organization of a SHM system

1.1 STRUCTURAL HEALTH MONITORING

The motivation of SHM in various applications of engineering field involves

maximum utilization of structure over a period of time with better serviceability and
improved maintenance without any major damage of structural components, or failure
of the entire structural systems, thus improving safety and reliability. The various
applications of SHM is shown in Fig.1.2, and a brief application of SHM is discussed
below (https://www.hbm.com/en/5530/structural-health-monitoring).

 Civil engineering structures (ie, bridges, buildings, tunnels, etc.) are

designed to resist enormous amount of applied loads and forces due to
natural disasters. With the application of SHM the structure is monitored

2
 by material testing and load assessment, measurement of displacements,
deflections and rotations for extending the performance of the structure.
 Application of SHM in railway industry includes, fatigue analysis of
structural members and railway components (ie, wheels and axels), remote
monitoring of tracks, measurement of forces and mechanical stresses, and
data determination for life cycle cost calculations.
 In wind energy sector, SHM is useful in condition monitoring of the
critical components (ie, rotor blades, drivetrains, inverters) which are
subjected to extreme mechanical stresses.
 In oil and gas industry, SHM has important role in pipe line monitoring
using fibre optic strain sensors, and also in efficient measurement of drive
power using torque transduces in gas compressor stations.

SHM

Civil Railway Wind Energy Oil and Gas

Engineering

Monitoring of
Rail Fatigue testing
Bridges Wind Turbines
measurements of pipelines
components

Pipeline
Structural Measurement of
Tunnels monitoring using
durability testing forces and torque
fibre optics

Increase the
Buildings efficiency of
wind generators

Fig. 1.2 Applications of SHM

The design of SHM system is influenced by different parts as follows (Hejll

2007).
 The structural phenomena to be studied (cracking, settlements, etc.)
 The time strategy (continuous, periodic or triggered monitoring)
 The condition of the phenomenon to be studied (global or local)
 The load effect (position/intensity/nature of loading)

3
  The evaluation method
 The model used to evaluate the cause-effect (i.e., knowledge of
geometrical, material, load data, etc.)

The SHM and the damage identification are gaining larger importance in civil
engineering. It is also defined as the use of in-situ, non-destructive sensing and
analysis of structural characteristics in order to identify any damages. In addition, it
defines the location, estimates the severity, and evaluates the consequences on the
residual life of the structure. In the field of structural engineering and design, with
respect to aging of steel and reinforced concrete (RC) structures, safety has become
the most important criterion. Prolonging the life of the structure is the major role of
SHM techniques by detecting the damages in the initial stages only.

The structural dynamic behaviour can be represented by many terms such as

natural frequencies, Eigen values, damping ratio, and frequency response functions.
From the dynamic analysis we can understand and evaluate the responses, and these
can be modified as per the requirements of engineering design, if required. The
dynamic analysis of structures can be done through either experimental route or by
using theoretical approach. In most of the cases, performing experimental
investigation requires more time and effort, and sometimes tedious due to extensive
procedure. Hence, theoretical approaches such as classical methods and finite element
method (FEM) are preferred. Classical methods are having limited applications,
which include the application in simple structural elements like plates, shells, beam
elements, laminates, composites, etc. For modelling and predicting the dynamic
behaviour of structures with complex shapes, boundary and loading conditions, FEM
is preferred (Sehgal and Kumar 2015).

1.2 MODEL UPDATE

In most of the cases where simplifications and assumptions are made while defining
the loads, modeling the joints and boundary conditions, optimizing the mesh pattern,
and characterising the damping, there exists a conflict with the results obtained from
FEM and the experimental observations. Inaccurate response prediction from finite
element (FE) model leads to difficulties in understanding the dynamic behaviour of

4
the system. In order to overcome this drawback, model updating techniques have
emerged to aid the FE model predictions with good accuracy, so that the vibrational
behaviour of FE model matches with the actual dynamic response obtained
experimentally. Modification of uncertain parameters to make certain analytical
responses in tune with the experimental observations is the key role that can be
considered as inverse method in finite element model updating (FEMU).

In the model updating procedure, the experimental results are considered as

targets, and the inputs of FE model are updated to obtain accurate parameters. Dealing
with FE model modifications, performing multiple runs needs more computational
effort and time. Hence to reduce the complexity, FE model is replaced with an
approximate mathematical expression which relates predetermined FE inputs and
outputs. Model updating methods may be classified as sensitivity or direct methods
(Sehgal and Kumar 2015). Sensitivity type methods rely on a parametric model of the
structure and the minimization of some penalty function based on the error between
the measured data and the predictions from the model. The alternative is the direct
updating methods that change complete mass and/or stiffness matrices, although the
updated models obtained are often difficult to interpret for health monitoring
applications.

The concept of FEMU has been applied to build efficient formulations for the
analysis of structural damages, for the investigation of material properties based on
NDE characterization, and for the design based on dynamic responses. Applications
of FEMU techniques can be found in various industries like, airport, automobile,
power plants, bridges, multi-storey steel/RCC structures, mechanical tools and
equipment etc.

1.3 STRUCTURAL DAMAGE IDENTIFICATION

The SHM applications have an important role in the field of composites and aircraft
industries. A significant amount of work has been conducted using SHM techniques
to determine the various defects, damages and critical size of damage, which
influences the strength and life of composite structures. Among the various NDE
methods, the important ones are as follows: visual inspection; optical methods; Eddy

5
current; ultrasonic inspection; laser ultrasonic; acoustic emission; vibration analysis;
radiography; thermography and Lamb waves (Balageas et al. 2006). Each method has
its own advantages and limitations. For instance, offshore platforms can be analysed
using visual examination. The acoustic emission technique plays an important role in
the inspection of secondary structures in nuclear reactor core, and in case of ultrasonic
method of SHM, it is necessary to know the damage location a priori, and it renders
the structure unavailable throughout the length of the test. Many of the NDE methods
can be applied only at the certain period of time. Application of impedance-based
health monitoring techniques has gained importance, as the conventional NDE
methods might be very tedious, expensive, or unreliable. Hence, automated NDE
techniques are developed to enable real time health monitoring of civil engineering
structures. While in operations, the techniques are embedded with built-in diagnostic
system, which can be placed at desired and inaccessible locations. This built-in
diagnostic system utilizes impedance-based damage detection technique, which uses a
smart piezoelectric ceramic material (PZT). The PZT patches have been bonded to the
surface of the structural member of interest or at critical locations, which will detect
any changes in the structural/ mechanical impedance due to external loads (Park et al.
2000).

Vibration based techniques are rapidly expanding in the field of structural

damage identification (SDI). In vibration based SHM techniques, the use of natural
frequencies is considered as the most important diagnostic parameter in assessment of
structural behaviour, since natural frequencies are sensitive to the modification in the
structural integrity. Hence, by conducting systematic periodical measurement of
changes in frequencies, the structure can be monitored. This method is more reliable,
and less expensive, and hence it can be carried out frequently (Salawu 1997). The
vibration based technique is based on the basic idea that, for any structure subjected to
dynamic loads, the frequencies, mode shapes and modal damping are functions of the
mass, damping and stiffness which constitute the physical properties of the structure.
Hence, changes in these physical properties result in changes in modal properties
(Doebling et al. 1998).

6
 In most of the cases, all the structural systems are subjected to various dynamic
loads. Hence, study of those parameters, which affect the dynamic behaviour of
structural systems, is an interesting area of research. In order to evaluate the accurate
dynamic parameters of intact and damage state of structural component or structure as
a whole, an efficient and cost effective tool is required to localize and quantify the
damage scenarios. Further, based on the dynamic response parameters, the method is
subdivided into modal analysis, time domain, frequency domain and impedance
domain (Zou et al. 2000). Among the various NDE techniques, vibration based
techniques have greater importance in identifying structural damages in the past
decade, and also greater advancements in instrumentation have been achieved.
Changes in the dynamic behaviour of any structural system due to localized damage is
associated with the reduction in stiffness, increase in damping, decrease in natural
frequencies, and variation of modes. Certain dynamic responses could not be
measured in the field experiments due to various causes. Hence, using FEM, any
complex structures can be modelled by considering all the degrees of freedom, so that
the exact dynamic behaviour can be assessed and quantified (Dutta and Talukdar
2004). Damage detection using inverse methods is carried out by various researchers.
A brief overview of the use of inverse methods in damage detection and location,
using measured vibration data is available in Friswell (2007).

Development of statistical model plays an important role in enhancing the SHM

process, and there is a lot of scope in this field since least attention is given to
implementation of statistical models in current and previous applications of SHM.
Generally, in all the engineering systems, a set of responses are dependent on certain
selected parameters, and a change in these parameters results in change in behaviour
of the systems. In majority of the cases, these parameters influence the type, location
and severity of the damage. In order to assess the changes in the selected features to
identify the damaged system, the statistical models can be effectively utilized. The
algorithms used to develop statistical model need data from both the undamaged and
the damaged structures. The statistical pattern recognition algorithms are generally
classified as supervised and unsupervised learning. Group classification and
regression analysis are supervised learning algorithms. Response surface (RS),
metamodeling, linear discriminants, neural networks (NN) and genetic algorithms

7
(GA) fall under the supervised learning. Unsupervised learning methods include
control chart analysis and novelty detection methods (Sohn et al. 2001).
Metamodeling has been widely used for design optimisation, where surrogate models
are built and applied in various engineering problems (Zhao et al.2011). Additionally,
surrogate models are utilized in the analysis of stochastic structures, where the
statistical properties of dynamic variables are obtained accurately. One of the methods
adopted by Liu et al. (2015) deals with problems in load identification for stochastic
structure by combining the Gegenbauer polynomial approximation and regularization
method. The response surface method (RSM) is widely adopted in many fields owing
to its numerical efficiency. Nonetheless, the RSM is time consuming for large-scale
applications, and sometimes indicates large errors in the calculation of the sensitivity
of the parameters. In order to overcome these problems, an improved method called
high dimensional model representation (HDMR), which is basically a dimension-
reduction technique widely used in kinetic chemistry and structural reliability areas, is
widely used in many areas of specializations.

1.4 NEED FOR THE PRESENT WORK

In general, most of the structural systems are simulated with the help of FE models to
predict static as well as dynamic behaviour of the systems with different boundary
conditions. Therefore, the FE models have to be in tune with the experimental
observation to facilitate any modifications in the systems so that the future responses
can be accurately predicted, and subsequently utilized in design optimization. Thus,
FEMU is effective in improving the correlation between predicted and observed ones
by correcting the inaccurate modelling assumptions.

Many damage detection methods have attempted to identify damage by solving

an inverse problem, which inevitably needs an analytical model. However, often the
construction of these analytical model requires considerable effort in building a
mathematical framework with acceptable level of accuracy and reliability which
makes these approaches less attractive. In order to circumvent this complexity, a
computationally efficient approach in SDI using HDMR is presented in this thesis.

8
1.5 THESIS ORGANISATION

The proposed methodology involves an integrated FE modeling, development of RS

model using HDMR, establishment of objective function, and minimization of the
function using GA in order to identify the damages by updating the FE model. An
attempt has been made to reduce the computational effort, with increase in the
accuracy of updated parameter. The thesis is organised as follows.

i) The first chapter describes a brief introduction to SHM, applications and

importance of SHM in SDI. Various damage identification methods have
been mentioned including NDE and vibrations based methods. Importance of
FEMU in SHM and damage identification has been discussed along with the
need of the present study.
ii) The second chapter presents a detailed review of relevant literature on FEMU,
SDI and HDMR, followed by summary of literature and objectives of the
proposed research work.
iii) The third chapter demonstrates the application of HDMR concepts in model
updating. Two numerical examples (simply supported beam and box culvert)
have been presented to verify the efficient application of HDMR in FEMU.
iv) The fourth chapter presents a detailed study on HDMR based damage
identification. A simulated numerical example of simply supported beam with
assumed damages is considered. To substantiate the merit of the method, three
case studies are considered from the literature to validate the proposed
methodology.
v) The last chapter presents the conclusions based on the key findings from the
present work, and also the scope for the future work.

9
10
 CHAPTER 2

LITERATURE REVIEW

A detailed review of relevant articles in the field of SDI is presented in this chapter.
Also, the importance of FEMU in SDI with respect to various techniques is explored.
The developments in RSM including HDMR is also reviewed and presented. The
literature study has mainly focused on three major groups (i.e., FEMU, SDI and
HMDR) by highlighting the research gaps. Figure 2.1 summarizes a brief list of
relevant articles reviewed in these areas.

2.1 FINITE ELEMENT MODEL UPDATING

In the modern applications of engineering and science, most of the effort and time has
been invested in developing numerical models based on finite element approach.
These necessity models have greater importance in predicting the responses of the
systems which will be utilized in model assessments, understanding the behaviour of
structures under untested loading conditions or modified structural configurations,
SHM and SDI. FEM is a numerical tool based on the numerous assumptions and
simplifications (Rombach 2004). Hence, there exists a lack of correlation between
predicted and experimental observations due to inaccuracies in numerical models. The
three model errors which are considered as the main cause of inaccuracies in
numerical models prediction are the model structure error, the model parameter error
and the model order error (Mottershead and Friswell 1993). The model structure error
is due to uncertainty in model parameters and input data, where accuracy of a model
structure depends on availability of field data. A framework which integrates
quantitative and qualitative uncertainty to estimate the impact of model structural
uncertainty model predictions was developed by Refsgaard et al. (2006).

11
 The nonlinear behaviour associated with the engineering systems may also
lead to model structure error. The possibility of occurring the model parameter error is
due to incorrect parameters and boundary conditions selected and applied to the given
model. The model parameter error occurs, when there is a limited amount of data
available to estimate the parameters. This error can be estimated using confidence
intervals, bootstrap technique and Bayesian estimation. In most of the finite element
modeling procedures, assumptions are made in order to simplify the model, and the
model parameter errors are expected to occur in case of inaccurate assumptions.

Structural Damage
Identification

Sensitivity based Response Surface

HDMR
FEMU methods

Development of
Damage signatures RSM
HDMR (Rabitz and
(Wang et al. 2001) Fang and Perera (2009)
Alis 1999)

Damage functions Hybrid HDMR

RSM and GA
(Teughels and Roeck (Tunga and Demiralp
(Deng and Cai 2010)
2004) 2006)

Modal flexibility Generalised RSMU HDMR in uncertainty

residuals (Jaishi and (Shahidi and Pakzad analysis
Ren 2006) 2014) (Balu and Rao 2012)

Damage RSM and D-Optimal HDMR in inverse

parameterization design reliability analysis
(Fang et al. 2008) (Perera 2011) (Balu and Rao 2013)

Multi-objective DOE Methods RS-HDMR

optimization, PSO (Mukhopadhyay et al. (Mukhopadhyay et al.
(Perera et al. 2010) 2015) 2016)

Fig. 2.1 Summary of articles

12
 The model order errors which arise in the discretization of complex systems
can result in a model of insufficient order. The model order may be considered as a
part of the model structure. The main objective of model updating is to improve the
correlation of results between the predicted and the observed by correcting the
inaccurate modelling assumptions, and not by making other alteration to the model.
Usually the parameters like mass, stiffness and damping of the numerical model are
modified so that they will be in tune with the experimental test results. With the good
correlation between the predicted and experimental observations, the application of
FE model can be utilized in future response prediction confidently.

The FEMU emerged and became prominent in 1990s. Application concepts of

FEMU techniques have gained a greater importance in the field of design and
maintenance of mechanical and civil engineering structures in order to improve the
performance of the products of engineering design. The procedure used to update the
model is called the FEMU (Friswell and Mottershead 1995). The FEMU techniques
can be broadly classified into direct and iterative techniques (Sehgal and Kumar
2015). Several FEMU techniques have been developed by different researchers on
direct techniques (Baruch and Bar-Itzhack 1978; Berman and Nagy 1983; Bucher and
Braun 1993; Friswell et al. 1998; Modak et al. 2002; Fang et al. 2011; Li et al. 2016)
and iterative techniques (Collins et al. 1974; Levin and Lieven 1998; Modak et al.
2000; Marwala 2005; Jaishi and Ren 2007).

Imregun et al. (1995b) developed a formulation for FEMU using frequency

response function (FRF), and applied for simple beam structure. The study was
carried out to find the effectiveness of updating technique by considering noise and
incomplete experimental data. Stability and convergence were also studied in
numerical simulations, where free–free beam model was considered with known
modeling errors using measured and simulated FRF data. Model updating based on
FRF data was further extended by Imregun et al. (1995a), and the method was applied
for a plate–beam structure consisting of 500 degrees of freedom, and after several
iterations, initial FRF is updated. Hence, a better correlation with measured FRF has
been obtained which was within engineering accuracy. From FRF based model
updating method, it was concluded that, updating of FE model can be done by

13
measured data if the size of FE model is relatively small and ratio of measurement co-
ordinates to the total number of degrees of freedom has to be high about 10%. Inverse
Eigen sensitivity method (IESM) was applied by Lin et al. (1995) to update the
analytical models of engineering structures. The IESM was improved by addressing
the drawbacks of classical IESM methods, which are having slower convergence, and
the calculated coefficients are based on modal data only. The improved IESM utilized
both experimental and analytical modal data to obtain the Eigen sensitivity
coefficients, which are found to be in tune with the true values. Also the method was
further investigated for the case where measured coordinates are incomplete, and
applied to FEMU of planar truss structure.

Structural model updating based on experimental data was carried out by

Sanayei et al. (1997). The results based on the experiments performed on small scale
steel frame model were presented to support the displacement equation error function,
displacement output error function, and strain output error function methods of
structural parameter estimation using static non-destructive test data. Parameters of
the structural elements were updated using experimental static measurements and the
stiffness of structural components was considered as unknown parameter. To evaluate
the unknown stiffness, the measurements obtained from static displacements and
static strains are used, and errors are reduced by weight factors obtained from analysis
of variance of measured data. The parameters were identified with good accuracy
having low deviation with respect to true values.Atalla and Inman (1998) presented
NN based model updating using frequency domain data. A similar study was
conducted by Chang et al. (2000) where concept of adaptive NN for model updating
was applied to suspension bridge and verified both numerically and experimentally.
A correlation-based model updating algorithm was proposed to update large structural
dynamics models using measured response functions by Grafe (1998). The
formulations developed are independent of number of measurements used, and are
resistant to measurement noise. Kenigsbuch and Halevi (1998) presented a
generalised reference basis approach of updating an analytical model from
experimental data. The optimization was carried out using general weighting matrices.
Accurate parameters like mode shapes or natural frequencies are taken while updating

14
the other parameters and constrained optimization problem has been solved in order to
obtain the updated parameters.

Levin and Lieven (1998) utilized GA and simulated annealing (SA) as

optimization algorithms for dynamic FEMU, and proposed a new variant for SA
called blended-SA algorithm, which performs better than GA. The algorithms
developed are based on probabilistic search approach, which are efficient in capturing
the global minimum from the set of local minima, and comparison has been made
between GA and SA algorithms. The choice of updating parameters has a greater
impact on the accuracy of the results obtained during the optimisation process.
Brownjohn and Xia (2000) investigated the application of sensitivity-based model
updating technique to the dynamic assessment of the Safti Link Bridge, a curved
cable-stayed bridge in Singapore. From the investigation, it was concluded that the
dynamic properties obtained by the finite element analysis (FEA) for the complex
structures such as the Safti Link Bridge, are not always consistent with the measured
results due to the modelling errors and the uncertainties in the structure. Therefore, it
is necessary to improve the FE model for successful dynamic assessment of the
structure. Hence, the model updating is a feasible and effective technology for
improvement of the FE model by modification of the parameters with uncertainties
existing in the structure based on the prototype testing data.

Modak et al. (2002) used IESM and the response function method (RFM) of
analytical model updating in their study. A detailed comparison of these two
approaches of model updating was made on the basis of computer simulated
experimental data. The main objective was to study the convergence of the two
methods and the accuracy in the prediction of required corrections in a FE model. The
updated models were compared on the basis of some error indices constructed to
quantify error in the predicted natural frequencies, mode shapes and response
functions. From the study it was concluded that RFM seems to have worked better
than the IESM for the case of incomplete experimental data. In order to understand
the structural behaviour and identify the parameters, the FEMU technique was
implemented to upgrade the existing highway bridge by conducting field dynamic
testing.

15
 Using FEMU techniques, a practical method was performed to assess the load
carrying capacity by utilizing the experimental data in order to upgrade and strengthen
the bridge. Also, the influence of girder stiffness on post-performance of guard rail in
the bridge was identified (Brownjohn et al. 2003). Particle swam optimization (PSO)
technique was utilized in FEMU by Marwala (2005). Bayesian probabilistic approach
to structural model updating associated with uncertainties was proposed by utilizing
the measured dynamic responses (Katafygiotis et al. 1998). A model updating
approach for linear structural models called Bayesian model updating based on Gibbs
sampler was proposed by Ching et al. (2006). This method not only updates the
optimal estimate of the structural parameters, but also updates the associated
uncertainties. Further the Gibbs sampler approach was applied to the health
monitoring of existing structures with high-dimensional uncertain parameters which
was effective in detecting location of the damage in an instrumented structure.

Eigen frequency residual and modal strain energy residual were used as two
objective functions of the multi-objective optimization in FEMU (Jaishi and Ren
2007). Selection of updating parameters was done based on predetermined concepts
of dynamic behaviour and sensitivity study. This FEMU technique based on Eigen
frequency and modal strain energy detected the damage effectively when applied to a
simulated simply supported beam with an assumed damage. Further the method was
applied to update on-field precast continuous box girder-bridge under working
conditions. In FE model applications, finding complex FRF and complex mode
shapes will be difficult when damping matrices are not accounted in FEMU
techniques. A method of FEMU by considering damping as a parameter was carried
out by Arora et al. (2009), wherein the damping matrix was identified by the updated
mass and stiffness matrices. The method was applied for a numerical beam with
viscous damping. The method was more useful when dealing with complex updating
parameter based FEMU, and a better matching of complex FRF with experimental
data was found.

Response surface (RS) models or meta-models are considered as most efficient

models in approximating the multivariate input-output relationships of a physical
system, which are utilized to update the FE models by replacing the time consuming

16
physical based computer models. Developing RS models for model updating mainly
involves implementing various sampling techniques with the aid of design of
experiments (DOE) parameter screening, where important parameters to be updated
are selected by screening out the non-significant ones and constructing the quadratic
polynomial RS. In FEMU, setting-up of an objective function, selecting updating
parameters and using robust optimization algorithm are the three crucial steps. The
FEMU in structural dynamics based on RSM was carried out by Ren and Chen
(2010), and applicability of the RSM was illustrated by considering a full scale
precast box-girder bridge, which was tested under operational vibration conditions.
Results showed faster convergence with RSM based FEMU than conventional
sensitivity based FEMU methods. Bridge model updating using RSM and GA was
proposed by Deng and Cai (2010), where parameters were updated using the GA by
minimizing the objective function. Numerical simulations were done using the
experimental design combinations of parameters and corresponding responses from
the structure of interest. Second-order RS equations were developed, and in order to
get accurate parameter values, third order RS equations were utilized to develop
objective function. The residuals between measured and predicted responses
constituted an objective function, and updated parameters were obtained by
optimizing the function.

Generalized response surface model updating (GRSMU) method was developed

by Shahidi and Pakzad (2014), where methodology was explained to formulate an
iterative based model updating in time domain state. Also the method was
implemented to update nonlinear FE model. A numerical case-study was carried out
considering a steel frame with global nonlinearity. Well prediction of unknown
parameters was observed using GRSMU in conjunction with optimization in
simulation case studies. The case studies with large estimation error were also
evaluated using GRSMU, where noise level was low. Xiao et al. (2014) presented a
new model-updating method for updating the multi-scale FE model of a long-span
bridge. The objective functions for model updating included both modal frequencies
and multi-scale influence lines (ILs). The results showed that the differences between
the measured and the computed modal frequencies, and between the measured and
computed multi-scale ILs were all reduced after using the model-updating method.

17
The comparison of the additional measured modal frequencies and ILs with the
corresponding computed results further confirmed the quality of the model-updating
method.

An automated FEMU technique was developed using data obtained from a set
of non-destructive tests conducted on a laboratory bridge model, where both stiffness
and mass parameters were updated at the element level, simultaneously. This
approach was utilized in software packages for automated and systematic FEMU
(Sanayei et al. 2015). Updating a structural model is as an optimization problem,
where parameters minimize the errors between the model and the actual structure. In
optimization, chances of obtaining the multiple solutions are more, and hence finding
the global minimum has its own importance. A combination of GA with sequential
niche technique was proposed to increase the chance of finding the global minimum
that best describes the system (Shabbir and Omenzetter 2016). Testing, modeling and
updating of laboratory and bridge structure was carried out, where laboratory structure
consists of four columns supporting the stainless plate, and bridge considered was a
full scale cable stayed bridge. To obtain FRF, the spectral analysis was carried out.
Further the sequential niche technique was applied for updating the numerical space
frame structure. Model updating using sequential niche technique in combination with
GA yielded with satisfactory results.

Pacini et al. (2017) utilized a computationally efficient modal nonlinear

identification technique in FEMU, and experimentally demonstrated the ability to
capture typical structural nonlinearity. Gautier et al. (2017) proposed a FE based
subspace fitting approach to identify the structural parameters based on the variance
analysis for model updating, where the data-related covariance was propagated to the
updated model parameters through first-order sensitivity analysis, and vibration
signals were used to demonstrate the accuracy and practicability of the method. In
recent past, Bayesian techniques have been widely used in FEMU (Boulkaibet et al.
2017). In order to update an FE model, the Bayesian formulation requires the
evaluation of the posterior distribution function. For large systems, this function is
difficult to solve analytically. In such cases, the use of sampling techniques often
provides a good approximation of this posterior distribution function. The hybrid

18
Monte Carlo method is a classic sampling method used to approximate high-
dimensional complex problems.

2.2 STRUCTURAL DAMAGE IDENTIFICATION

The FEMU applications have been extensively utilized in identifying the damage in
civil engineering structures and health monitoring of existing bridges. In the past
decade, model updating techniques have been extended not only to detect, but also to
localize and quantify the structural damage. In the FE model, the damage is
represented by a reduction of the stiffness properties of the elements, and by tuning
the FE model to the measured modal parameters, the damages can be identified. The
most common type of detecting the damages is by visual inspections. However, this
method requires skilled labours, and in order to access various key structural load
bearing components, dismantling of the ancillary components is must which leads to
the consumption of additional time, cost and effort. Further, in case of damages
detected in later stages, additional repairs and maintenance operations will become
important concerns to rectify the damages.

Wang et al. (2001) developed an algorithm based on static test data and changes
in natural frequencies. The method was improved by proper definition of measured
and predicted damage signatures. Teughels et al. (2002) proposed a sensitivity-based
FEMU method using experimental modal data to assess the damage including
localization and quantification, where damage pattern was represented by a reduction
factor of the element bending stiffness. Damage functions were used in order to
reduce the number of unknown variables. Teughels and Roeck (2004) extended the
use of damage functions in order to approximate the stiffness distribution, and
optimization was done using the trust region strategy in the implementation of the
Gauss–Newton method. The damage in the highway bridge was identified by
updating the Young’s modulus and the shear modulus whose distribution over the FE
model were approximated by piecewise linear functions. Damage detection based on
modal flexibility residuals (i.e., one of the sensitivity based FEMU techniques) was
carried out by Jaishi and Ren (2006). For minimization of developed objective
function and damage identification, the optimization algorithm was utilized which

19
also considered the effect of noise. The method was found susceptible to change in
the physical properties of the structure. Hence, the method identified the damage
location and severity with acceptable accuracy, and the crack pattern in the damaged
structure were similar to experimental observations. The complexity of finding the
damage in the structure increased when all the elements of FE model became
updating parameters in the SDI process. Ding et al. (2017) carried out the SDI based
on the modified artificial bee colony (ABC) algorithm using modal data, and
compared the method with other evolutionary algorithms.

Fang et al. (2008) investigated a RC frame, which was tested in the laboratory
for damage identification using damage parameterization. Further, RSM was adopted
by Fang and Perera (2009) to identify the damage. The quantification of structural
damage mainly depends on the quality of the damaged model and its ability to
describe the structural property changes due to damage in a physical meaningful way
(Link and Weiland 2009). Two different model updating techniques were summarized
in conjunction with damage identification using multi-model updating: first was based
on classical modal residuals by updating undamaged and damaged models
simultaneously, and second by updating the models using residuals composed of
measured and analytical time histories. The PSO method was utilized in damage
identification problems based on multi-objective FE updating procedures by
considering modelling errors and its performance was compared with GA (Perera et
al. 2010).

Damage identification by RS based model updating using D-optimal design was

carried out by Fang and Perera (2011). The advantage of D-optimal design is that,
updating of FE model can be done effectively with minimum number of numerical
sample when the availability of samples is limited. Before updating, the non-
significant parameters are screened out using D-optimal design, and FE models are
replaced with first order RS models. These RS models are used to predict the dynamic
behaviour of undamaged and damaged structures. The method was applied to a
numerical beam, tested RC frame and full scale bridge by considering Young’s
modulus, section inertia as input parameter and modal frequency was the only output
response. An improved PSO algorithm eas developed for FEMU using experimentally

20
obtained natural frequencies by Mohamed et al. (2013), where the inverse diagnostic
optimization procedure was adopted to detect and localise the crack in beams using
frequency measurements. Additionally, the damages were detected by reducing the
number of elements in FE model, where adaptive meshing was used to detect smaller
damage cracks in beams. A comparative assessment of the damage identification
capability of different DOE methods like 2k factorial design, central composite design
(CCD), Box-Behnken design, D-optimal design and Taguchi’s orthogonal array
design was carried out by Mukhopadhyay et al. (2015).

A framework for SHM and SDI of civil structures was presented by Ebrahimian
et al. (2016), which involves integration of advanced mechanics-based non-linear FE
model updating using batch Bayesian estimation approach technique to estimate time-
invariant model parameters used in the FE model. Non-linear FEMU was carried out
in order to minimize the discrepancies between predicted and measured response time
histories by considering excitations as input and dynamic responses as output. The
updated FE model was further utilized to identify, localize and quantify the damage
and to predict the remaining useful life of the structure. The application of non-linear
FEMU method was validated by considering realistic structural FE models of a bridge
pier and a moment resisting steel frame.

Single and multiple damage assessment was done for spring mass damper
system, a beam and a composite bridge deck. From the comparative study, it was
concluded that D-optimal design and CCD are efficient DOE methods for SDI.
Sensitivity based parameter screening was done using RSM by 2k factorial and D-
optimal design. The concepts of random sampling high dimensional model
representation (RS-HDMR) were utilized to develop damage identification algorithm
under the influence of noise, and also for the purpose of parameter screening, a global
sensitivity analysis based on RS-HDMR was adopted by Mukhopadhyay et al. (2016).
A damage identification methodology based on multivariate adaptive regression
splines in conjunction with a multi-objective goal-attainment optimization algorithm
was developed for the web core fibre reinforced polymer (FRP) composite bridges,
and was validated for several single and multiple damage cases (Mukhopadhyay
2016). Real-time vibration-based structural damage detection using one-dimensional

21
convolutional neural networks was studied by Abdeljaber et al. (2017). The method
adopted performs vibration-based damage detection and localization of the damage in
real-time. The advantage of this approach is the ability to extract optimal damage-
sensitive features automatically from the raw acceleration signals. A new sensitivity-
based damage detection method was proposed to identify and estimate the location
and severity of structural damage using incomplete noisy modal data. The accuracy
and performance of the sensitivity method were numerically examined by a planner
truss by incorporating incomplete noisy modal parameters and FE modeling errors
(Entezami et al. 2017). Damage detection using power spectral density of structural
response using FEMU approach was investigated both numerically and
experimentally which adopts sensitivity based damage detection methodology. The
method can be used to detect damages in lower frequency ranges with acceptable
accuracy (Pedram et al. 2017). Roy (2017) adopted the vibration-based damage
localization technique using mode shape slope and curvature, and formulated the
expressions for the derivatives of mode shapes. Damage functions were used in order
to reduce the number of unknown variables.

2.3 HIGH DIMENSIONAL MODEL REPRESENTATION

In engineering design, spending excessive amount of time on physical experiments or
expensive simulations makes the design costly and lengthy. The severity increases
when the design problem has a large number of inputs, or of high dimension. The
HDMR is a powerful method in approximating high dimensional, expensive, and
black-box problems (Rabitz and Alis 1999). It is a set of quantitative model
assessment and analysis tools for improving the efficiency of deducing high
dimensional input–output system behaviour stimulated by applications in chemistry.
The HDMR is an approximation tool, which expresses the input and output
relationships of complex and computationally burdensome models to form a function
having hierarchical correlation expansions. A family of nonparametric multivariate
approximation functions were developed in order to understand the hierarchy of
correlations amongst the input variables (Rabitz et al. 1999). A well-ordered mapping
strategy was developed among the inputs and outputs. Alis and Rabitz (2001)
assumed that the data was randomly scattered over the entire domain and formulated

22
HDMR expressions. Their prediction was that the dimensionality of the function was
not dependent on the number of samples needed for representation to a given
tolerance, which was the efficient means to perform high dimensional interpolation. It
was recognised and concluded from the various studies that, only lower order
interaction effect between the input variables will have an effective impact upon the
output responses of a precise and explicit physical system.

The HDMR can be applied for various well-defined physical systems by making
use of this property to develop a specific, ordered mapping between inputs and
outputs. The HDMR approximation techniques are very useful in many domains if
they can represent the output to good accuracy at sufficiently low orders. Two specific
HDMR expansions are developed, i.e., analysis of variance HDMR (ANOVA-
HDMR) for statistical applications and second the cut-HDMR expansion. ANOVA-
HDMR has its application in statistics which utilize the computation of
multidimensional integrals, where cut-HDMR does not necessitate the computation of
any integrals to represent the output of a physical system. A rapid convergence in cut-
HDMR approximations is found comparable to ANOVA-HDMR where
approximation function can be obtained in more efficient manner. In a monomial
based preconditioned HDMR method (Li et al. 2001), higher order terms of cut-
HDMR expansions are expressed as lower order terms with monomial multipliers.
Here additional input-output samples are used, where there is an inadequacy in the
approximations given by the first and second order cut-HDMR correlated functions,
which avoid the utilisation of higher order terms. The concept of HDMR is used to
build simplified and efficient meta-model by replacing the original model which is
complex and nonlinear in nature. The inputs may be in the form variables such as
initial boundary conditions, control variables as per field data, functions and its
parameters and response of the system or solutions would be the output variables (Li
et al. 2002). This mathematical assumption can dramatically reduce the sampling
effort in representing the multivariate function.

Sobol (2003) investigated mathematical models described by multivariable

functions, theorems and examples on model functions with separated variables, and
global sensitivity indices for approximations. In the study, the testing of the two

23
important assumptions in HDMR was done, i.e., a model can be approximated by
using arbitrary reference points and fault in approximations caused due incorrect
choice of reference points. Theorems and examples related to ANOVA HDMR, finite
difference HDMR, model functions with separated variables, and highest order
approximations for functions with separated variables were discussed by considering
many numerical examples.

Kaya et al. (2004) developed a computer program that computes individual

components of HDMR resolution of a given multivariate function, and also
calculateed the global sensitivity indices. Numerical experiments were considered,
where HDMR functions and sensitivity indices were computed, and examined the
effect of variables of different sets on the function outputs. In further experiments, the
closeness of HDMR approximation with the real functions was estimated. The kernel
function was taken as the main part of the developed algorithm where it receives a set
of inputs and returns corresponding output value as integrals. The advantage of the
algorithm developed was that, it can generate HDMR functions of zeroth order to nth
order. However, the program can be used for only model where explicit forms are
known and models with simple symbolic integration.

For partitioning the given multivariate data into low-variate data, HDMR and
generalized high dimensional model representation (GHDMR) methods are utilized.
The above two methods worked well when multivariate data was additive in nature,
and if multivariate data has multiplicative in nature then factorized high dimensional
model representation (FHDMR) can be used. But when the nature of multivariate data
and the sought multivariate function will have hybrid nature i.e., neither additive nor
multiplicative, hybrid high dimensional model representation (HHDMR) was
obtained to get the best value for the hybridity parameter (Tunga and Demiralp 2006).
Chowdhury et al. (2008) utilized the technique of HDMR approximation to obtain an
equivalent continuous function by replacing univariate or a multivariate piece-wise
continuous function. They concluded that the HDMR is a powerful approximation
tool to obtain equivalent continuous function from univariate and multivariate piece-
wise continuous function even when the original function is characterized with
sudden peak and fall in the domain. A dramatic reduction in approximation error can

24
be found in first order HDMR approximation with increase in number of samples or
with utilization of higher order (second) HDMR. And also from this study, it was
suggested to use Moving least square (MLS) interpolation scheme rather than
Lagrange interpolation, where approximation error can be drastically reduced from
MLS. A data partitioning method (Tunga 2011), which chooses the arbitrarily
distributed points from the given grid, constructs an approximate analytical structure
by interpolation at those chosen points of the grids by utilizing HDMR expansions to
partition the given multivariate data. The above method was used to increase the
approximation quality particularly for hybrid and purely multiplicative nature
structures.

In recent years, the application of HDMR has been extended to uncertainty

analysis. When the uncertainties are represented in terms of fuzzy membership
functions, analysis of response of the structures is done using HDMR based RS
models (Balu and Rao 2012). Implicit and explicit fuzzy analysis procedures are
developed using integrated FE modelling and HDMR based RS generation. It was
concluded that HDMR approach is mathematically elegant and computationally less
expensive for the approximation of fuzzy FE response quantity. In inverse reliability
analysis (Balu and Rao 2013), the HDMR is used to get the explicit expressions
without requiring the derivatives of the response functions with respect to uncertain
variables, and fast fourier transform (FFT) techniques are used to obtain the unknown
design parameters. It was concluded that optimum number of sample points in
approximating the HDMR component functions was the most important criterion.
Moreover, to capture the nonlinearity outside the domain of sample points, very small
number of sample points should be avoided during approximation and thereby
affecting the estimated solution. Efficient uncertainty analysis was performed for
estimating the possibility distribution of structural reliability in presence of mixed
uncertain variables (Balu and Rao 2014).

Stochastic free vibration analysis of angle-ply composite plates using RS-

HDMR approach has been carried out (Dey et al. 2015) by developing a meta-model
to express stochastic natural frequencies of the system, and performance of RS-
HDMR has been compared with full-scale Monte Carlo simulation results. An

25
efficient hybrid method based on RS-HDMR and GA coupled with a local
unconstrained multivariable minimization function was investigated by
Mukhopadhyay et al. (2015) for optimization of FRP composite web core bridge deck
panels. The application of HDMR in stochastic multiscale modelling conjunction with
multi-element least square approach was carried out by Jiang and Li (2015). A local
least square HDMR was constructed in subdomains which are constructed by
adoptive decomposition of randomly chosen main domains. These local HDMRs are
represented by a finite number of orthogonal basis functions defined in low-
dimensional random spaces, where the coefficients are determined using least square
methods. Hence a global approximation HDMR was obtained by summation of all the
local HDMR approximations.

An efficient uncertainty quantification scheme for frequency responses of

laminated composite plates was investigated by bottom up surrogate based approach
using general-high dimensional model representation (GHDMR) for achieving
computational efficiency in quantifying uncertainty (Dey et al. 2016). The uncertainty
based quantification using GHDMR is applied on laminated composite structure
having complex configuration. Effect of noise on quantification of uncertainty of
natural frequency was estimated using GHDMR. Convergence study on frequency
amplitude was done for combined variation of ply orientation and the method was
validated using FEA. A critical comparative assessment has been done by Dey et al.
(2017) for different meta-models including polynomial regression, Kriging, D-optimal
design, for stochastic natural frequency analysis of composite laminates. It was found
that regression based analysis using D-optimal design was proved to be a better
technique when there is individual as well as combined variation of parameters.
However, the artificial neural network (ANN) was found to be computationally more
expensive compared to other meta models. Further in order to construct an efficient
HDMR expansion, concept of support vector regression has been adopted by Li et al.
(2017) which enables efficient construction of high dimensional models with
satisfactory prediction accuracy from a modest number of samples.

26
2.4 SUMMARY OF LITERATURE REVIEW

The SHM is an important process in SDI of civil engineering structures. Various

techniques like vibration based and impedance based techniques have their
applications in real-time monitoring of engineering structures. Many researchers in
the literature have focused on the improvements with respect to the computational
efficiency in the damage identification apart from the studies on the utilization of
conventional NDE techniques. In the modern analysis of systems in engineering and
science, much effort has been invested in developing the complex FE models. The
main purpose of such models is to predict the responses of the system to disturbances
and the design advantage gained by the modifications in the configurations. But there
will be a lack of correlation between predictions and observations due to the
inaccuracies in numerical models.

Model updating is concerned with the correction of FE models and it is rapidly

developing technology. A number of techniques, as discussed in the literature review,
has been developed for model updating including sensitivity based techniques.
Further in the past decade, RSM based FEMU technique has become an important
tool in place of conventional methods, where response equations are developed with
the aid of CCD, and the optimization has been carried out using various optimization
techniques (i.e., GA, NN, and PSO). Further, in order to reduce the computational
effort of RSM, D-optimal design became popular. In order to increase the
computational efficiency, and to reduce the complexity of modeling the real life
structures, the FEMU techniques are adopted. Also the FEMU techniques are further
utilised to identify the structural damages for various structures including bridges.

The choice of the input and output features of the system to be updated is the
key aspect of FEMU process. Initially, before proceeding to the FEMU process, the
parameters which affect the output response of the systems are found by parameter
screening procedure. From the literature, it was found that, in majority of the case
studies the variation of parameters (like the Young’s modulus, the second moment of
area, and the density) affects the dynamic behaviour of systems, and dynamic
response measured is the fundamental natural frequency of the system. Most of the
methods are validated by considering the experimental responses as the target results

27
and thereby updating the FE model by incorporating the predicted values of the
parameters obtained using the FEMU techniques. The updated parameters are
obtained by optimizing the objective function developed, which is nothing but the
difference between the response equations developed by the approximation technique
and the experimental results of interest. The FEMU has also been carried out by
considering the non-linearity associated with the FE model.

In order to address the severity in the design problems having large inputs or
high dimension, an approximation technique called HDMR has been developed to
study the input-output relationships of a system under consideration. The concepts of
HDMR have been implemented in uncertainty analysis, inverse reliability analysis,
and stochastic free vibration analysis and applied to frame and bridge structures.

Hence, the concepts of HDMR have been applied in model updating in the
proposed methodology. The FEMU is carried out by considering the simulation
studies, and the study is further extended to the SDI. The methodology is validated
using case studies from the literature.

2.5 OBJECTIVES OF RESEARCH WORK

Based on the literature review, the objectives of the present research work are as
follows.

1. To utilize the concepts of HDMR for the best experimental design of the
parameters to be updated, so as to obtain explicit function interrelating the
responses and the parameters.
2. To optimize the objective function using genetic algorithm for obtaining the
updated parameters.
3. To apply the HDMR based model updating technique in structural damage
identification.

28
 CHAPTER 3

MODEL UPDATING USING HDMR

In any physical system, the output/response depends on the input variables. Therefore,
it is necessary to learn the input-output mapping for understanding the behaviour of
any physical system. The outputs of most physical systems are mathematically well
behaved and the scarcity of the data is usually compensated by additional assumptions
on the function. The HDMR is a particular family of representations where each term
in the representation reflects the individual or cooperative contributions of the inputs
upon the output. The main assumption for most well defined physical systems is that
the output can be approximated by the sum of these hierarchical functions whose
dimensionality is much smaller than the dimensionality of the output. The present
investigation is focused on applying concepts of HDMR is in FEMU by considering
the simulation study.

3.1 HDMR
The HDMR is an assumed form of mathematical expression in which the higher order
correlated effects of the inputs are expected to have negligible effect on the output
(Rabitz et al. 1999). This mathematical assumption can dramatically reduce the
sampling effort in representing the multivariate function. HDMR has a variety of
applications where an efficient representation of multivariate functions arises with
scarce data (Alis and Rabitz 2001).

Hence, HDMR is regarded as a general set of quantitative model assessment

and analysis tools for capturing the high-dimensional relationships between sets of
input and output model variables. It is a very efficient formulation of the system
response, if higher order variable correlations are weak, allowing the physical model

29
to be captured by the first few lower order terms. In HDMR background there stands
the simple observation: only low-order correlations amongst the input variables have
a significant impact upon the outputs. Such a presumption permits expressing single
multi-dimensional mapping as a sum of many low dimensional mappings. Its main
advantages are finite order of expansion and rapid convergence for “well-defined”
systems. The concepts behind HDMR aim to capitalize on the latter observations that
realistic physical systems generally do not call for an exponentially growing number
of samples to prescribe their input-output relationships. The HDMR technique can be
applied to complex models of nonlinear nature, where an efficient and accurate
simplified model to reflect the original model can be developed.

Depending on the method adopted to determine the component of functions of

HDMR there are two particular HDMR expansions: ANOVA-HDMR and cut-
HDMR. ANOVA-HDMR is useful for measuring the contributions of the variance of
individual component functions to the overall variance of the output. On the other
hand, cut-HDMR expansion is an exact representation of the output in the hyperplane
passing through a reference point in the variable space. Applications of the HDMR
tools can dramatically reduce the computational effort needed in representing the
input-output relationships of a physical system. HDMR applications include:

 Construction of a computational model directly from laboratory/field data

 Creating an efficient fully equivalent operational model to replace an
existing time consuming mathematical model
 Identification of key model variables and their interrelationships
 Assessment of global uncertainties, quantitative risks etc.
 Solving inverse problems
 In the fields of chemical kinetics, radiative transport, materials discovery,
molecular physics, statistical analysis, and financial and econometrics

3.2 HDMR EXPANSIONS

The HDMR expansions introduced here are especially useful for the purpose of
representing the outputs of a physical system when the number of input variables are

30
large. The notion of “high” dimensionality is system-dependent, with some situations
being considered high for practical reasons at N 3  5 , while others will only reach
that level of complexity for N 10 or more. For a high dimensional system, an output
f  x  is commonly a function of many input variables x   x1 , x2 ,, xN  with n 10 2

or larger. The HDMR approximations should not be viewed as first- or second-order

Taylor series expansions nor do they limit the nonlinearity of f ( x) . Furthermore, the
approximations contain contributions from all input variables. Thus, the infinite
number of terms in the Taylor series is partitioned into finite different groups, and
each group corresponds to one HDMR component function. The HDMR expresses the
output as a hierarchical correlated function expansion in terms of the input variables
as:

f ij  xi , x j   f ij  xi , x j , xk  
N
f  x   f 0   f i  xi    
i 1 1i  j  N 1i  j  k  N (3.1)
 f1,2,... N  x1 , x2 , , xN 

where, f 0 denotes the mean response to f  x  which is a constant. The function

fi  xi  is a first-order term expressing the effect of variable xi acting alone, although

generally nonlinearly, upon the output f  x  . The function f ij ( xi , x j ) is a second-

order term that describes the cooperative effects of the variables xi and x j upon the

output f  x  . The higher order terms give the cooperative effects of increasing

numbers of input variables acting together to influence the output. The last term
f12N  x1 , x2 ,  , xN  contains any residual dependence of all the input variables locked

together in a cooperative way to influence the output. To determine the component

functions in Eq. (3.1), cut-HDMR procedure is used in approximating a univariate or
a multivariate piece-wise continuous function with an equivalent continuous function.

Using the cut-HDMR method, first a reference point c  c1 , c2 cN  is defined in the
variable space. The expansion functions are determined by evaluating the input-output
responses of the system relative to the defined reference point c along associated
lines, planes, sub-volumes, etc. (i.e. cuts) in the input variable space. This process
reduces to the following relationship for the component functions in Eq. (3.1):

31
 f0  f  c  (3.2)


fi  xi   f xij , ci  f 0  (3.3)

fij  xi , x j   f x j1
i1 
, xi2j2 , ci1 ,i2  fi  xi   f j  x j   f 0 (3.4)

By evaluating the response quantities at all sample points of each variable including
the reference point the following expressions are obtained.

f  xij , ci   f  c1 ,..., ci 1 , xij , ci 1 ,..., cN  (3.5)

  
f xi1j1 , xi2j2 , ci1 ,i2  f c1 , ..., ci1 1 , xi1j1 , ci1 1 , ..., ci2 1 , xi2j2 , ci2 1 ..., cN  (3.6)

Using the Lagrange interpolation or the moving least squares interpolation yields Eq.
(3.7) and Eq. (3.8) for first-order and second-order expressions respectively as
follows.
N
f  xij , ci    j  xi  f  c1 , ..., ci 1 , xij , ci 1 , ..., cN  (3.7)
j 1

   
n n
f xi1j1 , xi2j2 , ci1 ,i2  j1 j2 xi1 , xi2
j1 1 j2 1 (3.8)
 j1 j2
f c1 ,  , ci1 1 , xi1 , ci1 1 ,  , ci2 1 , xi2 , ci2 1 ,  , cN 
By summing up the interpolated values of HDMR expansion terms from zeroth order
to the highest order retained in keeping with the desired accuracy, the first- and
second order approximations of the functions are as follows.
n
f  x     j  xi  f  c1 , ..., ci 1 , xij , ci 1 , ..., cN    N  1 f 0
N
(3.9)
i 1 j 1

   x , x  f  c ,  , c 
N n n
f  x  j1 j2 i1 i2 1 i1 1 , xi1j1 , ci1 1 ,  , ci2 1 , xi2j2 , ci2 1 ,  , cN
i1 1,i2 1, j1 1 j2 1
i1 i2
(3.10)
 N  1 N  2  f
 N  2  j ( xi ) f  c1 ,  , ci 1, xij , ci 1 ,  , cN 
N n
  0
i 1 j 1 2

The shape/interpolation function  j  xi  and j 1 j2 xi 1 , xi2 using the Lagrange  

interpolation is defined as:

32
  j  xi  
 x  x   x  x  x  x   x  x 
i
1
i i i
j 1
i i
j 1
i i
n

(3.11)
 x  x   x  x  x  x   x  x 
i
j 1
i i
j
i
j 1
i
j
i
j 1
i
j n
i

x   x  x
 xi1
1 j1 1
 x   
 xij 1  xi1  xin
1

  x x  x x
i1 i1 i1 i1
 j1 j2 xi1 , xi2
1 1

 j1
i1   1
i1 i1
j1 j1 1
i1  x  x  x  x 
j1
i1 i1
j1 1
i1
j1 n
i1
(3.12)


 x  x   x  x
i2
1
i2 i1
j1 1
i1  x  x  x  x 
i1
j1 1
i1 i1
n
i1

 x  x   x  x
j2
i2
1
i2
j2
i2
j2 1
i2  x  x  x  x 
j2
i2 i2
j2 1 j2
i2
n
i2

HDMR expressions are obtained by evaluating the component functions in Eq. (3.1),
which can be utilized to replace the original, complex, and expensive methods
efficiently without compromising with the accuracy of the model. Expression for first-
order approximation with uniformly distributed sample points (n) given by
µi   n  1  i / 2, µi   n  3  i / 2,..., µi ,..., µi   n  3  i / 2, µi   n  1  i / 2 , where

n  3,5, 7 are deployed along the variable axis xi with mean ( i ) and standard

deviation (  i ) through the reference point. The sampling scheme for first-order

HDMR for a function having one variable  x  and two variables  x1 and x2  is shown

in Fig. 3.1 (a) and (b), respectively. Similarly, Fig. 3.2 represents sampling scheme for
second-order HDMR for a function with two variables.

x2

c x c x1

(a) (b)

Fig. 3.1. Sampling scheme: First-order HDMR: (a) function with one variable ( x ) and (b)
function with two variables  x1 and x2 

33
 x2

c x1

Fig. 3.2. Sampling scheme: Second-order HDMR

3.3 MODEL UPDAITNG USING HDMR

This section explains the proposed approach of utilizing the concepts of HDMR in
model updating adopted for damage identification. FEMU has been carried out using
first order HDMR approximation functions. The HDMR based SDI procedure
consists of six steps as presented in Fig. 3.3.

Objective Structure

Feature Selection

Parameter screening

Development of HDMR equations

Optimization by Genetic algorithm

Model updating using HDMR

No
Convergence

Yes

Damage Identification

End

Fig. 3.3 Flow chart of HDMR based model updating and damage identification

34
In feature selection, the material properties, such as the Young’s modulus, the density,
and the geometric properties, are usually adopted as the input parameters in model
updating. And, for the output features, time and frequency domain features are two
feasible options. In parameter screening, non-significant inputs that have least
influence on output responses are screened out using regression analysis. Further,
HDMR response equations are developed to map the input-output relationships.
Objective functions are built using HDMR response equations, which are optimized
using GA to obtain the updated parameters. Further the updated FE model is utilized
for structural damage identification.

3.4 NUMERICAL EXAMPLES

Two numerical examples are considered in order to study the efficiency of HDMR in
predicting the updating parameters of interest. Without any physical tests, the
damages are assumed at specific locations and response quantities are evaluated for
further investigations.

3.4.1 Example 1: Simply Supported Beam

To illustrate the applicability of HDMR technique in FEMU, a simply supported

beam (Deng and Cai 2010) is considered for the present simulation study. This
example demonstrates how RS generation can be done using proposed HDMR, in
conjunction with the GA in order to obtain accurate parameters. The cross section and
material property of the beam is assumed to be uniform throughout its length. Three
parameters, the Young’s modulus ( E ) , the density (  ) and the Poisson’s ratio (  ) of
the material, are chosen as the input parameters. Figure 3.4 shows the beam
considered for the present study, which is divided in to 15 elements.

200 N 200 N

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
6m

Fig. 3.4 Simply supported concrete beam

35
The span and cross section of the beam are taken as 6 m and 200 mm × 200 mm
respectively. For the simulation study Young’s modulus  E  , density    and
Poisson’s ratio    are taken as 20 GPa, 2,400 kg/m3 and 0.2 respectively

For the present case, it is assumed that, elements at location 4 and 8 are
damaged. The initial values of E at two locations (4 and 8) are taken as 20 GPa and ρ
as 2,400 kg/m3 respectively. Assuming a unit change for each of the three parameters
to be 50%, 50%, and 20% of the baseline values respectively, the first three natural
frequencies from the modal analysis and the deflection at the bottom of the section
near the mid span of the beam from the static test are obtained as responses.

3.4.1.1 Response surface generation

For development of approximation equation, number of variable (N) for the static test
is taken as 3 and number of sample points (n) as 3. The function evaluations required
for developing first order HDMR approximation equation is  n  1 N  1 . Hence seven

function evaluations are performed using FEA package. Considering first-order

HDMR approximation technique:
N
f  x   f  x1 , x2 , x3   f 0   f i  xi    N  1 f 0 (3.13)
i 1

f  x   f 0  f1  x1   f 2  x2   f3  x3    N  1 f 0 (3.14)

f 0  f  c1 , c2 , c3  (3.15)

f  x   f  x1 , x2 , x3   f  c1 , c2 , c3   f1  x1   f 2  x2   f 3  x3    N  1 f 0 (3.16)

where

f  xi   f  xij , ci   f 0 (3.17)

f  xi , c   j  xi  f  c1 , , ci 1, xij , ci 1, , cN 

n
j i
(3.18)
j 1

Therefore, from Eq. (3.17)

36
 f  x1   f  x1 , c1 
 1  x1  . f  x11 , c2 , c3   2  x1  . f  x12 , c2 , c3  (3.19)
3  x1  . f  x13 , c2 , c3   f 0

f  x2   f  x2 , c 2 
 1  x2  . f  c1 , x12 , c3   2  x2  . f  c1 , x22 , c3  (3.20)
3  x2  . f  c1 , x23 , c3   f 0

f  x3   f  x3 , c 3 
 1  x3  . f  c1 , c2 , x31   2  x3  . f  c1 , c2 , x32  (3.21)
3  x3  . f  c1 , c2 , x33   f 0

Let c1 , c2 , c3 be the reference points where output response of the system is

evaluated. The values of c1 , c2 , c3 are taken as 20 GPa, 20 GPa and 2200 kg/m3

respectively. x1 , x2 , x3 are the three parameters considered i.e., E at two locations (4

and 8) , and ρ of beam respectively, where x11 , x12 , x13 are three sample points taken as

10 GPa, 20 GPa, 30 GPa respectively. Similarly x12 , x22 , x23  10, 20,30  GPa and

x31 , x32 , x33  (1920, 2200, 2880) kg/m3. Using the above values, the required responses
are found out using FEA for the component functions in Eq. (3.18) which constitute
the approximation function.

Hence,
f  x11 , c2 , c3   f 10, 20, 2200   7.11 Hz

f  x12 , c2 , c3   f 10, 20, 2200   7.32 Hz

f  x13 , c2 , c3   f 1920, 20, 2200   7.39 Hz

f  c1 , c2 , c3   f  20, 20, 2200   7.32 Hz

Using Lagrange Interpolation function,

1  x1    x1  20 x1  30   0.005  30  2.5 x1  0.05 x12

2  x1    x1 10 x1  30  0.01   3  0.4 x1  0.01x12 

37
 3  x1    x1  10 x1  20  0.005  0.15  0.15 x1  0.005 x12

From Eq. (3.16) and Eq. (3.19),

f  x   f  x1 , x2 , x3 
 f  c1 , c2 , c3   f1  x1   f 2  x2   f 3  x3   ( N  1) f 0

f  x1   f  x1 , c1 
 1  x1  . f  x11 , c2 , c3   2  x1  . f  x12 , c2 , c3   3  x1  . f  x13 , c2 , c3   f 0

f  x1 , c1   192.448  15.955x1  0.31875x12

Similarly, all the component functions are obtained using Eq. (3.20) and (3.21) as:
f  x2 , c2   185.562  15.394 x2  0.3082 x22

f  x3 , c3   3.119  0.23x3  0.009 x32

Substituting the component functions in Eq. (3.19), the expression for first response

i.e., first natural frequency Y1  is obtained as follows.

Y1  360.252  15.955 x1  15.394 x2  0.223x3  0.3192 x12

(3.22)
0.308 x22  0.0088 x32

Similarly using the above procedure, approximation functions for second natural

frequency Y2  third natural frequency Y3  and the deflection at mid span Y4  are

obtained as follows.

Y2  1457.93  61.451x1  65.3054 x2  1.03054 x3  1.2303x12

(3.23)
1.30613x22  0.05452 x32

Y3  3215.443  141.238 x1  138.429 x2  3.6532 x3  2.8267 x12

(3.24)
2.77145 x22  0.20964 x32

Y4  380.4051  15.386 x1  17.003 x2  4.209346 x3  0.3075 x12

(3.25)
0.3394 x22  0.8162 x32

3.4.1.2 Optimization using GA

Since no physical tests are conducted, an assumption has been made that the damage
in numerical beam at locations 4 and 8 is due to reduction in E from 20 GPa to 15 and

38
12 GPa respectively with an assumed value of  as 2,200 kg/m3. The reduction of E
and  at location 4 and 8 are taken as true values. An objective function is then built
up which is nothing but the difference between the responses predicted from HDMR
approximation function and the true values.

The responses considered for the present simulation case are natural frequencies
from the first three modes from modal analysis and deflection at the mid-span of the
beam due to application of the static loads, which are found to be 6.954, 28.459,
69.916Hz, and 7.249 mm, respectively. Further, the effect number of HDMR response
quantities required to build an objective function for accurately predicting the values
of the updating parameters is studied. Four objective functions are built with each
function having one to four response quantities, and are shown below.

 Y1  6.954 
1 2
Fobj (3.26)

 Y1  6.954   Y4  7.249 

2 2 2
Fobj (3.27)

 Y1  6.954   Y2  28.459   Y4  7.249 

3 2 2 2
Fobj (3.28)

 Y1  6.954   Y2  28.459   Y3  69.916   Y4  7.249 

4 2 2 2 2
Fobj (3.29)

The updated parameters obtained by minimising the all the four objective
function with different number of response equations are presented in Table 3.1.
Results from Table 3.1 indicate that the number of responses needed in objective
function should be at least no less than the number of parameters to be identified.
4
Hence, for Fobj case, updated parameters are close to true values. Further the study is

extended to utilize HDMR response equations in FEMU.

In order to know the effect of number of sample points on the accuracy of the
updated parameters, same simulated beam as shown in Fig. (3.4) is considered and in
this case, the responses considered are first four natural frequencies of the beam. Also
the efficiency of HDMR is compared with the RSM in terms of difference values with
respect to true parameters, and number of function evaluations.

39
 Table 3.1 Identified values of parameters using GA

Identified values
Objective
Function True HDMR
Parameters
Value 1st Order
1
Fobj x1 (GPa) 15 18.98
x2 (GPa) 12 15.46
x3 (103 kg/m3) 2.2 2.57
2
Fobj x1 (GPa) 15 15.52
x2 (GPa) 12 19.29
x3 (103 kg/m3) 2.2 2.3
3
Fobj x1 (GPa) 15 14.31
x2 (GPa) 12 21.5
x3 (103 kg/m3) 2.2 2.48
4
Fobj x1 (GPa) 15 15.94
x2 (GPa) 12 11.55
x3 (103 kg/m3) 2.2 2.65

Before proceeding to RS generation, a three factor CCD with 18 trials (i.e.,

eight corner points, six star points and four replicates at centre points) are utilized to
perform the simulation study (Deng and Cai 2010). Parameter screening has been
carried out using regression analysis to find the percentage contribution of each
parameter to the total variance of the output. From the initial values, the parameter
bounds (±1) are fixed to ±30%. Frequencies from the first four modes is evaluated
using numerical simulations, and screening results are presented in Fig. 3.5. From the
screening results, the values of E and  are considered for further investigation, and
 is screened out as it has zero contribution towards the total variance.

40
 60
Young's Modulus
Density
Poissions ratio

Contribution to the total Variance (%)

50

40

30

20

10

0
f1 f2 f3 f4
Frequency

Fig. 3.5 Parameter screening results

Based on the assumption that the beam is damaged at element locations 4 and
8, the true values are as follows. E4  15 GPa , E8  12 GPa and 4,8  2200 kg/m3 .

Further in order to obtain the true values of the parameters  E4 , E8 and 4,8  , the

concept of model updating using HDMR is applied to the numerical beam. Let
c1 , c2 , c3 be the reference points where output response of the system is evaluated. The

values of c1 , c2 , c3 are taken as 20 GPa, 20 GPa and 2200 kg/m3 respectively. The

beam is updated by considering different number of sample points (ie, n  3,5 & 7 ).
The HDMR approximation equations for first four natural frequencies are developed.
Then an objective function which is the difference between the responses obtained
(natural frequencies: 6.951 Hz, 10.771 Hz, 28.415 Hz and 35.667 Hz) based on the
true parameter values  E4  15 GPa , E8  12 GPa and 4,8  2200 kg/m3  and the

HDMR response equations. Table 3.2 shows the responses obtained from function
evaluations.

Y1  6.951  Y2  10.771  Y3  28.415  Y4  35.667 

2 2 2 2
F  (3.30)
obj

where Y1 , Y2 , Y3 and Y4 denote the HDMR response equations, for first four natural
frequencies.

The objective function developed in Eq. (3.30) is optimized using GA by defining the
lower and upper bounds of three parameters ( E4 , E8 and  4,8 ) as [10 GPa, 10 GPa,

1920 kg/m3] and [30 GPa, 30 GPa, 2880 kg/m3] respectively. Numerical beam is

41
investigated by updating the model for n  3,5 & 7 in order to minimize the errors of
all the parameters. The sample points for n  3,5 & 7 considered are within the
parameter bounds. Results obtained for all the sample points are compared with RSM
(Deng and Cai 2010).

Table 3.2 Responses for development of HDMR expression (Beam example)

f  c1 ,..., ci 1 , xij , ci 1 , ..., cN 

f1 f2 f3 f4
(Hz) (Hz) (Hz) (Hz)

f 10, 20, 2400  7.04 10.84 27.24 34.49

f 13.33, 20, 2400  7.15 11.07 28.00 35.40

f 16.66, 20, 2400  7.21 11.22 28.51 36.02

f  20, 20, 2400  7.25 11.32 28.87 36.47

f  23.33, 20, 2400  7.28 11.40 29.14 36.81

f  26.66, 20, 2400  7.31 11.47 29.35 37.08

f  30, 20, 2400  7.33 11.50 29.52 37.29

f  20,10, 2400  6.82 10.61 28.82 36.06

f  20,13.33, 2400  7.02 10.95 28.85 36.26

f  20,16.66, 2400  7.16 11.17 28.86 36.38

f  20, 23.33, 2400  7.32 11.44 28.88 36.54

f  20, 26.66, 2400  7.38 11.52 28.89 36.59

f  20,30, 2400  7.42 11.59 28.90 36.63

f  20, 20,1920  7.40 11.52 29.26 37.03

f  20, 20, 2080  7.35 11.45 29.13 36.84

f  20, 20, 2240  7.30 11.33 29.00 36.65

f  20, 20, 2560  7.21 11.26 28.75 36.29

f  20, 20, 2720  7.16 11.20 28.62 36.11

f  20, 20, 2880  7.12 11.14 28.50 35.93

Table 3.3 presents the variation of values obtained using HDMR, and the
second- and third-order RSM (Deng and Cai 2010). The model updating using
HDMR with n  5 results in less absolute error of 0.5%, 1.71% and 0.07% for E4 ,

42
E8 and  4,8 respectively, which is more accurate than the results reported using

conventional RSM (Deng and Cai 2010). Also, the computational effort is calculated
in terms of number function evaluations required for construction of the functions.

Table 3.3 Variation of values (Error percentage) for Beam

% Error
True RSM HDMR
Variables
Values (Deng and Cai 2010)
Second order Third order n=3 n=5 n=7

E4 (GPa) 15 5.99 2.22 3.59 0.50 0.45

E8 (GPa) 12 13.31 0.55 4.7 1.71 1.70

 4,8 (kg/m3) 2200 0.55 0.91 0.19 0.07 0.06

Table 3.4 presents the number of function evaluations required for RSM and HDMR
methods. The conventional RSM requires 18 function evaluations whereas the
HDMR requires only 7 function evaluations, and provides more accurate prediction
than the RSM. In the proposed work, number of sample points in an axis considered
for evaluation is increased to investigate the parametric study.

Table 3.4 Computational effort (Beam Example)

RSM HDMR
(Deng and Cai 2010)
Second Order Third Order n3 n=5 n=7
# Fn.
18 18 7 13 19
Evaluations
It is observed that the increase in number of sample points results in more accurate
prediction of results. However, a little more computational effort is required which is
still less than the RSM. Therefore, the proposed approach is able to obtain the
accurate parameters with less computational effort.

3.4.2 Example 2: Reinforced Concrete Box Culvert

The concept of HDMR based model updating has been applied for an existing RC box
culvert located at Surathkal, India. Over the RC box culvert a four lane highway
traffic flow is present which is along the length of the culvert. The RC box culvert has
an expansion joint, hence two box culvert structures are present, each providing a path

43
for traffic traveling in one direction. Hence only one box culvert is considered for the
present investigation. The RC box culvert structure under study has a total width of 54
m, where expansion joint is located a 27 m and length of each span measures 6 m, and
both the thickness of deck slab and RC wall are 0.8 m. Based on the configuration of
the box culvert, a numerical model was created using a commercial FE software. A
fixed boundary condition has been assigned between RC slab and vertical wall
system. Similar to the simply supported beam study, the box culvert is assumed to
have damages or stiffness changes due to moving load of the vehicle at two locations,
(location A and location B) one at the middle of first span and second at the corner of
the second span as shown in Fig. 3.6. Three parameters are considered as variables
i.e., Young’s modulus of concrete at location A  EA  , Young’s modulus of concrete

at location B  EB  and the density of concrete at two locations  A,B  .

Fig. 3.6 Finite element model of RC box culvert

For the simulation study, the values of E ,  , and  are taken as 30 GPa, 0.2,
and 2,500 kg/m3 respectively. Truck loadings are considered as per IRC: 3-1983 and
type of vehicle considered is Type 3-S2. Let c1 , c2 , c3 be the reference point (30 GPa,
30 GPa, 2500 kg/m3) where output responses of the system is evaluated. It is assumed
that damages have occurred at location A and location B, and the true values for the E
at two locations are taken as 25 GPa and 22 GPa. The true value for ρ is taken as
2,300 kg/m3.

44
The expressions for first, second and third natural frequencies are developed using
first-order HDMR as follows.

Y1  3.97 105 x16  6.22 1010 x26  2.35 1030 x36  0.069 x15  1.133  107 x25
5.832 1020 x35  0.468 x14  8.388  106 x24  7.29  1016 x34  17.8 x13
(3.31)
3.22 104 x23  3.62  1012 x33  356.63x12  0.0068 x22  8.97  109 x32
3701.3x1  0.0749 x2  1.896  105 x3  15563

Y2  6.76 105 x16  1.51109 x26  1.03 10 18 x36  0.011x15  2.76 10 7 x25
1.55 1014 x35  0.827 x14  2.04 10 5 x24  9.68 10 11 x34  30.27 x13
(3.32)
7.88 104 x23  3.20 10 7 x33  606.19 x12  0.0165 x22  5.94 10 4 x32
6291.4 x1  0.178 x2  0.585 x3  26694

Y3  8.47 105 x16  3.88 1019 x36  0.0147x15  8.00 109 x25
5.899 1015 x35  1.036 x14  1.266  106 x24  3.711 1011 x34
(3.33)
37.91x1
3
 7.83 105 x23  1.23 107 x33  759.163x12  0.0023x22
2.31104 x32  7878.9 x1  0.0331x2  0.2302 x3  33224

The HDMR functions are developed for deflection at the center of each span (Y4 and
Y5). Deflections are taken on the upper portion of the concrete slab where moving
loads are considered. The response functions for deflection at center of the first span
are developed using first-order HDMR as follows.

Y4  8.390  108 x16  5.333 10 33 x36  1.456 10 5 x15  8.094 10 29 x35
0.001x14  5.10110 25 x34  0.0375 x13  1.708 10 21 x33  0.7518 x12 (3.34)
3.207 10 18 x32  7.803 x1  3.199 10 15 x3  32.834

Y5  9.165  10 8 x16  1.777 10 11 x26  5.42110 33 x36  1.590 10 5 x15
3.466 10 9 x25  8.228 10 29 x35  0.001x14  2.711 10 7 x24
5.186 10 25 x34  0.041x13  1.083 10 5 x23  1.737 10 21 x33 (3.35)
0.821x12  2.312 10 4 x22  3.260 10 18 x32  8.524 x1  0.002 x2
3.252 10 15 x3  35.834

The HDMR approximation equations for five responses are developed i.e., first three
natural frequencies and deflection at the center of two spans of box culvert. Then an
objective function is developed which is the difference between the responses
obtained (21.319 Hz, 36.251 Hz, 45.328 Hz, 0.159 mm, 0.157 mm) based on the true

45
parameter values  EA , EB and A,B  ,and the HDMR response equations. The objective

function is as shown in Eq. (3.36):

Y1  23.319   Y2  36.251  Y3  45.328

2 2 2

F  (3.36)
obj  Y4  0.159   Y5  0.157 
2 2

where Y1 , Y2 and Y3 denote the first three natural frequencies of the bridge, Y4 denotes

deflection at center of first span, and Y5 denotes the deflection at center of second
span obtained using HDMR.

The objective function developed in Eq. (3.36) is optimized using the GA by

defining the lower and upper bounds as [30 GPa, 30 GPa, 2500 kg/m3] and [45 GPa,
45 GPa, 3000 kg/m3] respectively. The predicted values and the percentage variation
of true values of EA , EB and  A,B for n  3,5 & 7 are presented in Table 3.5.

The values of absolute errors diminish when sample points (n) are increased
from n  3 to n  5 and converges at sample point n  7 . It is observed that for n  7 ,
the error with respect to  A,B is only 0.26% while for EA and EB , it is around 2.63%

and 2.21% respectively. Since there is no significant improvement in reduction of

percentage error between n  5 and n  7 , the optimum number of sample points is
considered as 5. The function evaluations required for the HDMR functions are 7, 13
and 19 for n  3,5 & 7 respectively. It is observed that the increase in the number of
sample points results in more accurate prediction of results, but with a little more
computational effort.

Table 3.5 Variation of values (Error percentage) for Box-culvert

True % Error
Variables
Values n3 n5 n7

EA (GPa) 25 35.88 2.64 2.63

EB (GPa) 22 6.16 2.32 2.21

 A,B (kg/m3) 2300 13.04 0.28 0.26

46
In order to make the parameters in Eq. (3.36) dimensionless, the normalization
procedure is tested. Since the identified parameters are insensitive to the
normalization procedure, the results are not presented. Therefore, the methodology
proposed is suitable for accurate prediction of parameters for model update, and the
method is computationally efficient. Hence the HDMR based FEMU is further
employed in SDI.

3.5 SUMMARY

The concept of HDMR is applied to update the FE model in order to obtain the
accurate parameters. Two numerical simulations (a simulated simply supported beam
and an existing RC box culvert) have been carried out. In the numerical simulations,
damages are assumed at particular locations due to reduction of stiffness. From the
simulation study, the applicability of HDMR in FEMU has been studied by
conducting a parametric study with respect to number of sample points. And the
results of the proposed method are compared with conventional RSM.

47
48
 CHAPTER 4

STRUCTURAL DAMAGE IDENTIFICATION

USING HDMR

In order to validate the proposed methodology, HDMR based FEMU is applied to find
out the damage pattern, location and severity. Unlike numerical beam examples, the
damages are not known in advance, while updating FE model of the real structures.
Hence, it is interesting to solve the case-study example with respect to identification
of the damage location, which defines the objective of the problem as shown in Fig.
3.2. Three case studies have been considered in this work to validate the proposed
methodology. The damage pattern, location and severity are correlated with the
experimental investigations from the literature.

4.1 SIMULATION STUDY: SIMPLY SUPPORTED BEAM

In this simulation study, selected elements of the simply supported beam are assumed
to have damage at three locations, and rest of the beam was considered with no
damage elements (Jaishi and Ren 2006). The density    and Young’s modulus  E 

of the concrete are 2500 kg/m3 and 3.2e+10 N/m2, respectively. Size of the beam
considered is 250 mm × 200 mm as shown in Fig. 4.1. Modal analysis is performed
by using FEA package for both damaged and undamaged beam. Damages are
assumed by reducing the E of the three elements (3, 8 & 10) by 20%, 50% and 30%
respectively.

Fig. 4.1 A simulated simply supported beam

49
 To develop the response equations using HDMR, the elastic modulus is
considered as variable, and elastic moduli of the 15 elements of the beam are
considered as updating parameters  x1 to x15  . The first ten fundamental natural

frequencies are considered as output responses. The objective of the proposed

methodology is to find out the damage location and severity by updating the
undamaged beam by considering the first order approximation equation with number
of variables, N  15 and sample points n  3,5 & 7 . Using the above values,
responses are found out for functions using FEA for all the three parameters to
constitute the component functions of HDMR.

Let c1 to c15 be the reference point, where output response of the system is
evaluated and taken as 32 GPa. The lower and upper bounds considered for E are 16
GPa and 48 GPa respectively and sample point for n  3,5, &7 are taken within the
parameter bounds. The HDMR expressions are developed for first ten natural
frequencies and objective function can be written as follows:

Y1  8.25  Y2  34.85  Y3  74.68  Y4  135.43

2 2 2 2

Fobj   Y5  141.06   Y6  204.69   Y7  298.58 

2 2 2
(4.1)
 Y8  386.98   Y9  417.32   Y10  494.62 
2 2 2

where, Y1 to Y10 denote the first ten natural frequencies of the beam obtained using
HDMR. Objective function developed in Eq. (4.1) is optimized using the GA. Table
4.1 shows the initial frequencies of undamaged and damaged beam where frequencies
of damaged beam are taken as true values.

The beam is updated first by considering three sample points  n  3 . It is

found that, after FEMU, the frequency values are close to damage state model with
the maximum difference of 5.92% in mode 1 (Fig. 4.2).

50
 Table 4.1 Initial frequencies of FE model: Simulated beam

Undamaged Damaged Difference in

Mode
beam (Hz) beam (Hz) frequencies (%)
1 9.00 8.26 8.22
2 35.86 34.84 2.84
3 80.13 74.69 6.78
4 141.03 135.43 3.97
5 149.00 141.06 5.32
6 217.39 204.69 5.84
7 307.52 298.58 2.90
8 409.13 386.98 5.41
9 445.38 417.32 6.30
10 519.16 494.62 4.72

Figure 4.2 indicates the variation of percentage frequencies with respect to

damaged beam frequencies in ten modes after model updating for n  3 and the results
show the reduction in percentage difference after updating. However, the values are
slightly high, and it requires further refinement of parameter values. Also from Fig.
4.3, the prediction of damage location and severity is found to be less accurate, since
percentage damage at location 3, 8 and 10 are found to be 5.3%, 26.9% and 4.1%
respectively, which less than the assumed damage percentage. Hence the HDMR
based model updating is carried out for n  5 .

Before updating
8 After updating (n=3)
Difference in frequencies (%)

0
1 2 3 4 5 6 7 8 9 10
Mode

Fig. 4.2 Difference in frequencies after model updating  n  3 

51
 50

40

30
Damage %

20

10

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Element Location

Fig. 4.3 Identified damage locations and stiffness distribution  n  3

When number of sample points is increased from 3 to 5, significant

improvement in the predicted frequencies are found, since reduction of difference in
frequencies is witnessed from 5.92% for n  3 (Fig. 4.2) to 0.48% for n  5 (Fig. 4.4)
in mode 1, and also in other modes. Also improvement in prediction of damage
location is observed. The variation of difference in frequencies after updating, using
n  5 is presented in Fig. 4.4, which shows a good agreement with the true responses
in the damaged beam.

Before updating
8 After updating (n=5)
Difference in frequencies (%)

0
1 2 3 4 5 6 7 8 9 10
Mode

Fig. 4.4 Difference in frequencies after model updating  n  5  

52
 Results obtained by updating the value of E of the 15 beam elements using
five sample points are presented in Table 4.2. 

Table 4.2 Updated frequencies of simulated beam  n  5 

After Difference in
Before Damaged
Mode updating frequencies
updating (Hz) beam (Hz)
(Hz) (%)
1 9.00 8.26 8.22 0.48
2 35.86 34.84 34.21 1.81
3 80.13 74.69 75.58 1.19
4 141.03 135.43 135.47 0.02
5 149.00 141.06 141.32 0.18
6 217.39 204.69 204.22 0.22
7 307.52 298.58 298.41 0.05
8 409.13 386.98 387.19 0.05
9 445.38 417.32 418.76 0.34
10 519.16 494.62 495.62 0.20

Figure 4.5 indicates that the damage has been located at element 3, 8 and 10
with the damage percentage of 31% 45% and 34% respectively. Values of damages
have appeared on the undamaged elements i.e., 5, 11 and 12, damage is found to be
13% and for element location 6 and 15 it is found to be 6% and 9% respectively.
50

40

30
Damage %

20

10

0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Element Location

Fig. 4.5 Identified damage locations and stiffness distribution  n  5 

53
 Even though the damage percentage in three elements are higher than the
assumed ones, the results are acceptable due to a significant reduction in the
differences observed in frequencies. Hence the updated FE model is able to predict
the responses with good accuracy with the difference in frequencies less than 0.5%,
except for mode 2 and 3 which is found to be 1.81% and 1.19% respectively as shown
in Table 4.2. From the parametric study, it is found that, increase in number of sample
points yields better responses with lesser percentage difference in frequencies in
comparison with damaged beam frequencies. Hence after updating, the average
difference between the experimental and numerical frequencies decreases from 4.42%
(Fig. 4.2) to 0.45% (Table 4.2) with the increase in sample points from n  3 to 5
respectively. Similar study is extended to update the model for finding the damage
locations and severity using n  7 . As the obtained results for n  7 do not exhibit
any remarkable changes in the percentage of difference in frequencies compared with
n  5 , considering the efficiency with respect to computational effort as crucial, the
higher number of sample points is not reported in this work. Hence optimum number
of sample points for better prediction of true response is taken as n  5 .

4.2 CASE STUDY 1: EXPERIMENTAL BEAM

The cross section of the RC beam tested in laboratory (Jaishi and Ren 2006) is shown
in Fig. 4.6. The RC beam consists of three 16 mm diameter bars provided on tension
and compression sides with 1.4% of reinforcement ratio. Two legged stirrups of 8 mm
diameter with 200 mm centre to centre is provided as shear reinforcement along the
length of the beam. The beam is with a mass of 750 kg and ρ value of 2500 kg/m3.
Two point symmetric loading has been applied which are at distance of 2 m.

Fig. 4.6 Beam cross section and with two point loading (Jaishi and Ren 2006)

54
 Damages at regular intervals are introduced into the beam by a six-step static
load arrangement (Fig. 4.7). Dynamic measurements are recorded after each loading
intervals to get the dynamic responses of the beam at different damage conditions.
The magnitude of the loads applied during each load step is given in Table 4.3.

Table 4.3 Six step static load magnitude

Load Step No. 1 2 3 4 5 6

Load (kN) 4.0 6.0 12.0 18.0 24.0 26.0

In the test setup, first for simply supported beam a static load is applied at two
locations and after successive damage the static load is removed. Then beam was
rested on flexible springs in order to eliminate the effect of inadequate support
conditions on dynamic characteristics (modal frequencies). A non measured dynamic
input in the form of impulses was created and supplied to the beam resting on flexible
springs. With the said test procedure, initially the modal frequencies are recorded for
simply supported beam without having damage, which was called as initial state of
the beam. Before FE model updating, the responses from the initial state of the beam
is considered to build the reference state FE model. Similar tests are carried out for all
the static load steps and dynamic measurements are done. The dynamic responses
from 5th load step are considered for updating the reference state model in order to
obtain the final damage state FE model. For updating process, the test beam is
analytically modelled with 30 beam elements using commercial FEA software.

Fig. 4.7 Crack pattern at each load steps (Jaishi and Ren 2006)

55
 The modulus of elasticity (E) of 38 GPa and second moment of area (I) of
66×10─4m4 as material and inertial properties are considered in the original FE model
respectively. The E of all the 30 elements of numerical model is used as updating
parameters (N=30). Acceptable bounds are considered for all the parameters and
HDMR response equations are used to develop the objective function. The function is
minimized using GA to obtain the updated values of E. In this case study x1  x30 are
the 30 parameters considered which are the elastic modulus of individual elements of
FE model. The values of the reference points  c1  c30  is taken as 38 GPa for

evaluating the output response of the system. Sample points for n  3 are 22, 38, 54
GPa, and similarly the sample points for n  5 are 22, 30, 38, 46, 54 GPa. The
spacing between the two sample points in an axis is maintained as constant.

50

40

30

20
% Damage

10

-10

-20
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Element location

Fig. 4.8 Stiffness distribution: Reference state model (After updating: n  3 )

The model updating procedure is first implemented for n  3 . The HDMR

response equations are developed for first four modal frequencies. Once HDMR
response equations are developed, the objective function is built up and the same is
optimized using GA. From Fig. 4.8, it is observed that the reference state model has
several damages at different location at the mid span of the beam between element 10,

56
14 and 21 and damage percentage is found to be high with the values of 16.4%,
21.7% and 9% respectively.

Moreover, the damage at support is detected about 16% in element 29 and

which needs further refinement. Hence study is extended to n  5. The SDI based on
HDMR is carried out for n  5 in order to obtain a better reference state model so
that, damage pattern is found out which can be correlate with the experimental
observations. Considering first-order HDMR expression from Eq. (3.9), for
N  30 and n  5 :

f  x     j  xi  f  c1..., ci 1..., xij ..., ci 1..., cN    N  1 f 0

30 5
(4.3)
i 1 j 1

f  x     j  x1  f  x1j , c2 , c3 ,..., c38     j  x2  f  c1 , x2j , c3 ,..., c38 

5 5

j 1 j 1
(4.4)
....    j  x30  f  c1 , c2 ,..., x   30  1 f  c , c ,..., c 
5
j
30 2 3 38
j 1

In the above expression the functions f  c1..., ci 1..., xij ..., ci 1..., cN  are evaluated from

FEA tool and responses of first four bending modes in vertical direction are
considered to develop HDMR expansion. The evaluations of first five component
functions are presented as below.

f  x11 , c2 , c3 ,..., c38   f  22,38,38,...,38 


 21.947 Hz,60.071 Hz,116.670 Hz, 190.578 Hz

f  x12 , c2 , c3 ,..., c38   f  30,38,38,...,38 


 21.946 Hz,60.072 Hz,116.679 Hz, 190.615 Hz

f  x13 , c2 , c3 ,..., c38   f  38,38,38,...,38 


 21.947 Hz,60.073 Hz,116.673 Hz, 190.636 Hz

f  x14 , c2 , c3 ,..., c38   f  46,38,38,...,38 

 21.947 Hz,60.073 Hz,116.673 Hz, 190.650 Hz

f  x15 , c2 , c3 ,..., c38   f  54,38,38,...,38 

 21.947 Hz,60.073 Hz,116.673 Hz, 190.689 Hz

57
 In first order-approximations,  n  1 N  1 function evaluations are required,

hence 131 function evaluations are done to obtain the HDMR expression for four
bending modes. Based on the experimental results, the objective function for the first
four bending frequencies is written as follows:

Fobj  Y1  18.005  Y2  50.204   Y3  98.219   Y4  161.876  (4.5)
2 2 2 2

Initial model is updated to obtain reference state FE model. The stiffness

distribution of reference state FE model is shown in Fig. 4.9. A decrease in stiffness
of 9.7% is observed at element location 24. Also, increase in stiffness has been
observed in some elements of the FE model. For example, the stiffness at element 13
is found to be 14.9%, which is higher compared to all other elements.

50

40

30

20
% Damage

10

-10

-20
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Element location

Fig. 4.9 Stiffness distribution: Reference state model (After updating: n  5 )

Table 4.4 shows the recognized modal parameters for the experimental beam,
the frequencies of the initial FE model and its percentage differences. From the results
it is observed that, initial FE model is in tune with the reference state of the test beam
having variation of percentage frequencies from 0.191 to −0.723. The FE model has to
be updated since the difference in damage state frequencies are found to be high

58
(21.18%). Frequencies from first four bending modes are utilized to construct
objective functions and further optimization using GA.

Further by considering the E values obtained in reference state model, new

lower bound, upper bound and mean [23 GPa, 43 GPa and 33 GPa] for all the 30
updating parameters are fixed to carry out the updating process for second iteration.
Table 4.5 presents the frequencies of the damage state FE model obtained by updating
the reface state FE model using HDMR, and this model is called reference damage
state (Ref.-damage).The stiffness distribution of Ref.-damage state model is shown in
Fig.4.10. From Table 4.5 the percentage difference between experimental values and
updated model in Ref.-damage state is found to be more than 5%, which is slightly
high.

Table 4.4 Frequencies and their differences: Experimental beam  n  5 

Difference in
Mode Experimental Value (Hz) Initial FE Value (Hz)
frequencies (%)

Reference State

1 21.904 21.946 0.191

2 60.329 60.072 −0.425
3 117.022 116.680 −0.292
4 192.026 190.636 −0.723

Damaged State
1 18.005 21.819 21.183
2 50.204 60.285 20.080
3 98.219 116.949 19.069
4 161.876 191.944 18.574

59
 50

40

30
% Damage

20

10

0
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Element location

Fig. 4.10 Stiffness distribution: Ref.-damage state model (After updating: n  5 )

Hence in order to reduce the percentage difference and to obtain the better
correlation between experimental and FE model, the values of E from Ref.-damage
state model is further updated to obtain the modal parameters.

Table 4.5 Frequencies and its differences of Ref.-damage model: Experimental beam ( n  5 )

Differences in
Mode Experimental Value (Hz) After updating (Hz)
frequencies (%)

Reference State

1 21.904 21.819 ─ 0.388

2 60.329 60.285 ─ 0.072
3 117.022 116.949 ─ 0.062
4 192.026 191.944 ─ 0.042

Damaged State
1 18.005 18.980 5.415
2 50.204 51.928 3.433
3 98.219 101.314 3.151
4 161.876 166.357 2.768

60
 The comparison of Table 4.4 and 4.5 shows that a significant improvement in
tuning the natural frequencies is observed except for mode 1 in reference state. And
for other modes, the difference in frequencies are reduced further, where experimental
and FE values are almost found equal with a minimum difference of −0.042% and
maximum difference of −0.388%. From the model updating procedure using HDMR,
accurate modal frequencies are obtained which are having values closer to
experimental damage beam. By updating, percentage differences are found to reduce
from 21.81% to 5.41% in the first mode and also in the higher modes. The detected
damage distribution is shown in Fig. 4.10 without the assumed damage pattern.

New lower bounds, upper bounds and mean [22.8 GPa, 41.8 GPa and 32.3 GPa]
are considered for third iteration for further refinement of responses. After updating
for third iteration, good convergence of frequencies between experimental and FE
model is obtained. Table 4.6 presents the updated modal parameters obtained and
variation in comparison with experimental values. It can be observed that the
percentage difference reduces from 5% to ─1.89% in first mode, and in other three
modes also a reduction in difference is observed as shown in Table 4.6.

Table 4.6 Frequencies and its differences in damage state model: Experimental beam  n  5 

Differences in
Mode Experimental Value (Hz) After updating (Hz)
frequencies (%)

Reference State

1 21.904 21.819 ─0.388

2 60.329 60.285 ─0.072
3 117.022 116.949 ─0.062
4 192.026 191.944 ─0.042

Damaged State
1 18.005 18.347 1.899
2 50.204 50.735 1.057
3 98.219 99.090 0.886
4 161.876 163.32 0.892

61
 50

40

30
% Damage

20

10

0
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Element location

Fig. 4.11 Stiffness distribution: Damage state model (After updating: n  5 )

Figure 4.11
2 shows
4 6the 8damage distribution
10 12 14 16 18 in terms of percentage
20 22 24 26 28reduction
30 in
E for all the 30 elements of the FE model. Reduction in E is found to be more than
30% between the elements 7 and 23.

Further in order to know the effect of increase in number of sample points on

accuracy of the updated FE model, study is extended to n  7 . From the obtained
results no significant reduction of difference in frequencies for damage state model is
found when sample points are increased from n  5 to n  7 . Hence similar damage
pattern is observed as in case of n  5 (Fig. 4.12). Hence five sample points are
considered as optimum for this study.

Tables 4.7 and 4.8 present the comparison between FEMU based on modal
flexibility residual method (Jaishi and Ren 2006) and HDMR. From the results, it is
observed that the HDMR based SDI predicts results with less percentage variation
than modal flexibility in both reference state and damage state. In final damage state
model, the mean error of frequency is found to be 1.18% using HDMR based SDI in
comparison with modal flexibility residual method wherein the error is 4.86%.

62
 Table 4.7 Differences in frequencies of Modal flexibility residual (Jaishi and Ren 2006) and
HDMR before updating: Experimental beam (Reference state)

Differences in Differences in
Mode Experimental Value (Hz) frequencies (%) frequencies (%)
(Jaishi and Ren 2006) HDMR

Reference State
1 21.904 1.410 0.191
2 60.329 1.219 ─0.425
3 117.022 1.898 ─0.292
4 192.026 2.187 ─0.723
Damaged State
1 18.005 21.466 21.183
2 50.204 21.416 20.080
3 98.219 20.361 19.069
4 161.876 19.953 18.574

Table 4.8 Differences in frequencies of Modal flexibility residual (Jaishi and Ren 2006) and
HDMR after updating: Experimental beam (Damaged state)

Differences in Differences in
Mode Experimental Value (Hz) frequencies (%) frequencies (%)
(Jaishi and Ren 2006) HDMR

Reference State

1 21.904 ─0.155 ─0.388

2 60.329 1.039 ─0.072
3 117.022 1.022 ─0.062
4 192.026 1.119 ─0.042
Damaged State
1 18.005 ─1.144 1.899
2 50.204 4.923 1.057
3 98.219 6.845 0.886
4 161.876 6.519 0.892

Figure 4.12 shows the variation of E at different updating states, where

straight line (i.e, horizontal line) indicates the E values are taken as 35.5 GPa
throughout the length of the beam for 30 elements. However in real time situations, it
is not the case, and there is a variation of E in reference state model. Further in

63
iteration 2 after updating the reference state model, values of E will converge, so that,
modal frequencies are in tune with the experimental observations and the variation is
found to be 5% more than the experimental values. Also the updated values of E are
reduced (Ref.-Damage) compared to reference state. In order to reduce the variation
further, iteration 3 is performed, where excellent convergence of modal frequencies
are found. The final variation of E values of damage state model for iteration-3 is
shown in Fig. 4.13. The damage distribution for n  5 and n  7 is found to be
similar with no significant improvement. Hence, with minimum number of sample the
frequency responses from updated FE model are found to be in tune with the
experimental observations.
50
Initial FE
Reference
Ref.Damage
45 Damage
Young's modulus (GPa)

40

35

30

25

0 5 10 15 20 25 30
Element No.

Fig. 4.12 Variation of Young’s modulus at different updating states

In order to know the exact damage pattern due to variation of Young’s modulus
(E), the damaged beam is modelled using FEA software ABAQUS (Hibbitt et al.
2000). Total number of nodes used are 8092, and total number of elements are 5656.
To model the concrete beam, 4800 linear hexahedral elements of type C3D8R
(Continuum 3D 8-node Reduced integration) are used. Reduced integration elements
increase computational efficiency without losing the accuracy. Figure 4.13 shows the
discretisation of FE model into 30 elements in which updated values of E are given.

64
Figure 4.14 shows the reinforcement provided as per the experimental beam.
Reinforcements are modelled using 856 linear line elements of type T3D2 i.e., 2-
noded linear 3-D stress/displacement truss element.

Fig. 4.13 3D Numerical model of experimental beam

Fig. 4.14 Reinforcement bars (Linear line elements of type T3D2)

Figure 4.15 shows the tension cracks generated due to static load step 5 in
experimental case study and damage state FE model. The crack pattern obtained in the
final updated FE model is found to be in well agreement with the experimental
observations, where most of the cracks are under the applied load and no cracks are
found at the supports for both the cases. Hence, SDI based on HDMR can be
effectively utilized in SDI and damage pattern recognition.

65
 Fig. 4.15 Comparison of cracks

4.3 CASE STUDY 2: REINFORCED CONCRETE FRAME

A single storey and single bay laboratory-scale RC frame tested under static and
dynamic conditions (Fang et al. 2008) is considered for further validation of the
proposed methodology. The geometric dimensions, reinforcement layout and in-situ
modal test setup are as shown in Figs.4.16 and 4.17 respectively. Modal parameters of
the intact and damaged frame are obtained by performing dynamic test and are used to
validate the proposed method. The RC frame is modelled using commercial FE
software with two-dimensional beam element. The actual E of concrete is taken as
35.5 GPa and  as 2400 kg/m3 with  of 0.2.

E1 E1
E2 E2

E3 E3

E4
E4

Fig. 4.16 RC frame: Geometric dimensions and accelerometer arrangement (Fang et al. 2008)

66
 Fig. 4.17 Modal test of an RC frame: (a) In-situ modal test with spring boundary condition
(top right corner). (Fang et al. 2008)

100
E1
90 E2
Contirbution to the total Variance (%)

E3
80
E4
70

60

50

40

30

20

10

0
f1 f2 f3 f4 f5
Frequency

Fig. 4.18 Parameter screening results (RC frame)

Before proceeding to further development of HDMR response equations,

parameter screening has been carried out to know the effect of E on response

67
characteristics of four substructures of RC frame. Figure 4.18 shows the contribution
of four substructures on first five frequencies of the RC frame. From the results of
parameter screening, E4 substructure has no contribution on variation of the
frequencies. Since, the Young’s modulus is the only property used as updating
parameter, all the four substructures are considered as for identifying the damage
location. Young’s modulus of the four substructure  E1 , E2 , E3 and E4  is considered

as input parameters and nine modal frequencies are considered as output responses.
The study is carried out for n  3,5 & 7 . The lower bound, upper bound values of E
considered are 30.2 GPa, 40.8 GPa respectively, and 35.5 GPa as mean (reference
point, c in HDMR expansions). The damage identification is performed in three
updating steps. In the first step, the four substructures E1, E2, E3 and E4 having 35.5
GPa each are updated. These updated values are used to build the damage state model.

Table 4.9 Updated values of E for reference state model: Experimental RC frame

Reference State Reference State

Initial State
Substructure (GPa) (GPa)
(GPa)
(n=3) (n=5)

E1 35.5 38.56 31.23

E2 35.5 40.55 40.80
E3 35.5 40.80 40.49
E4 35.5 36.59 39.81

Initially, SDI using HDMR technique is implemented in order to update the

model and to identify the damage location and severity using n  3 . From Table 4.9,
it can be observed that the updated values of E of reference state model has no
significance, since the values of E are higher than 35.5 GPa in all the substructures.

The frequency errors of reference state model is found to be very high i.e.,
16.81% in first mode itself (Table 4.10), which is not acceptable, and a maximum
error of 5.33% in mode 7 is found (Table 4.10). Hence, the possibility of detecting the
damage location is found to be less. Therefore the investigation is further carried out
for n  5 , where an attempt has been made to reduce the percentage error and to

68
detect damage location in order to make FE model in tune with the experimental
damage frame.

By considering the parameter bounds similar to the case for n  3 , the study is
carried out for n  5 . In this case, from the updated values in the first iteration,
reduction in E for substructure one (E1) is observed and the corresponding values of E
for all the substructures are shown in Table 4.9. From Table 4.10 frequency
differences are found to reduce significantly from 16.81% to 3.85% in reference state
with the increase in sample points from n  3 to n  5 respectively. Also from Fig.
4.19, it is clear that, in the reference state model, the absolute errors significantly
reduce except for first frequency.

Table 4.10 Modal frequencies of reference state model: Experimental RC frame

Undamaged
Mode Experimental FEA Initial FEA Ref (Hz) Error (%) FEA Ref (Hz) Error (%)
(Hz) (Hz) n=3 n=3 n=5 n=5

1 30.16 30.51 35.23 16.81 31.32 3.85

2 69.34 65.51 72.64 4.76 70.17 1.19
3 178.11 163.58 173.57 2.54 167.36 6.04
4 339.06 317.01 348.00 2.63 337.44 0.48
5 348.74 322.43 351.14 0.68 339.67 2.60
6 509.78 463.40 494.58 2.98 480.40 5.76
7 709.86 682.30 747.70 5.33 686.30 3.32
8 790.07 750.52 796.76 0.84 768.00 2.79
9 948.13 890.53 965.45 1.82 952.60 0.47

As observed from Table 4.9, the updated values of E1–E4 obtained from
reference state model, it is clear that the substructure E1 is damaged due to reduction
in E. Hence, the reference state model is further updated to find out the exact location
of damage in substructure E1. Further the substructure E1 is subdivided into three
divisions (E11, E12 and E13) and updated for the new bounds (lower bound 25.056
GPa, upper bound 31.32 and mean 28.18 GPa).

69
 Initial State
Reference State
8

Absolute Error (%)

0
1 2 3 4 5 6 7 8 9
Frequency Order

Fig. 4.19 Comparison of absolute errors (%) of initial and reference state model  n  5 

The updated values of three substructures (E11, E12 and E13) are shown in Table
4.11. From the results it is clear that, damage at the substructure E11 is found to be
more due to reduction in E from 31.23 GPa to 26.79 GPa. Hence, the damage is
located at the mid span of the beam element. Finally, the E of four substructure and
frequencies of damage state is shown in Table 4.12 and 4.13 respectively

Table 4.11 Updated values E of E1 substructures: Experimental RC frame  n  5 

Initial State(GPa) Reference State (GPa) Damage State (GPa)

E1 E1 E1’=E11-E12-E13

35.5 31.23 26.79, 29.77, 29.77

Table 4.12 Values of E at different updating states: Experimental RC frame  n  5 

Substructure Initial (GPa) Reference State (GPa) Damage State (GPa)

E1 35.5 31.23 26.79

E2 35.5 40.80 40.80
E3 35.5 40.49 40.49
E4 35.5 39.81 39.81

70
 Table 4.13 Modal frequencies of damage state model: Experimental RC frame  n  5 

Damaged

Mode Experimental FEA Ref Error FEA Dam Error

(Hz) (Hz) (%) (Hz) (%)

1 29.10 31.32 7.64 30.81 5.88

2 68.05 70.17 3.11 70.17 3.11
3 170.24 167.36 1.69 166.82 2.01
4 335.53 337.44 0.57 336.11 0.17
5 344.65 339.67 1.44 339.43 1.52
6 499.53 480.40 3.83 480.48 3.81
7 679.41 686.30 1.01 674.92 0.66
8 768.70 768.00 0.09 763.44 0.68
9 933.98 952.60 1.99 953.50 2.09

From the results of reference state (Table 4.10) and damage state (Table 4.13)
reduction in mean error is observed, i.e., the mean error is found to be 2.94% in
reference state and in case of damage stage model it is 2.21%. Also frequencies of the
damage state model are in tune with the experimental observations. With maximum
error of 5.88% in mode one and 0.17% to 3.81% in remaining modes.

Figure 4.20 shows the absolute percentage variation of reference state and
damage state beam with respect to results of test beam for nine frequencies. It is found
that absolute error is found to reduce in damage state model from 7.64% to 5.88% in
mode 1, and reduction in absolute error is also found in higher modes. Finally, Fig.
4.21 represents the variation of E at different updating states.

The study has been extended for n  7 , to know any further improvement in
updated parameters. From the updated values of E1  E4 it is observed that, there is no
significant improvement in the responses of the damage state model. The mean error
for n  5 and n  7 is found to be the same i.e., 2.2%. Hence the optimum number of
sample points is found to be n  5 for this case study.

71
 Reference State
Damage State
8
Absolute Error (%)

0
1 2 3 4 5 6 7 8 9
Frequency Order

Fig. 4.20 Comparison of absolute error (%) of reference and damage state model ( n  5 )

50

48 Initial State
Reference State
46 Damage State
44
Young's Modulus (GPa)

42

40

38

36

34

32

30

28

E1 E2 E3 E4
Element No.

Fig. 4.21 Variation of E at different updating states

72
4.4 CASE STUDY 3: BRIDGE STRUCTURE

In order to find out the efficiency of HDMR based SDI, a tested (Farrar et al. 1996)
full scale I – 40 bridge constructed over the Rio Grande is considered (Fig. 4.22). In
order to investigate the dynamic behaviour of the bridge, different damage scenarios
are introduced in web and flange portion of a girder. First stage of damage in the form
of two foot web cut at girder mid span, second stage of damage as six foot cut from
the centre of the web to bottom flange at girder mid span, third Stage of damage as six
foot cut in the web plus half bottom flange cut at girder mid-span, and fourth stage of
damage as six foot cut in the web plus full bottom flange cut at girder mid-span (Figs.
4.23 and 4.24).

Fig. 4.22 General view of I – 40 Bridge

(a) (b)

Fig. 4.23 Damage scenarios (a) First stage damage (b) Second stage damage

73
 (a) (b)

Fig. 4.24 Damage scenario (a) Third stage damage (b) Fourth stage damage

The bridge has three continuous spans each of 39.9 m, 49.7 m and 39.9 m (Fig.
4.25). To understand the dynamic behaviour of the bridge, three torsional and three
bending modes are recorded from the experimental investigation. To apply the
concept of HDMR based SDI, the three dimensional model is simplified to two-
dimensional beam model. Hence, only bending modes are considered as the target
values which are obtained from extraction of frequencies due to fourth damage
scenario.

Figure 4.26 shows the simplified two dimensional beam model resting on three
piers, which have a rigid connection at the bottom and rotatable connections between
top of the pier and the girder. The beam model is divided into 9 substructures (thin
girders 7 numbers and thick girders 2 numbers-S3 and S7). The values of E for both
girders are taken as 200 GPa and flange dimensions of thick girder are higher than
thin girders. The section properties considered in the present study are cross sectional
area  A and second moment of area  I  . The initial values of cross section area of

thin girder  A1  and thick girder  A2  are taken as 0.5685 and 0.6436 respectively.

The values of second moment of area of thin section  I1  and thick section  I 2  is

considered as 0.512 m4 and 0.812 m4 respectively. A total of 61 beam elements were

generated to simulate the equivalent girder, and 10 beam elements were used for each
equivalent pier.

74
 Fig. 4.25 Simplified I-40 bridge model with accelerator layout

Fig. 4.26 Two dimensional beam model of bridge

Similar to the first two case studies, a parametric study has been carried out
for n  3,5 & 7 . In this case-study, as damage locations are predetermined in the field
study, first the second moment of area of thin girder is taken as updating parameter
 I1  . Lower bound and upper bound for  I1  is taken as 0.340 m4 and 0.684 m4

respectively with mean value of 0.512 m4 for all the sample points. Frequencies of
initial state model before updating for first three bending modes are presented in
Table 4.14.

75
 Table 4.14 Frequency errors before updating: Bridge example

Experimental Un-damage Frequency (Hz) Frequency

Mode Frequency (Hz) (Initial State) Errors (%)

1 2.48 2.45 −1.2

2 3.50 3.69 5.1

3 4.08 3.90 −4.6

50
8

7
40
6

I reduction (%)
I reduction (%)

5 30

20
3

2
10
1

0 0
1 2 3 4 5 6 7 8 9 S5 (1) S5 (2) S5 (3) S5 (4) S5 (
Substructure Substructure
(a) (b)

Fig. 4.27 Identified damage location at substructure S5  n  3

From Fig. 4.27 it is observed that the damage is detected at substructure S5

which agrees with the experimental value. However, the damage percentage is found
to be less than 2% with mean percentage error 6.01%. Hence the study is extended
using n  5 to obtain the accurate reference state model, with less frequency errors.

From model updating using HDMR, for n  5 , reduction in section inertia of I1

(7.5%) is found at S5 substructure, and hence damage is found to be at S5 location as
shown in Fig. 4.28. Table 4.15 presents percentage difference of frequencies between
predicted and experimental observations in reference and damage state. Also the mean
error is found to be reduced from 6%  n  3 to 5.8% in reference state model for

n  3 . Model is further updated to obtain damage distribution (Fig. 4.29) and

frequency errors of damage stage model is presented in Table 4.15.

76
 50
8

7
40
6

I reduction (%)
I reduction (%)
5 30

20
3

2
10
1

0 0
1 2 3 4 5 6 7 8 9 S5 (1) S5 (2) S5 (
Substructure Sub
(a)
Fig. 4.28 Identified damage location at substructure S5  n  5 

50

40
I reduction (%)

30

20

10

0
3 4 5 6 7 8 9 S5 (1) S5 (2) S5 (3) S5 (4) S5 (5) S5 (6)
Substructure Substructure
(a) (b)
Fig. 4.29 Stiffness distribution at S5 substructure  n  5 

77
 Table 4.15 Frequency errors after updating: Bridge model in damage state  n  5 

Experimental Damage Frequency Errors (%) Frequency Errors (%)

Mode
Frequency (Hz) (Reference state) (Damage State)

1 2.30 6.90 6.88

2 3.49 7.90 3.59

3 3.99 −3.00 −2.57

Finally, by further updating the S5 substructure elements of FE model, the mean

error is found to be reduced to 4.34%. The study is further carried out for n  7 and no
reduction in mean error of percentage difference is found out. Values of parameter
bounds and percentage frequency error of initial state is taken similar to n  5 . From
the results observed it is found that, the mean error of damage state model for
n  5 and n  7 is found to be 4.34% and 4.29% respectively. Hence the values of
frequency are converged for n  5 .

4.5 SUMMARY
The concept of HDMR is applied to three case studies in this chapter. In case-study 1
an attempt has been made to update the initial FE model to final damage state model
so that the dynamic responses obtained are in tune with the experimental
observations. The dynamic responses i.e., frequencies at different modes are in well
agreement with the experimental ones, also the crack pattern in damaged beam
matches with the experimental beam.

In case-study 2, the responses from single bay single story tested under
laboratory conditions (Fang et al. 2008) are considered to build objective function,
where the function is minimized using GA to obtain the updated parameters. In this
case-study, location of damage has been identified using HDMR accurately.

In case-study 3, HDMR based SDI is extended to identify the damage location

in a real bridge structure, where damage location is found with minimum number of
sample points  n  3 . However, the precise location is identified when the number of

78
sample points are increased. Since the variation of experimental frequencies between
the undamaged and damaged states are very less, percentage frequency errors are
found to be slightly high (i.e., 6.8%) in mode 1. And subsequently, the percentage
errors are reduced in other two modes in damage state. Using second-order HDMR,
the errors can be further minimized, with additional computational effort.

79
80
 CHAPTER 5

CONCLUSIONS

5.1 SUMMARY AND RESEARCH FINDINGS

The HDMR techniques are effectively applied for construction of a computational
model directly from laboratory or field data, creating an efficient fully equivalent
operational model to replace an existing time-consuming mathematical model, and for
identification of key model variables, global uncertainty assessments, and efficient
quantitative risk assessments, etc.

In this work, a computationally efficient FEMU method using the HDMR is

applied in conjunction with GA. Initially, HDMR response equations are generated by
considering a simulated beam using three sample points and development of an
objective function for minimum number of sample points is developed. Based on the
accuracy of the results it is concluded that, in order to obtain acceptable true
parameters, response equations in objective functions required should be greater than
the number of parameters utilized. The method has been applied to model updating of
a simply supported beam as well as a RC box culvert. Suitable parameters are first
selected for model updating and numerical simulations are performed using the
combinations of parameters. First order HDMR expressions are used to develop the
response equations. An objective function is built up using the residuals between
measured and predicted responses from the developed HMDR equations, and updated
parameters are obtained using the GA. From the numerical examples presented, the
computational efficiency is studied.

In the proposed work, HDMR based FEMU is carried out for SDI. To locate the
damage, and to observe its pattern, all elements in FE model are considered as
parameters to be updated. First the proposed damage detection procedure was

81
illustrated with an example of simulated simply supported concrete beam, and damage
patterns are identified using HDMR based model updating. The SDI procedure is then
verified with the experimental results available in the literature.

The identified damage distribution obtained from HDMR (without assuming the
damaged pattern of the tested beam) is compared with the test results reported in
literature. The comparison of the damage patterns obtained using the proposed
method with the experimental observations shows well agreement. The minimum and
maximum percentage differences between the experimental values and HDMR in
updating the reference state model are −0.72% and 0.19%, which is much better than
the values obtained using modal flexibility residual that ranges between 1.41% and
2.19%. Similarly, in case of damaged state model after updating, the minimum and
maximum differences in frequencies between the experimental values and HDMR are
found to be −1.899 and −0.892% respectively, and −1.14 % and 6.52 % for the modal
flexibility residual. Hence, it is concluded that the proposed method using HDMR
predicts the frequencies of damaged beam more efficiently than other methods such as
modal flexibility residual method.

The methodology is further applied to RC frame, where damage is located using

substructure based approach, where the frame is divided in to four substructures
considering Young’s modulus as variable. It is found that the damage is occurred with
reduction in Young’s modulus at beam mid-span. The frequencies are found to be in
tune with the experimental observations with percentage error of 5.48% in mode 1
and 0.15% to 3.98% in all other modes.

Finally, the concept of HDMR based SDI is applied to real bridge structures, in
which the frequencies are efficiently computed with less computational effort, with a
mean error of 4.29%. Mean error is converged with the sample points n=7. Also in
the present work, a parametric study is conducted with respect to the number of
sample points used in approximation of HDMR component functions in model
updating, and its effect on absolute error of updating parameters with respect to true
values. It is observed that an improvement in accuracy is witnessed with increase in
number of sample points from three to five. With the increase in sample point from
five to seven, no significant tuning of parameters and responses are identified.

82
Therefore, the optimum number of sample points is taken as five for all the cases.
Hence the proposed FEMU using the HDMR is promising in SDI.

5.2 SUGGESTIONS FOR FUTURE WORK

i. In the present work, first order HDMR expansions are utilized to develop the
response equations. Further, the accuracy can be significantly improved by
employing the second order HDMR, but with slightly increased
computational effort. Finally, the updated FE model can be utilized in
structural damage identification.

ii. In the present study, Lagrange interpolation functions are utilized. Further,
other interpolation techniques like moving least squares can be used and the
accuracy of the parameters and responses of updated FE model can be
investigated.

iii. The present study has been conducted with limited information. However,
this could be extended by incorporating changes in damping, and also the RS
designs could be coupled with neural networks.

iv. The stochastic variation of material parameters can also be considered by

proper uncertainty analysis tools.

v. Different kinds of uncertainties can be characterised and included in the

model based on the real time data.

83
84
 APPENDIX

DEVELOPMENT OF HDMR APPROXIMATION EQUATIONS

USING THREE SAMPLE POINTS (n=3) IN CONJUNCTION
WITH GENETIC ALGORITHM

Case study – 1: SDI of experimental beam

The general form of HDMR expression is given as:

f ij  xi , x j   f ijk  xi , x j , xk  
N
f  x   f 0   f i  xi    
i 1 1i  j  N 1i  j  k  N

 f12N  x1 , x2 , , xN 

Considering first order HDMR expression to develop response equations:

N n
f  x    j (xi ) f (c1..., ci 1..., xij ..., ci 1..., cN )  ( N  1) f 0
i 1 j 1

For N = 30 and n = 3:
30 3
f  x    j (xi ) f (c1..., ci 1..., xij ..., ci 1..., cN )  ( N  1) f 0
i 1 j 1

The expanded form the above expression is given as:

3 3
f  x    j ( x1 ) f ( x1j , c2 , c3 ,..., c38 )   j ( x2 ) f (c1 , x2j , c3 ,..., c38 )
j 1 j 1
3
....   j ( x30 ) f (c1 , c2 ,..., x30j )  (30  1) f (c2 , c3 ,..., c38 )
j 1

For i = 1 and j = 1 – 3
Expansion  1
1 ( x1 )  f ( x11 , c2 , c3 ,..., c30 )  2 ( x1 )  f ( x12 , c2 , c3 ,..., c30 )  3 ( x1 )  f ( x13 , c2 , c3 ,..., c30 )

For i = 2 and j = 1 – 3

85
Expansion  2
1 ( x2 )  f (c1 , x12 , c3 ,..., c30 )  2 ( x2 )  f (c1 , x22 , c3 ,..., c30 )  3 ( x2 )  f (c1 , x23 , c3 ,..., c30 )

Similarly from i = 3 to 30 and j = 1 – 3 expression are given by

Expansion  3
1 ( x3 )  f (c1 , c2 , x31 , c4 ,..., c30 )  2 ( x3 )  f (c1 , c2 , x32 , c4 ,..., c30 )  3 ( x3 )  f (c1 , c2 , x33 , c4 ,..., c30 )

Expansion  4
1 ( x4 )  f (c1 , c2 , c3 , x14 ,..., c30 )  2 ( x4 )  f (c1 , c2 , c3 , x42 ,..., c30 )  3 ( x4 )  f ( c1 , c2 , c3 , x43 ,..., c30 )

Expansion  5
1 ( x5 )  f (c1 , c2 ,...x51 ,..., c30 )  2 ( x5 )  f (c1 , c2 ,...x52 ,..., c30 )  3 ( x5 )  f (c1 , c2 ,...x53 ,..., c30 )
Expansion  6
1 ( x6 )  f (c1 , c2 ,...x61 ,..., c30 )  2 ( x6 )  f (c1 , c2 ,...x62 ,..., c30 )  3 ( x6 )  f (c1 , c2 ,...x63 ,..., c30 )
Expansion  7
1 ( x7 )  f (c1 , c2 ,...x71 ,..., c30 )  2 ( x7 )  f (c1 , c2 ,...x72 ,..., c30 )  3 ( x7 )  f (c1 , c2 ,...x73 ,..., c30 )
Expansion  8
1 ( x8 )  f (c1 , c2 ,...x81 ,..., c30 )  2 ( x8 )  f (c1 , c2 ,...x82 ,..., c30 )  3 ( x8 )  f (c1 , c2 , ...x83 ,..., c30 )
Expansion  9
1 ( x9 )  f (c1 , c2 ,...x91 ,..., c30 )  2 ( x9 )  f (c1 , c2 ,...x92 ,..., c30 )  3 ( x9 )  f (c1 , c2 ,...x93 ,..., c30 )
Expansion  10
1 ( x10 )  f (c1 , c2 ,...x10
1
,..., c30 )  2 ( x10 )  f (c1 , c2 ,...x102 ,..., c30 )  3 ( x10 )  f (c1 , c2 ,...x103 ,..., c30 )
.
.
i  30 : j  1  3
1 ( x30 )  f (c1 , c2 , c3 ...x30
1
)  2 ( x30 )  f (c1 , c2 , c3 ...x30
1
)  3 ( x30 )  f (c1 , c2 , c3 ...x30
1
)

In order to obtain the HDMR expression for the desired response, the functions
f (c1..., ci 1..., xij ..., ci 1..., cN ) are evaluated using FEA package. The responses in Hz
for the function evaluations in the above expansion are as below:

86
Function Evaluations in Expansion  1
f  x11 , c2 , c3 ,..., c30   f (22,38,38,...,38)  21.947, 60.071, 116.670, 190.578
f  x12 , c2 , c3 ,..., c30   f (38,38,38,...,38)  21.947, 60.073,116.684, 190.636
f  x13 , c2 , c3 ,..., c30   f (5438, 38,...,38)  21.947, 60.073 ,116.689, 190.659
Function Evaluations in Expansion  2
f (c1 , x12 , c3 ,..., c30 )  f (38, 22,38,...,38)  21.945, 60.050, 116.537, 190.083
f (c1 , x12 , c3 ,..., c30 )  f (38,38,38,...,38)  21.947, 60.073, 116.684, 190.636
f (c1 , x12 , c3 ,..., c30 )  f (38,54,38,...,38)  21.949, 60.082, 116.743, 190.856

Function Evaluations in Expansion  3

f (c1 , c2 , x31 , c4 ,..., c30 )  f (38,38, 22,38,...,38)  21.940, 59.972, 116.105, 188.737
f (c1 , c2 , x31 , c4 ,..., c30 )  f (38,38,38,38,...,38)  21.947, 60.073, 116.684, 190.636
f (c1 , c2 , x31 , c4 ,..., c30 )  f (38,38,54,38,...,38)  21.949, 60.113, 116.914, 190.389
Function Evaluations in Expansion  4
f (c1 , c2 , c3 , x14 ,..., c30 )  f (38,38,38, 22,...,38)  21.928, 59.799, 115.331, 187.009
f (c1 , c2 , c3 , x42 ,..., c30 )  f (38,38,38,38,...,38)  21.947, 60.073, 116.684, 190.636
f (c1 , c2 , c3 , x43 ,..., c30 )  f (38,38,38,54,...,38)  21.954, 60.182, 117.228, 192.144
Function Evaluations in Expansion  5
f (c1 , c2 ,...x51 ,..., c30 )  f (38,38,...22,...,38)  21.904, 59.528, 114.460, 186.097
f (c1 , c2 ,...x52 ,..., c30 )  f (38,38,...38,...,38)  21.947, 60.073, 116.684, 190.636
f (c1 , c2 ,...x53 ,..., c30 )  f (38,38,...54,...,38)  21.964, 60.293, 117.609, 192.686

Function Evaluations in Expansion  6

f (c1 , c2 ,...x61 ,..., c30 )  f (38,38,...22,...,38)  21.866, 59.195,113.880, 186.612
f (c1 , c2 ,...x62 ,..., c30 )  f (38,38,...38,...,38)  21.947, 60.073,116.684, 190.636
f (c1 , c2 ,...x63 ,..., c30 )  f (38,38,...54,...,38)  21.976, 60.433,117.910, 192.586

Function Evaluations in Expansion  7

f (c1 , c2 ,...x71 ,..., c30 )  f (38,38,...22,...,38)  21.813, 58.872, 113.849, 188.166
f (c1 , c2 ,...x72 ,..., c30 )  f (38,38,...38,...,38)  21.947, 60.073, 116.684, 190.636
f (c1 , c2 ,...x73 ,..., c30 )  f (38,38,...54,...,38)  22.001, 60.577, 117.983, 191.851

Function Evaluations in Expansion  8

f (c1 , c2 ,...x81 ,..., c30 )  f (38,38,...22,...,38)  21.747,58.635,114.372, 189.761
f (c1 , c2 ,...x82 ,..., c30 )  f (38,38,...38,...,38)  21.947,60.073,116.684, 190.636
f (c1 , c2 ,...x83 ,..., c30 )  f (38,38,...54,...,38)  21.028,60.692,117.772, 191.040

87
Similarly, all the function evaluations are carried out up to i = 30 and j = 1 – 3

Function Evaluations in Expansion  30

1
f (c1 , c2 , c3 ,..., x30 )  f (38,38,38,..., 22)  21.947,60.071,116.670, 190.578
2
f (c1 , c2 , c3 ,..., x30 )  f (38,38,38,...,38)  21.947,60.073,116.684, 190.615
3
f (c1 , c2 , c3 ,..., x30 )  f (38,38,38,...,54)  21.947,60.073,116.689, 190.636

The shape/interpolation function  j  xi  is evaluated using the Lagrange

interpolation:

j x  
 x  x   x  x  x  x   x  x 
i
1
i i i
j 1
i i
j 1
i i
n

i
 x  x   x  x  x  x   x  x 
i
j 1
i i
j
i
j 1
i
j
i
j 1
i
j n
i

Considering the first component function

For i = 1 and j = 1 – 3, the expression of Expansion 1 is given by

Expansion  1  1 ( x1 )  f ( x11 , c2 , c3 ,..., c30 )  2 ( x1 )  f ( x12 , c2 , c3 ,..., c30 )
3 ( x1 )  f ( x13 , c2 , c3 ,..., c30 )
 0.002 x12  0.179 x1  4.007  f (22,38,38,...,38)  0.004 x12  0.296 x1
4.640  f (38,38,38,...,38)  0.031x12  1.875 x1  26.125  f (5438,38,...,38)
 0.002 x12  0.179 x1  4.007  21.947  0.004 x12  0.296 x1
4.640  21.947  0.031x12  1.875 x1  26.125  21.947
Hence from the Expansion – 1 first component function in HDMR expansion is given
by:

HDMR Component function 1  0.643x12  38.578x1  559.477

Similarly,

HDMR Component function  2  3.9 106 x22  3.593 104 x2  21.939

HDMR Component function  3  9.756 106 x32  0.001x3  21.922

HDMR Component function  4  2.343104 x42  0.0026x4  21.882

HDMR Component function  5  5.078 105 x52  0.0057x5  21.802

Similarly all the 30 HDMR component functions are evaluated and summation of all
gives the HDMR approximation equation for first response ie natural frequency in
mode 1 Y1  as below:

88
Y1  0.64 x16  38.58 x1  3.90 10 6 x22  3.59 10 4 x2  9.76 10 6 x32  1.02 10 3 x3
1.56 10 4 x72  0.017 x7  9.76 10 6 x28
2
 1.023 10 3 x28  3.90 10 6 x29
2

3.59 10 4 x29  2.59 10 3 x4  5.73 10 3 x5  0.01x6  0.143 x8  0.036 x9
0.0468 x10  0.057 x11  0.066 x12  0.074 x13  0.079 x14  0.082 x15  0.082 x16
0.079 x17  0.074 x18  0.063 x19  0.057 x20  0.047 x21  0.036 x22  0.026 x23
0.134 x24  0.014 x25  5.73 10 3 x26  2.59 10 3 x27  2.34 10 5 x42
5.078 105 x52  1.015 10 4 x62  2.18 103 x82  3.18 10 4 x92  4.08 10 4 x102
4.98 104 x112  5.76 10 4 x122  6.38 10 4 x132  6.83 10 4 x142  7.07 10 4 x152
7.07 104 x162  6.83 10 4 x172  6.83 10 4 x182  5.76 10 4 x192  4.98 10 4 x20
2

4.08 104 x21

2
 3.18 104 x22
2
 2.32 10 4 x23
2
 2.10 10 4 x24
2
 1.58 10 4 x25
2

5.07 105 x26

2
 2.34 10 5 x27
2
 530.32

Similarly, the HDMR approximation equations are developed for the frequency responses of
second, third and fourth mode ie. Y2 , Y3 and Y4 and utilized to develop objective functions.

Then the objective function is minimized using GA to obtain the updated parameters. Using
updated parameters, initial FE model is updated to obtain reference state, reference damage
state and damage state models, so that, responses of FE model were in tune with the
experimental observations.

89
90
 REFERENCES

[1] Abdeljaber, O., Avci, O., Kiranyaz, S., Gabbouj, M. and Inman, D.J. (2017).
“Real-time vibration-based structural damage detection using one-dimensional
convolutional neural networks.” J. Sound Vib., 388, 154–170.
[2] Alis, O.F. and Rabitz, H. (2001). “Efficient implementation of high
dimensional model representations.” J. Math. Chem., 29(2), 127–142.
[3] Arora, V., Singh, S.P. and Kundra, T.K. (2009). “Comparative study of
damped FE model updating methods.” Mech. Syst. Signal Process., 23, 2113–
2129.
[4] Atalla, M.J. and Inman, D.J. (1998). “On model updating using neural
networks.” Mech. Syst. Signal Process., 12(1), 135–161.
[5] Balageas, D., Fritzen, C. and Guemes, A. (2006). “Structural health
monitoring.” ISTE Ltd., London.
[6] Balu, A.S. and Rao, B.N. (2012). “High dimensional model representation
based Formulations for fuzzy finite element analysis of structures.” Finite
Elem. Anal. Des., 50, 217–230.
[7] Balu, A.S. and Rao, B.N. (2013). “Confidence bounds on design variables
using high-dimensional model representation based inverse reliability
analysis.” J. Struct. Eng., 139(6), 985–996.
[8] Balu, A.S. and Rao, B.N. (2014). “Efficient assessment of structural reliability
in presence of random and fuzzy uncertainties.” J. Mech. Des., 136, 051008.
[9] Baruch, I.Y. and Bar-Itzhack (1978). “Optimal weighted orthogonalization of
measured modes.” Am. Inst. Aeronaut. Astronaut. J., 16(4), 346–351.
[10] Berman, A. and Nagy, E.J. (1983). “Improvement of a large analytical model
using test data.” Am. Inst. Aeronaut. Astronaut. J., 21, 1168–1173.
[11] Boulkaibet, I., Mthembu, L., Marwala, T., Friswell, M.I. and Adhikari, S.
(2017). “Finite element model updating using Hamiltonian Monte Carlo
techniques.” Inv. Prob. Sci. Eng., 25 (7), 1042–1070.

91
[12] Brownjohn, J and Xia, P. (2000). “Dynamic assessment of curve cable-stayed
bridge by model updating.” J. Struct. Eng., 126(2), 252–260.
[13] Brownjohn, J.M.W., Moyo, P., Omenzetter, P. and Lu, Y. (2003).
“Assessment of highway bridge upgrading by dynamic testing and finite-
element model updating.” J. Bridge Eng., 8(3), 162–172.
[14] Bucher, I. and Braun, S. (1993). “The structural modification inverse problem:
an exact solution.” Mech. Syst. Signal Process., 7(3), 217–238.
[15] Chang, C.C., Chang, T.Y.P. and Xu, Y.G. (2000). “Adaptive neural networks
for model updating of structures.” Smart Mater. Struct., 9, 59–68.
[16] Ching, J., Muto, M. and Beck, J.L. (2006). “Structural model updating and
health monitoring with incomplete modal data using Gibbs sampler.”
Comput.-Aided Civ. Infrastruct. Eng., 21(4), 242–257.
[17] Chowdhury, R., Rao, B.N. and Prasad, A.M. (2008). “High dimensional model
representation for piece-wise continuous function approximation.” Commun.
Numer. Methods Eng., 24, 1587–1609.
[18] Collins, J.D., Hart, G.C., Haselman, T.K. and Kennedy, B. (1974). “Statistical
identification of structures.” Am. Inst. Aeronaut. Astronaut. J., 12(2) 185–190.
[19] Deng, L. and Cai, C.S. (2010). “Bridge model updating using response surface
method and genetic algorithm.” J. Bridge Eng., 15(5), 553–564.
[20] Dey, S., Mukhopadhyay, T. and Adhikari, S. (2015). “Stochastic free vibration
analysis of angle-ply composite plates – A RS-HDMR approach.” Compos.
Struct., 122, 526–536.
[21] Dey, S., Mukhopadhyay, T. and Adhikari, S. (2017). “Metamodel based high-
fidelity stochastic analysis of composite laminates: a concise review with
critical comparative assessment.” Compos. Struct., 171, 227–250.
[22] Dey, S., Mukhopadhyay, T., Spickenheuer, A., Adhikari, S. and Heinrich, G.
(2016). “Bottom up surrogate based approach for stochastic frequency
response analysis of laminated composite plates.” Compos. Struct., 140, 712–
727.

92
[23] Ding, Z., Yao, R., Li, J. and Lu, Z. (2017). “Structural damage identification
based on modified artificial bee colony algorithm using modal data.” Inv.
Prob. in Sci. Eng., 1–21.
[24] Doebling, S.W., Farrar, C.R. and Prime, M.B. (1998). “Summary review of
vibration-based damage identification methods.” Shock and Vib. Digest, 30,
91–105.
[25] Dutta, A. and Talukdar, S. (2004). “Damage detection in bridges using
accurate modal parameters.” Finite Elem. Anal. Des., 40, 287–304.
[26] Ebrahimian, H., Astroza, R., Conte, J.P. and Callafon, R.A. (2016).
“Nonlinear finite element model updating for damage identification of civil
structures using batch Bayesian estimation.” Mech. Syst. Signal Process., 84,
194–222.
[27] Entezami, H., Shariatmadar and Sarmadi, H. (2017). “Structural damage
detection by a new iterative regularization method and an improved sensitivity
function,” J. Sound Vib., 399, 285–307.
[28] Fang, H., Wang, T.J. and Chen, X. (2011). “Model updating of lattice
structures: a substructure energy approach.” Mech. Syst. Signal Process.,
25(5), 1469–1484.
[29] Fang, S.E. and Perera, R. (2009). “A response surface methodology based
damage identification technique.” Smart Mater. Struct., 18, 1–14.
[30] Fang, S.E. and Perera, R. (2011). “Damage identification by response surface
based model updating using D-optimal design.” Mech. Syst. Signal Process.,
25, 717–733.
[31] Fang, S.E., Pereraa, R. and Roeck, G.D. (2008). “Damage identification of a
reinforced concrete frame by finite element model updating using damage
parameterization.” J. Sound Vib., 313, 544–559.
[32] Farrar, C. R., Baker, W. E., Bell, T. M., Cone, K. M., Darling, T. W., Duffey,
T. A., Eklund, A. and Migliori, A. (1996) “Dynamic characterization and
damage detection in the I-40 bridge over the Rio Grande.” Los Alamos
National Laboratory Report, LA–12767–MS.

93
[33] Friswell, M.I. (2007). “Damage identification using inverse methods.” Philos.
Trans. R. Soc. A., 365, 393–410.
[34] Friswell, M.I. and Mottershead, J.E. (1995). “Finite element model updating in
structural dynamics.” Kluwer Academic Publishers, Dordrecht, The
Netherlands.
[35] Friswell, M.I., Inman, D.J. and Pilkey, D.F. (1998). “The direct updating of
damping and stiffness matrices,” Am. Inst. Aeronaut. Astronaut. J., 36(3),
491–493.
[36] Gautier, G., Mevel, L., Mencik, J., Serra, R. and Döhler, M. (2017). “Variance
analysis for model updating with a finite element based subspace fitting
approach.” Mech. Syst. Signal Process., 91, 142–156.
[37] Grafe, H. (1998). “Model updating of large structural dynamics models using
measured response functions.” Doctoral thesis, Imperial College of Science,
Technology and Medicine University of London, South Kensington.
[38] Hejll, A. (2007). “Civil structural health monitoring - strategies, methods and
applications.” Doctoral Thesis, Luleå Univ. of Tech., Sweden.
[39] Hibbitt, Karlsson and Sorensen (2000). “ABAQUS/Standard user's manual,
version 6.14.” Simulia, Providence, RI.
[40] Imregun, M., Sanliturk K.Y. and Ewins, D. J. (1995a). “Finite element model
updating using frequency response function data - II. Case study on a medium
– size finite element model.” Mech. Syst. Signal Process., 9(2), 203–213.
[41] Imregun, M., Visser, W. J. and Ewins, D. J. (1995b). “Finite element model
updating using frequency response function data - I. Theory and initial
investigation.” Mech. Syst. Signal Process., 9(2), 187–202.
[42] Jaishi B. and Ren, W. (2006). “Damage detection by finite element model
updating using modal flexibility residual.” J. Sound Vib., 290, 369–387.
[43] Jaishi, B. and Ren, W. (2007). “Finite element model updating based on
eigenvalue and strain energy residuals using multi-objective optimization
technique.” Mech. Syst. Signal Process., 21(5), 2295–2317.

94
[44] Jaishi, B. and Ren, W. (2007). “Finite element model updating based on
eigenvalue and strain energy residuals using multiobjective optimization
technique.” Mech. Syst. Signal Process., 21(5), 2295–2317.
[45] Jiang, L. and Li, X. (2015). “Multi-element least square HDMR methods and
their applications for stochastic multiscale model reduction.” J. Comput. Phys.
294, 439–461.
[46] Katafygiotis, L.S., Papadimitriou, C. and Lam, H. (1998). “A probabilistic
approach to structural model updating.” Soil Dyn. Earthquake Eng., 17, 495–
507.
[47] Kaya, H., Kaplan, M. and Saygin, H. (2004). “A recursive algorithm for
finding HDMR terms for sensitivity analysis.” Comput. Phys. Commun., 158,
106–112.
[48] Kenigsbuch, R. and Halevi, Y. (1998). “Model updating in structural
dynamics: A generalized reference basis approach.” Mech. Syst. Signal
Process., 12(1), 75–90.
[49] Levin, R.I. and Lieven, N.A.J. (1998). “Dynamic finite element model
updating using simulated annealing and genetic algorithm.” Mech. Syst. Signal
Process., 12(1), pp. 91–120.
[50] Li, G., Wang, S., Rosenthal, C. and Rabitz, H. (2001). “High dimensional
model representations generated from low dimensional data samples. I. mp-
Cut-HDMR.” J. Math. Chem., 30(1), 1–30
[51] Li, G., Wang, S.W., Rabitz, H. and Wang, S. Jaffé P. (2002), “Global
uncertainty assessments by high dimensional model representations
(HDMR)”. Chem. Eng. Sci., 57(21), 4445–4460.
[52] Li, G., Xing, X., Welsh, W. and Rabitz, H. (2017). “High dimensional model
representation constructed by support vector regression. I. Independent
variables with known probability distributions.” J. Math. Chem., 55(1), 278–
303.
[53] Li, J., Hao, H. and Fan, G. (2016). “Damping ratios identification by
sensitivity-based model updating: Experimental investigation.” Proceedings of

95
 the 24th Australian Conference on the Mechanics of Structures and Materials:
Advancements and Challenges, 1349–1354.
[54] Lin, R.M., Lim, M.K. and Du, H. (1995). “Improved inverse eigen sensitivity
method for structural analytical model updating.” J. Vib. Acou., 117, 192–198.
[55] Link, M. and Weiland, M. (2009). “Damage identification by multi-model
updating in the modal and in the time domain.” Mech. Syst. Signal Process.,
23, 1734–1746.
[56] Liu, J., Sun, X., Han, X., Jiang, C. and Yu, D. (2015). “Dynamic load
identification for stochastic structures based on Gegenbauer polynomial
approximation and regularization method.” Mech. Syst. Signal Process., 56,
35–54.
[57] Marwala (2005). “Finite element model updating using particle swarm
optimization.” Int. J. Eng. Simul., 6(2), 25–30.
[58] Modak, S.V., Kundra, T.K. and Nakra, B.C. (2000). “Model updating using
constrained optimization.” Mech. Res. Commun., 27(5), 543–551.
[59] Modak, S.V., Kundra, T.K. and Nakra, B.C. (2002), “Comparative study of
model updating methods using simulated experimental data.” Compu. Struct.,
80, 437–447.
[60] Modak, S.V., Kundra, T.K. and Nakra, B.C. (2002). “Prediction of dynamic
characteristics using updated finite element models.” J. Sound Vib.,
254(3),447–467.
[61] Mohamed, M. S., Mustafa, H. A. and Ashraf, O. N. (2013) “Finite element
model updating approach to damage identification in beams using particle
swarm optimization.” Eng. Optim., 45(6), 677–696.
[62] Mottershead, J.E. and Friswell M.I. (1993). “Model updating in structural
dynamics: A Survey.” J. Sound Vib., 167(2), 347–375.
[63] Mukhopadhyay, T. (2016). “A multivariate adaptive regression splines based
damage identification methodology for web core composite bridges including
the effect of noise.” J. Sandwich Struct. Mater., doi: 10.1177/
1099636216682533.

96
[64] Mukhopadhyay, T., Chowdhury, R. and Chakrabarti, A. (2016). “Structural
damage identification: a random sampling-high dimensional model
representation approach.” Adv. Struct. Eng., 19(6), 908–927.
[65] Mukhopadhyay, T., Dey, T.K., Chowdhury, R. and Chakrabarti, A. (2015).
“Structural damage identification using response surface-based multi-objective
optimization: a comparative study.” Arab. J. Sci. Eng., 40(4), 1027–1044.
[66] Mukhopadhyaya, T., Dey, T.K., Chowdhury, R., Chakrabarti, A. and
Adhikari, S. (2015). “Optimum design of FRP bridge deck: an efficient RS-
HDMR based approach.” Struct. Multidiscip. Optim., 52, 459–477.
[67] Pacini B.R., Mayes R.L., Owens B.C. and Schultz R.A. (2017). “Nonlinear
finite element model updating, Part I: Experimental techniques and nonlinear
modal model parameter extraction.” Dynamics of Coupled Struct., 4, 263–274.
[68] Park, G., Harley, H., Cudney and Inman, D.J. (2000). “Impedance-based
health monitoring of civil structural components.” J. Infrastruct. Syst., 6(4),
153–160.
[69] Pedram, M., Esfandiari, A. and Khedmati, M.R. (2017). “Damage detection by
a FE model updating method using power spectral density: Numerical and
experimental investigation.” J. Sound Vib., 397, 51–76.
[70] Perera, R., Fang, S.E. and Ruiz, A. (2010). “Application of particle swarm
optimization and genetic algorithms to multi-objective damage identification
inverse problems with modelling errors.” Meccanica, 45, 723–734.
[71] Rabitz, H. and Alis, O.F. (1999). “General foundations of high-dimensional
model representations.” J. Math. Chem., 25, 197–233.
[72] Rabitz, H., Alis, O.F., Shorter, J. and Shim, K. (1999). “Efficient input-output
model representations.” Comput. Phys. Commun., 117, 11–20.
[73] Refsgaard, J.C., Sluijs, J.P., Brown, J. and Keur P. (2006). “A framework for
dealing with uncertainty due to model structure error.” Adv. Water Resour.,
29, 1586–1597.

97
[74] Ren, W.X. and Chen, H. B. (2010). “Finite element model updating in
structural dynamics by using the response surface method.” Eng. Struct., 32,
2455–2465.
[75] Rombach, G.A. (2004). “Finite Element Design of Concrete Structures:
Practical problems and their solutions” Thomas Telford Ltd., London.
[76] Roy, K. (2017). “Structural damage identification using mode shape slope and
curvature.” J. Eng. Mech., 143(9), 04017110.
[77] Salawu, O.S. (1997). “Detection of structural damage through changes in
frequency: a review.” Eng. Struct., 19 (9), 718–723.
[78] Sanayei, A. Khaloo, M. Gul and Catbas, F.N. (2015). “Automated finite
element model updating of a scale bridge model using measured static and
modal test data,” Eng. Struct., 102, 66–79.
[79] Sanayei, M., Imbaro, G.R., McClain, J.A.S. and Brown, L.C. (1997).
“Structural model updating using experimental static measurements.” J.
Struct. Eng., 123 792–798.
[80] Sehgal, S. and Kumar, H. (2015). “Structural dynamic model updating
techniques: a state of the art review.” Arch. Computat. Methods Eng., 23(3),
515–533.
[81] Shabbir, F. and Omenzetter, P. (2016). “Model updating using genetic
algorithms with sequential niche technique.” Eng. Struct., 120, 166–182.
[82] Shahidi, G.S. and Pakzad S.N. (2014). “Generalized response surface model
updating using time domain data.” J. Struct. Eng., 140(8):A4014001.
[83] Sobol, I.M. (2003). “Theorems and examples on high dimensional model
representation.” Reliab. Eng. Sys. Saf., 79, 187–193.
[84] Sohn H., Farrar, C., Hunter, N.F. and Worden, K. (2001). “Structural health
monitoring using statistical pattern recognition techniques.” J. Dyn. Syst.
Meas. Contr., 123, 706–711.
[85] Teughels, A. and Roeck, D.G. (2004). “Structural damage identification of the
highway bridge Z24 by FE model updating.” J. Sound Vib., 278, 589–610.

98
[86] Teughels, A., Maeck, J. and Roeck, D.G. (2002). “Damage assessment by FE
model updating using damage functions.” Comput. Stru., 80, 1869–1879.
[87] Tunga, M.A. (2011). “An approximation method to model multivariate
interpolation problems: Indexing HDMR.” Math. Comput. Modell., 53, 1970–
1982.
[88] Tunga, M.A. and Demiralp, M. (2006). “Hybrid high dimensional model
representation (HHDMR) on the partitioned data.” J. Comput. Appl., Math.,
185, 107–132.
[89] Wang, X., Hu, N., Fukunaga, H. and Yao Z.H. (2001). “Structural damage
identification using static test data and changes in frequencies.” Eng. Struct.,
23, 610–621.
[90] Xiao, X., Xu, Y.L. and Zhu, Q. (2014). “Multiscale modelling and model
updating of a cable-stayed bridge. II: Model updating using modal frequencies
and influence lines.” J. Bridge Eng., 20(10), 1–12.
[91] Zhao, L., Choi, K.K. and Lee, I. (2011). “Metamodeling Method Using
dynamic kriging for design optimization.” Am. Inst. Aeronaut. Astronaut. J.,
49(9), 2034–2046.
[92] Zou, Y., Tong, L. and Steven, G.P. (2000). “Vibration-based model-dependent
damage (delamination) identification and health monitoring for composite
structures – a review.” J. Sound Vib., 230(2), 357–378.

99
 PUBLICATIONS

International Journals
1. Naveen, B.O. and Balu, A.S. (2018). “HDMR based model update in structural
damage identification.” International Journal of Computational Methods, 15(2),
1–14.
2. Naveen, B.O. and Balu, A.S. (2017). “High dimensional model representation
based bridge model update for structural damage identification.” International
Journal of Advances in Engineering Sciences and Applied Mathematics. (Under
review)
3. Naveen, B.O. and Balu, A.S. (2017). “Efficient finite element model update
using HDMR for damage identification in bridges.” Engineering Structures.
(Under review)
4. Naveen, B.O. and Balu, A.S. (2018). “Application of HDMR in assessment of
structural damages.” Advances in Structural Engineering. (Under review)

International Conferences
1. Naveen, B.O. and Balu, A.S. (2016). “HDMR based model updating in
structural damage identification.” Proceedings of Sixth International Congress
on Computational Mechanics and Simulation, June 27–July 1, 2016, IIT
Bombay, India.
2. Naveen, B.O. and Balu, A.S. (2016). “Structural damage identification by
model updating using HDMR.” Structural Engineering Convention, December
21–23, CSIR, Chennai, India.
3. Naveen, B.O. and Balu, A.S. (2017). “HDMR based finite element model
update in structural damage identification.” Seventh International Conference
on Theoretical, Applied, Computational and Experimental Mechanics,
December 28–30, 2017, IIT Kharagpur, India.

100
 CURRICULUM VITAE

1. NAME : NAVEEN, B.O.

2. DATE OF BIRTH : October 13, 1984

3. EDUCATIONAL QUALIFICATIONS
Degree : Bachelor of Engineering (B.E.)
Specialization : Civil Engineering
Institute : B.I.E.T Davangere, Visvesvaraya Technological
University, Belgavai, India
Year : 2006
Degree : Master of Engineering (M.E.)
Specialization : Industrial Structures
Institute : S.J.C.E Mysuru under Visvesvaraya Technological
University, Belagavi, India.
Year : 2009
Degree : Doctor of Philosophy (Ph.D)
Specialization : Structural Engineering
Institute : National Institute of Technology Karnataka, Surathkal,
India
Registration Date : July 17, 2014

4. ADDRESS : S/o B.D.Onkareswara,

#3371/3B, 3rd Cross, M.C.C. B-Block,
Davangere - 577 004, Karnataka, India
Mobile: +91 9611691111/8762097978
Email: [email protected]
____________________________________________________________________

101