1.
i
A TECHNICAL SEMINAR REPORT
ON
STRUCTURAL HEALTH MONITORING OF BRIDGES
USING ARTIFICIAL NEURAL NETWORK.
Submitted in partial fulfillment for the award of the degree
of
BACHELOR OF TECHNOLOGY
in
CIVIL ENGINEERING
Submitted by
J.SREEJA
19B81A01B2
Under the Guidance of
Mr. T.MANOJ
Assistant Professor
Department of Civil Engineering
CVR COLLEGE OF ENGINEERING
ACCREDITED BY NBA, AICTE & Affiliated to JNTUH
Vastunagar, Mangalpalli (V)
Ibrahimpatnam (M), R.R. District, PIN – 501 510

2.
ii
CVR COLLEGE OF ENGINEERING
ACCREDITED BY NBA, AICTE &Affiliated to JNTUH
Vastunagar, Mangalpalli (V), Ibrahimpatnam (M), R.R. District, PIN – 501 510
Web: http://cvr.ac.in, email: info@cvr.ac.in
Department of Civil Engineering
CERTIFICATE
This is to certify that the seminar report titled “STRUCTURAL HEALTH
MONITORING OF BRIDGES USING ARTIFICIAL NEURAL NETWORK. ”
is a bonafide work done as Technical Seminar and submitted by
J.SREEJA 19B81A01B2
in partial fulfillment of requirement for the award of Bachelor of Technology degree
in Civil Engineering, CVR College of Engineering, Ibrahimpatnam. This work has not
been submitted to any university or institution for the award of any degree or diploma.
Technical Seminar Supervisor Head of the Department
Mr. T.MANOJ Dr. T. Muralidhara Rao
Assistant Professor

3.
iii
ACKNOWLEDGEMENT
I owe a great many thanks to a great many people who helped and supported me during
the Technical Seminar work.
I express my earnest gratitude to my technical seminar supervisor, Mr. T.MANOJ,
Assistant Professor, Department of Civil Engineering, for his constant support,
encouragement and guidance. I am grateful for his cooperation and his valuable
suggestions.
I express my thanks to my technical seminar coordinators, Mr. M. Ashok Kumar
(Assistant Professor) and Mr. K. Mahesh (Assistant Professor), Department of Civil
Engineering for the encouragement and support given to me.
I also express my thanks to Dr. T. Muralidhara Rao, Head of the Department of Civil
Engineering for the encouragement and support given to me.
I express my gratitude to Dr. Ram Mohan Reddy, Principal, CVR college of
engineering for the encouragement.
Finally, I express my gratitude to all other members who are involved either directly or
indirectly for the successful completion of this seminar.

4.
iv
DECLARATION
I, the undersigned, declare that the seminar entitled “STRUCTURAL HEALTH
MONITORING OF BRIDGES USING ARTIFICIAL NEURAL NETWORK. ”,
being submitted in partial fulfillment for the award of Bachelor of Technology Degree
in Civil Engineering, affiliated to Jawaharlal Nehru Technological University
Hyderabad, is the work carried out by me.
J.SREEJA
19B81A01B2

5.
v
ABSTRACT
The condition of a bridge is critical in quality evaluations and justifying the
significant costs incurred by maintaining and repairing bridge infrastructures.
Using bridge management systems, the department of transportation in the United
States is currently supervising the construction and renovations of thousands of
bridges. The inability to obtain funding for the current infrastructures, such that
they comply with the requirements identified as part of maintenance, repair, and
rehabilitation (MR&R), makes such bridge management systems critical. Bridge
management systems facilitate decision making about handling bridge
deterioration using an efficient model that accurately predicts bridge condition
ratings. The accuracy of this model can facilitate MR&R planning and is used to
confirm funds allocated to repair and maintain the bridge network management
system. In this study, an artificial neural network (ANN) model is developed to
improve the bridge management system (BMS) by improving the prediction
accuracy of the deterioration of bridge decks, superstructures, and substructures.
A large dataset of historical bridge condition assessment data was used to train and
test the proposed ANN models for the deck, superstructure, and substructure
components, and the accuracy of these models was 90%, 90%, and 89% on the
testing set, respectively.
Keywords: bridge deterioration; bridge condition; artificial neural network;
machine learning; condition prediction.

6.
vi
TABLE OF CONTENTS
CHAPTER 1 INTRODUCTION 1
1.1 STRUCTURAL HEALTH MONITORING 1
1.1.1. PURPOSE OF STRUCTURAL HEALTH MONITORING 1
1.1.2 COMPONENTS 2
1.1.3. STRUCTURAL HEALTH MONITORING TESTING CATEGORIES 5
1.2. ARTIFICIAL NEURAL NETWORK 6
1.2.1 ARTIFICIAL INTELLIGENCE INTO CIVIL ENGINEERING 7
1.2.2 ADVANTAGES OF AI IN CIVIL ENGINEERING 8
CHAPTER 2 LITERATURE REVIEW 9
CHAPTER 3 OBJECTIVES 16
CHAPTER 4 METHODOLOGY 17
4.1 ARCHITECTURE OF NEURAL NETWORK 17
CHAPTER 5 CASE STUDY 19
5.1 DETERMINATION OF OPTIMAL MODEL 19
5.2 CONSTRUCTION OF CONFUSION MATRIX 21
5.3 ANAYSIS OF MODEL 22
CHAPTER 6 CONCLUSIONS 25
CHAPTER 7 REFERENCES 26

7.
1
CHAPTER 1
INTRODUCTION
1.1. Introduction to Structural health monitoring
Structural health monitoring (SHM) has the potential to transform the bridge
engineering industry by providing stakeholders with additional information to inform
decisions about the design, operation, and management of bridges throughout the
structures lifespans. This chapter gives guidance on SHM for engineers who design,
build, operate, and maintain bridges. There remain numerous technical challenges to
overcome when deploying SHM systems; however the most important issues to
consider are how to decide what information is required, and then how to develop a
strategy to deliver this information in a form that is easy to interpret and can inform
decision making. The structural health monitoring process includes installing sensors,
data acquisition, data transfer, and diagnostics through which the structure's safety,
strength, integrity, and performance are monitored. If overloading or any other defects
are observed, proper correction measures are suggested.
Fig 1.1 Introduction to SHM
1.1.1. Purpose of Structural Health Monitoring
1. Improve performance (safety and functionality) of existing structures.

8.
2
2. The placement of sensors during construction works enables observers to assess the
structure's condition and specify its remaining life span.
3. Evaluate the integrity of a structure after earthquakes.
4. Structural monitoring and assessment are essential for on-time and cost-effective
maintenance. So, it reduces construction work and increases maintenance activities.
5. The SHM process collects data on the realistic performance of structures. This data
can help design better structures in the future.
6. Shift towards a performance-based design philosophy.
Fig 1.2 bridge health monitoring
1.1.2. Components of Structural Health Monitoring System
The structural health monitoring system consists of several components which are
presented schematically in figure and discussed below:
i. Structure
The critical structures like bridges, tunnels, dams, and wind turbines are mostly
monitored as they are a vital part of the national infrastructure.
ii. Data Acquisition System
Data acquisition addresses the number and type of sensors, how to activate sensors, and
techniques to save data. The placement of sensors should not alter the behavior of the
structure. This can be achieved by considering the placement of wiring, boxing, etc., at
the design stage.

9.
3
Fig 1.3 components
Sensors need to be appropriate and robust and serve their function adequately for a
specified duration. Each sensor may evaluate a particular aspect of the structure. They
measure strain, deflection, rotation, temperature, corrosion, prestressing, etc.
Several types of sensors are available, such as those provided in Table-1, to be
employed, but fiber optic sensors are the latest ones suitable for infrastructure.
Table-1: Measurements of Structural Response Using Various Sensors
Measurement of
structural aspects
Measuring device
Measuring device
output
Reasons
Loads Load cells
Assess the
magnitude and
distribution of forces
on the structure.
To check whether the
loads are expected and
how they are distributed
on the structure.
Deformation Transducers Deflection
To ensure whether
deflection is within the
tolerable limit or not.
Otherwise, repair and
rehabilitation may be
needed.
Strain Strain gauges
Magnitude and
variation of strains
To check the safety and
integrity of the
structure.

10.
4
Temperature
Thermistors,
thermocouples,
integrated
temperature circuits
Temperature and
temperature
variations
Temperature variation
causes thermal
expansion, and repeated
cycles can damage
structures. It affects
strains.
Acceleration Accelerometer
Acceleration of
structure under
loads, especially in a
seismic prone area.
To examine how the
structure resists
acceleration and
subsequent loads.
Wind speed and
pressure
Anemometer
Measure speed and
pressure at various
locations.
Wind load can govern
the design of long-span
bridges and tall
buildings.
Acoustic emission Microphone
Measure noises and
determine the
location of the noise
using triangulation.
To detect the breakage
of elements in a structure
and determine its
location. It is highly
applicable in prestressed
members and cable-
stayed bridges.
Video monitoring
Internet camera
technology
Record videos of
extreme conditions
of the structure
Record extreme
loadings and detect
overloaded trucks on
bridges.
iii. Data Transfer
The transfer of data can be done through a wire which is common and cost-effective
but may not be practical for large structures. Alternatively, wireless communication can
be considered, which is suitable for large structures, but it is slower and more expensive
than the wired method. Telephone lines are another option to transfer data from site to
the offsite offices. These data transfer techniques eliminate the need to visit the field
for collecting and transferring data.
Fig 1.4 data transfer

11.
5
iv. Digital Processing
After the data is transferred, digital processing is carried out to eliminate unwanted
effects such as noises. It should be carried out before the information is stored. Digital
processing will make the interpretation of data easier, faster, and more accurate.
v. Storage of Data
The processed data can be stored for a long time and retrieved in the future for analysis
and interpretation.
vi. Data Diagnostics
Diagnostic processes involve the conversion of abstract data to useful information
about the structure's condition and its responses to loads. So, the final data obtained
from structural health monitoring should be detailed and physical, based on which
rational and knowledge-based engineering decisions can be made.
The methodology used for the diagnostics process is dependent on the structure type,
location and types of sensors, monitoring purpose, and structural response under
consideration.
1.1.3. Structural Health Monitoring Testing Categories
Testing categories of the structural health monitoring system can be classified as
follows:
Based on a timescale of monitoring:
1. Continuous testing
2. Periodic testing
Based on the manner the response is invoked in the structure:
3. Static load
4. Dynamic load
5. Ambient vibrations
1.1.4. Advantages of Structural Health Monitoring
 Improved understanding of field structural behavior.
 Detect damages at an early age of problem initiation.
 Reduced inspection and repair time.
 Encourage the use of innovative materials.
 Help to develop rational management and maintenance strategies.
1.1.5. Disadvantages
 High installation costs.

12.
6
 Vulnerable to ambient noise corruption.
 Vulnerable to earthquake conditions.
 Challenges with the application of SHM like building accessibility,
manipulating the huge amount of data generated by sensors, environmental
conditions, etc.
 The size and complexity of large structures need a great number of sensing
points to be installed.
1.2. ARTIFICIAL NEURAL NETWORK
Neural networks, also known as artificial neural networks (ANNs) or simulated neural
networks (SNNs), are a subset of machine learning and are at the heart of deep learning
algorithms. Their name and structure are inspired by the human brain, mimicking the
way that biological neurons signal to one another.
Artificial neural networks (ANNs) are comprised of a node layers, containing an input
layer, one or more hidden layers, and an output layer. Each node, or artificial neuron,
connects to another and has an associated weight and threshold. If the output of any
individual node is above the specified threshold value, that node is activated, sending
data to the next layer of the network. Otherwise, no data is passed along to the next
layer of the network.
Fig 1.5 artificial neural networks

13.
7
Neural networks rely on training data to learn and improve their accuracy over time.
However, once these learning algorithms are fine-tuned for accuracy, they are powerful
tools in computer science and artificial intelligence, allowing us to classify and cluster
data at a high velocity. Tasks in speech recognition or image recognition can take
minutes versus hours when compared to the manual identification by human experts.
One of the most well-known neural networks is Google’s search algorithm.
1.2.1 Artificial Intelligence into Civil Engineering
Artificial Intelligence (AI) is a specialized system that can recognize intelligent entities,
make decision-making easier, faster, and more efficient. Artificial intelligence is
concerned with the roboticization of intelligent behaviour that thinks and acts the same
way people do. Artificial intelligence is a broad concept that has become firmly
ingrained in our daily lives. It is built on the collaboration of numerous fields, including
computer science, cybernetics, information theory, psychology, and neurophysiology,
among others. As a result, artificial intelligence is a discipline of science concerned
with the study, design, and implementation of time-saving technologies. AI is
concerned with machines that carry out tasks.
Artificial Intelligence is mostly used in civil engineering applications like construction
management, building materials, hydraulic optimization, geotechnical and
transportation engineering, and is also useful in developing robots and automated
systems. AI models in civil engineering can be used for accurate, cheaper, and less
disruptive construction projects. In modern structures, artificial intelligence is being
utilised to plan the routing of electrical and plumbing systems. Artificial intelligence
(AI) is being used to track real-time interactions between personnel, machinery, and
items on the job site and supervisors for potential safety hazards, construction errors,
and productivity concerns. Simulated intelligence makes it simpler for those who
engage with the development business by making it more sensible. It gives more open
doors in a structural design by making it an appealing field of work.
On a fundamental level, AI is a technology that enables machines to execute tasks
commonly undertaken by humans. In this technology, machines are programmed to
examine data streams to discover new information, and/or perform routine operations.
From a civil engineering perspective, AI can enable civil engineers to uncover new
knowledge hidden in our systems (i.e., in the form of new theories, finding patterns that

14.
8
may help establish hypotheses, and so forth) and/or perform computationally structured
or heuristic processes (e.g., design beams, identify commuter patterns, and so forth). A
key point to remember is that although civil engineers often develop automated
spreadsheets, such spreadsheets are built to follow existing codal provisions or
established prior engineering judgment. On the other hand, automation via AI is not
bound by the availability of existing codal procedures.
1.2.2 Advantages of AI in Civil Engineering
 AI is used to Prevent Cost Overruns.
 AI is used to Reduce the Risk of Accidents.
 AI empowers contractors and project managers to monitor and prioritize risks
on the job site and allows the team to focus on them.
 AI is used for Efficient Project Planning.
 AI is used to Increase Productivity on Job Sites

15.
9
CHAPTER 2
LITERATURE REVIEW
Literatures collected for this present seminar is presented below:
SN
O.
Title of the
paper
Authors Name of the
Publisher
and Year of
Publication
Tests conducted/
Modelling
Resultsand
Discussions
Important
Conclusions
 1.  Developing
Bridge Dete
rioration M
odels Using
an Artificial
Neural Net
work
Essam A
lthaqaf
And
Eddie
Chou

Marinella
Fossetti and
Mi G. Chorz
epa
2022

 ANN models
were developed
to predict the
condition ratings
of three different
bridge
components, i.e.,
the bridge deck,
superstructure,
and substructure.
 TensorFlow and
Keras were used
to implement the
ANN models,
and these models
were trained and
tested on
Google’s
Colab platform.
 the data used to
formulate
the deterioration
models were
sourced from the
National Bridge
Inventory (NBI) d
atabase.

For deck
model, input
layer containi
ng 11
neurons,
7 hidden
layers and an
output layer
1 neuron
obtained
the overall
smallest
MAE
value (0.10)a
nd the
highest
R2 value(0.8
9).
For
superstructur
e
and substruct
ure models
had an input
layer 11
neurons,
six hidden
layers &
output layer
1 neuron obt
ained the
lowest MAE
value(0.10 &
0.11) and the
highest R2
value((0.89
& 0.88)
ANN
architecture
to obtain
optimal
accuracy for
a given
number
of hidden
layers
and neurons.
the accuracy
of the
final model
trained was f
ound to be
greater
than 89% on
the
NBI dataset

16.
10
SN
O.
Title of the
paper
Authors Name of the
Publisher
and Year of
Publication
Tests conducted/
Modelling
Resultsand
Discussions
Important
Conclusions
2.  The
Health Mo
nitoring
of Bridges
using Artif
cial
Neural Net
works.
 Pei-Ling
Liu and
Shyh-
Jang Sun
Journal
of Mechani
cs
2011

The longitudinal e
longations at
the bottom of the
9 zones are used
as the
input features of
the monitoring
system.
The amplitude of
a curve can
by represented by
the maximum or
minimum of the
curve.
Because of
geometrical sym
metry, o. ne
may develop an
ANN system
using the
18 extreme
elongations as the
input vector.
Based on observat
ions,
the 9 damage
zones
are divided into
five groups:
(1L), (1M, 2L),
(1R, 2M,
3L), (2R, 3M) and
(3R).
The extreme
elongations
of
the undamag
ed
bridge, meas
ured in
advance, are
used as a
baseline for
comparison.
For Single
Zone Damag
ed, 163
test cases,
errors of
the output
ratios are
less than 5%.
For Two
Zone Damag
ed, 75 cases,
errors
of the output
ratios are les
s than 10%.
The
longitudinal e
longations of
a
bridge subjec
ted to a
moving truck
can reflect
the damage
condition of
the bridge
The
monitoring
system devel
oped in this
study can
detect the
locations and
levels of
damages accu
rately.
The system
is effective
even in the
case of
multiple
damages
or partial
damage.

17.
11
SNO. Title of the
paper
Authors Name of the
Publisher
and Year of
Publication
Tests conducted/
Modelling
Resultsand
Discussions
Important
Conclusions
 3.  Archetypal
Use of Artif
icial Intellig
ence for Br
idge Struct
ural Monit
oring
Bernardi
no Chiai
a
and Vale
rio De B
iagi

DISEG, De
partment of
Structural
2020

Generation of
a synthetic dataset
then
the training of
learning
tools to determine
whether a structur
e is
damaged and whe
re the
damage is located
.
Evaluation of disp
lacements and vib
ration
frequencies of un
damaged
and damaged stru
ctural
members onto wh
ich a moving
load is acting was
first performed on
a simply supporte
d beam.
A set of
30k simulations w
ith moving load
were performed u
sing Matlab.
The analys
es showed
a large vari
ety of
trends and
dependenci
es on
the accurac
y of the esti
mated frequ
encies and
measured r
otation
A confusio
n
matrix is pl
oted. The
testing data
set was
constituted
by 1455
undamaged
and 1545
damaged be
ams, for a
total of
3000 struct
ures.
A contour
plot of the
error for
the conside
red ranges
of
variables is
made.
the finer the
evaluation
of the freque
ncies, the mo
re
precise the a
mount
of damage.
Tests on
a simply sup
ported beam
subjected to
damage prov
ided interesti
ng insights in
to
the possibilit
y of impleme
nting a hybri
d method.

18.
12
SN
O.
Title of the
paper
Authors Name of the
Publisher
and Year of
Publication
Tests conducted/
Modelling
Resultsand
Discussions
Important
Conclusions
 4.  Estimating
Bridge Dete
rioration
Age
Using Artifi
cial Neural
Networks
 Aseel H
ussein a
nd
Abid Ab
u Tair
Internationa
l Journal
of Engineer
ing and Tec
hnology
Feb 2019

Feedforward Neur
al Network
is selected.
dataset has huge
information about
the 400 bridges
with 60 years of
defect histories
four important
parameters
to design a
network-
Learning Algorith
m, Number
of hidden layers
,Number
of Neurons, Trans
fer Function.
Factorial des
ign of (24) is
used
on ANN
design para
meters with
the defects
age as
the target.
measure the
difference
between targ
ets (true
age) and
outputs
(ANN age)
ten neurons
is better for
ANN
outcome.
The neural
network results
reveal
that most
of the models
had
similar range
in the
similarity
and differences
.
Three
validation
methods were
taken to
measure
the performanc
e of
ANN, Mean
Square Error.
 5. Applicatio
n Researc
h
of Neural
Network
Algorithm
in Bridge
Health Sta
tus Monit
oring Syst
em

 Zhang Q
ingchun
,Lu Ying
hu
,Gan Ha
oyu
,Wang Ji
nya
Internationa
l Conferenc
e on
Robots & In
telligent Sy
stem (ICRI
S)
2019

Sensor nodes
collect
bridge operation
parameters.
The measured
data are
uploaded to
OneNET Internet
of Things cloud
service platform
through GPRS
and EDP protocol
provided
by Internet of
Things
cloud platform
Android APP
The
monitoring s
oftware
is developed
by Python
language.
The interface
is mainly
based
on TabWidg
et tab layout,
UI design
is carried out
through QT
Designer,
and UI files
in
XML format
are generate
d.
PyQtgraph is
used to draw
icons
to make data
viewing mor
e intuitive.
the accuracy is
above 80%, so
80% is set
as the
threshold.
When the
health status is
more than
80%,
the bridge is
judged to
be healthy.
when the
maintenance st
atus is more
than 70%, the
notice is sent
to inform the
staff to
take measures
for maintenanc
e.

19.
13
SN
O.
Title of the
paper
Authors Name of the
Publisher
and Year of
Publication
Tests conducted/
Modelling
Resultsand
Discussions
Important
Conclusions
 6. Artificial
Neural Ne
twork
Model of
Bridge De
terioration

 Ying-
Hua Hua
ng
ASCE
2010

records of 942
decks with 2,300
inspection record
s that have been
treated with
concrete overlay
were studied.
Two null
hypotheses
were tested.
Records of the
condition
of 1,241 decks
with
no maintenance
were used
to identify and
avoid
the influences of
maintenance
on the condition
of decks.
for data set,
any number of
hidden layers
greater than
5 does not
yield
better results.
the average
training classif
ication
rate increases
when
the number of
hidden neuron
s increases
the number of
neurons
increases. The
best
average testing
classification r
ate
obtained, 75.3
9%
study
develops an
ANN predicti
on model for
deck deterior
ation by
adopting BP-
MLP
classifier.
These results
indicate
the potential
of the
ANN model
with BP-
MLP classifie
r as a
prediction too
l.
The 11
factors
identified as
statistically
significant
 7. Artificial
Neural Net
work
Model
for Bridge
Deteriorati
on and
Assessment
Gasser
Galal Al
i,Amr El
sayegh,
Rayan H
. Assaad

 Laval (Greate
r Montreal)
 2019
Software used is
the programming
language, Cran
R,used to train
neural networks
Data Collection fr
om the US.
Department
of Transportation,
Federal Highway
Administration (F
HWA 2018)
properties that are
relevant to the
condition.
Several
NN configurati
ons were
tested,
the residuals
of
those configur
ations were
investigated, a
nd
comparisons w
ith the LM
were performe
d, in order to
select the
best model.
It can be seen
that as the
complexity of
the
NN increases.
investigated s
everal
NN that
output the
condition of
bridge
elements base
d on
selected para
meters.
The develope
d model was
compared
with a LM &
found that
both methods
have almost
the same acc
uracy.
However, the
accuracy of
the NN Good

20.
14
SN
O.
Title of the
paper
Authors Name of the
Publisher
and Year of
Publication
Tests conducted/
Modelling
Resultsand
Discussions
Important
Conclusions
 8. Structural
health mon
itoring
of bridges:
a model-
free ANN-
based appr
oach
to damage
detection
 A. C.
Neves,
I. Gonza
´lez2
, J. Lean
der1
Springer
2017

The first stage of
damage identifica
tion uses
methods which
provide a
qualitative indicat
ion of the
presence of
damage in the
structure.
A ROC curve is a
two-dimensional
graphic in which
the true positive
rate (TPr) is
plotted against
the false positive
rate (FPr) for a
given threshold.
Root Mean
Squared Error
(RMSE) is
one way of
evaluating the
performance
of the
trained &Com
pared to
the simple
Mean Absolut
e Error
the sensors
situated closer
to
the geometric
middle of the
bridge seem to
be associated
with networks
that yield
a better
separation bet
ween
structural state
s
the use of past
recorded deck
accelerations in
the bridge as
input to
an Artificial
Neural
Network that,
after being
properly trained,
is able to
predict forthcomi
ng accelerations
the statistical
evaluation of the
prediction errors
of the network by
means of
a Gaussian
Process,
after which one
can select
the detection
threshold
in regard to a
Damage Index.

21.
15
SN
O.
Title of the
paper
Authors Name of the
Publisher
and Year of
Publication
Tests conducted/
Modelling
Resultsand
Discussions
Important
Conclusions
 9.  Application
of Neural
Networks
in Bridge
Health Pred
iction
based
on Accelera
tion
and Displac
ement
Data Domai
n
Reni
Suryanit
a and
Azlan
Adna

IMECS
2013

The best
performances
of BPNN depend
on the selection of
suitable
initial weight,
learning
rate, momentum,
networks architect
ure model
and activation
function.
The architecture
model for this
system has n
number of input
neurons, one and
two hidden layers
with n neurons
and an output.
output layer is the
level of a bridge
health condition
due to an
earthquake,
to implement the
ANN models,
and these models
were trained and
tested on
Google’s
Colab platform.
The bridge
health syste
m used
several sens
ors
to detect
the behavio
r of
a bridge su
ch
as bridge d
eformation
and damage
The
sensors con
nected to
data logger
and sent
the informa
tion data
such as
displaceme
nt
and acceler
ation to
the local se
rver.
The
architecture
of neural
network me
thod in this
study used
one and
two hidden
layers.
models with
one hidden
layer
for accelerati
on
data domain
had
the discrepan
cy of
the MSE
validation.
The compari
son
of acceleratio
n
and displace
ment
data domain
for one
and two
hidden
layers’ mode
l has
been conclud
ed based
on MSE
mean
value, regres
sion
mean value
and CPU
time of the
network mod
el.
Both compar
isons
show the
MSE mean
value decreas
es since
the epoch
increases.

22.
16
CHAPTER 3
OBJECTIVES
 To measure the extent of structure deterioration.
 To predict structure's failure in advance.
 To apply techniques of neural network in structural health monitoring.
 To analyse the efficiency of ANN model in predicting the failure of structure.
 To enhance the conventional technique of damage detection by introducing ANN techniques.

23.
17
CHAPTER 4
METHODOLOGY
4.1 ARCHITECTURE OF NEURAL NETWORK
An ANN model uses different mathematical layers to learn various features in the data
being processed. Typically, an ANN model comprises thousands to tens of thousands
of manmade neurons arranged in layers as units. The input layer is designed to receive
multiple types of data and extract various features in the data. The extracted information
is then passed to a hidden layer. The feature learning and transformation processes
occur in a hidden layer, and the output layer converts the processed information to its
original format to facilitate easy interpretation. However, due to the exponential
increase in computational power over the last few years, ANN models have evolved to
incorporate multiple hidden layers. In addition, many design optimizations, e.g.,
architectural variations, and preprocessing techniques are being incorporated into these
models. Consequently, these advanced ANN models can produce highly accurate
results; thus, ANN researchers can improve their models by performing a variety of
simulations based on combining a set of hyperparameters, e.g., the total number of
layers (i.e., up to 14 hidden layers) and neurons (i.e., up to 512 neurons), activation
functions, biases, and weights. In this study, TensorFlow and Keras were used to
implement the ANN models, and these models were trained and tested on Google’s
Colab platform, which has the computational power of several graphical processing
units. As a result, the optimal model that can provide the most valuable results can be
realized.
 It is used to determine both the input size and output size.
 The aim is to identify the optimal number of hidden layers and the optimal number of
neurons in each hidden layer.
 This evaluation is conducted after executing various systematic variations to identify the
optimal number of hidden layers and the optimal number of neurons in each hidden layer.
 Here, the goal was to select the ANN model with the optimal architecture.
 The various networks affiliated with the deck, superstructure, and substructure models are
trained and validated, where various nodes in the hidden layers are executed.
 Two factors are considered for determining the architecture of neural network:

24.
18
 The MAE metric was used to determine the prediction accuracy of the models. Essentially,
MAE represents the mean (average) deviation of the predicted values from observation
values.
 The second set of results is related to the coefficient of determination, i.e., the R2 value,
which represents a comparison of the mutual relationship between the predicted and real
values.
 The model with the smallest mean absolute error (MAE) and highest coefficient of
determination (R2 ) simulate to determine the verification set in order to identify which
deck, superstructure, and substructure models demonstrated the best performance.

25.
19
CHAPTER 5
CASE STUDY
In this case study an overpass of Highway A21 “Torino-Piacenza-Brescia” in the Northwest
of Italy is considered. Here, there are several overpasses that are coeval and have the same
structural scheme, details, and techniques. The design of such infrastructures was
performed by Dott. Ing. Gervaso.
5.1 DETERMINATION OF OPTIMAL MODEL
First, a 32-neuron ANN model with a single hidden layer was subjected to the training and
validation processes. Then, this model was modified to include two hidden layers with 16
and 32 neurons, respectively. In the subsequent configuration, the number of neurons in
both hidden layers was increased by 32 and 256 neurons in each layer. Then, a third hidden
layer was introduced, where the number of neurons in the first, second, and third hidden
layers was 16, 32, and 64, respectively. In addition, the model with three hidden layers was
subjected to training and validation, and the resulting number of neurons in the respective
layers was 32, 256, and 512 neurons. Then, a configuration with four hidden layers was
subjected to training and validation with 32, 256, 512, and 512 neurons, respectively.
Fig 1.6 overpass of highway A21

26.
20
In the subsequent step, it is attempted to evaluate the effect of varying the number of layers.
Here, the total number of layers was increased to 5, 6, 7, 8, 9, and 14, where the
combinations of neurons were 32, 256, 512, 512, and 512, respectively. All the models with
hidden layer counts beyond four layers were subjected to training and validation using 512
neurons. The architectures of the different models and their respective validation are
detailed below.
table 1. 2 model architecture
The model with seven hidden layers obtained the smallest MAE value of 0.10 for the
bridge deck model. In another case, the model with six hidden layers obtained MAE
values of 0.10 and 0.11 for the bridge superstructure and substructure models,
respectively
Similar to the MAE analysis, the highest R2 value for the bridge deck model was
obtained by the model with seven hidden layers (0.89). In addition, the highest R2 value
for both the superstructure and substructure models was obtained by the models with
six hidden layers (0.89 and 0.88).
The deck model with an input layer containing 11 neurons, seven hidden layers with 32,
256, 512, 512, 512, 512, and 512 neurons, respectively, and an output layer with one neuron
and a linear activation function obtained the overall smallest MAE value and the highest
R2 value. In addition, the superstructure and substructure models had an input layer
containing 11 neurons, six hidden layers with 32, 256, 512, 512, 512, and 512 neurons,
respectively, and an output layer with a single neuron and a linear activation function
obtained the lowest MAE value and the highest R2 value.

27.
21
A confusion matrix is applied to evaluate the performance of the ANN models in terms of
predicting the condition ratings according to FHWA ratings for bridge conditions (FHWA,
1995) of the different bridge components.
 Essentially, confusion matrices consist of tabular illustrations of MLP predicting
capability
 The models were trained and tested on Google’s Colab platform which has the
computational power of several graphical processing units.
table 1.3 FHWA ratings
5.2 CONSTRUCTION OF CONFUSION MATRIX
The computation of functionality variables for algorithms was performed by evaluating
the-
 true positive (TP-the NN predicts damage on a structure that is really undamaged)
 false positive (FP-the NN predicts damage on a structure that, on the contrary, is
undamaged)
 true negative (TN-the NN predicts no damage on a structure that is really undamaged)
 false negative (FN-the NN predicts no damage on a structure that, on the contrary, is
damaged)
 In addition, accuracy, effectiveness, recall, the F1-score, and the macroaverage are
investigated.

28.
22
Fig 1.7
 The total number of true positives TP value represents the number of correct predicted
values, i.e., the values lying in the diagonal of the confusion matrix.
 The total number of false negatives FN for a class is the sum of values in the corresponding
row (excluding the TP).
 The total number of false positives FP for class is the sum of values in the corresponding
column (excluding the TP).
 The total number of true negatives TN for a certain class will be the sum of all columns
and rows excluding that class's row and column.
Fig 1.8
5.3 ANAYSIS OF MODEL
 The parameters like accuracy, effectiveness, recall, the F1-score are calculated using
following formulas.

29.
23
Confusion matrix for different segments of bridge are:
table 1.4

30.
24
 All three models obtained recall values that were greater than 81%, which implies that if a
particular section of the bridge is characterized by repair issues (TP) and the corresponding
model falsely predicts this (FN), then the bridge may suffer some damage conditions,
particularly where the issue can affect other sections of the bridge.
 The F1-score of all three bridge component models was greater than 85%, which means
that the number of FPs and FNs was low. In addition, it signifies that the model has
predicted both correct and incorrect predictions. From these results, we conclude that both
the recall and precision of the models are significantly high.
 The model obtained the highest accuracy (90%) for both the deck and superstructure
components, and the model obtained 89% accuracy for the substructure component. These
results imply that training was performed accurately with an error rate of approximately
10%.
 According to the findings, the age and type of support have a high impact on the deck
condition. In addition, the age and type of design have a high impact on the condition of
the superstructure. Age and year built have great effects on the condition of the
substructure, whereas the average daily traffic has the least impact on bridge condition.

31.
25
CHAPTER 6
CONCLUSIONS
 ANN modelling techniques in AI can incorporate multiple parameters into a model that is
both sophisticated and nonlinear. In terms of bridge engineering, an ANN model can be
trained and tested on data available in the NBI database in order to predict the deterioration
of a bridge.
 The most appropriate ANN architecture to obtain optimal accuracy for a given number of
hidden layers and neurons.
 Finally, with the optimal ANN configuration, the proposed model is trained and tested for
various components for the dataset.
 Overall, the accuracy of the final model trained for the deck, substructure, and
superstructure components was found to be greater than 89% on the NBI dataset. Similarly,
the F1-score, recall, and precision values were all greater than 81%. Therefore, we consider
that the proposed ANN models have allowed us to identify a strong correlation between the
predicted values and actual values in the NBI dataset.
 Finally, the proposed ANN models were used to develop a bridge deterioration model to
predict deterioration in all bridge systems. Consequently, this model is perceived to assist
in furnishing an ideal plan for bridge maintenance, scheduling any bridge with an imminent
requirement to earmark appropriate funds for its maintenance and repair.

32.
26
CHAPTER 7
REFERENCES
1. American Society of Civil Engineers (ASCE). Infrastructure Report Card: A
Comprehensive Assessment of America’s Infrastructure; ASCE: Reston, VA, USA,
2017.
2. Madanat, S.; Mishalani, R.; Ibrahim, W.H.W. Estimation of infrastructure transition
probabilities from condition rating data. J. Infrastruct. Syst. 1995, 1, 120–125.
3. Thompson, P.D.; Small, E.P.; Johnson, M.; Marshall, A.R. The Pontis bridge
management system. Struct. Eng. Int. 1998, 8, 303–308.
4. Thompson, P.D. Decision Support Analysis in Ontario’s New Bridge Management
System. In Structures 2001: A Structural Engineering Odyssey; ASCE: Reston, VA,
USA, 2001; pp. 1–2.
5. Urs, N.; Manthesh, B.S.; Jayaram, H.; Hegde, M.N. Residual life assessment of
concrete structures-a review. Int. J. Eng. Tech. Res. 2015, 3, iss3.
6. Kayser, J.R.; Nowak, A.S. Reliability of corroded steel girder bridges. Struct. Saf.
1989, 6, 53–63.
7. Adams, T.M.; Sianipar, P.R.M. Project and network level bridge management. In
Transportation Congress, Volumes 1 and 2: Civil Engineers—Key to the World’s
Infrastructure; ASCE: Reston, VA, USA, 1995; pp. 1670–1681.
8. Morcous, G.; Rivard, H.; Hanna, A.M. Modeling bridge deterioration using case-
based reasoning. J. Infrastruct. Syst. 2002, 8, 86–95.
9. Sanders, D.H.; Zhang, Y.J. Bridge deterioration models for states with small bridge
inventories. Transp. Res. Rec. 1994, 1442, 101–109.
10. Jiang, Y.; Saito, M.; Sinha, K.C. Bridge Performance Prediction Model Using the
Markov Chain, no. 1180; NASEM: Washington, DC, USA, 1988.
11. Tolliver, D.; Lu, P. Analysis of bridge deterioration rates: A case study of the
northern plains region. J. Transp. Res. Forum. 2012, 50.
12. Chase, S.B.; Small, E.P.; Nutakor, C. An in-depth analysis of the national bridge
inventory database utilizing data mining, GIS and advanced statistical methods. Transp.
Res. Circ. 1999, 498, 1–17.
13. Morcous, G.; Lounis, Z.; Mirza, M.S. Identification of environmental categories for
Markovian deterioration models of bridge decks. J. Bridg. Eng. 2003, 8, 353–361.

33.
27
14. Abdelkader, E.M.; Zayed, T.; Marzouk, M. A computerized hybrid Bayesian-based
approach for modelling the deterioration of concrete bridge decks. Struct. Infrastruct.
Eng. 2019, 15, 1178–1199.
15. Jiang, Y.; Sinha, K.C. Bridge service life prediction model using the Markov chain.
Transp. Res. Rec. 1989, 1223, 24–30.
16. Lee, Y.; Chang, L.M. Econometric model for predicting deterioration of bridge deck
expansion joints. Transp. Res. Circ. No. E-C049 2003, No.E-C049, 255–265.
17. Madanat, S.M.; Karlaftis, M.G.; McCarthy, P.S. Probabilistic infrastructure
deterioration models with panel data. J. Infrastruct. Syst. 1997, 3, 4–9.
18. Kong, J.S.; Frangopol, D.M. Life-cycle reliability-based maintenance cost
optimization of deteriorating structures with emphasis on bridges. J. Struct. Eng. 2003,
129, 818–828.
19. Saydam, D.; Bocchini, P.; Frangopol, D.M. Time-dependent risk associated with
deterioration of highway bridge networks. Eng. Struct. 2013, 54, 221–233.
20. Frangopol, D.M.; Bocchini, P. Bridge network performance, maintenance and
optimisation under uncertainty: Accomplishments and challenges. Struct. Infrastruct.
Eng. 2012, 8, 341–356.