Documente Academic
Documente Profesional
Documente Cultură
MASTEROF ENGINEERING
JUNE, 1997
ST. JOHN'S
NEWFOUNDLAND
CANADA
National Library
Bibliothque nationale
du Canada
Acquisitions and
Bibiiographic Services
Acquisitions et
services bibliographiques
Canada
Canada
ABSTRACT
Reinforced concrete slabs srhibit complexities in their structural behavior due to the
composite nature of the materid and the multitude and variety of factors that affect such
behavior. As such, current rnethods for the design and anaiysis of reinforceci concrete
slabs are limited in scope and are approximate at best as they must rely on the renilts of
experimental tests, which are both costly and timeansuming to perform. The research
embodied by this document imrestigates the use of a branch of artificid intelligence known
as Neural Networks
Four neural network models are deveIoped to predict the following aspects of the overall
behavior of a concrete SM: 1) loaddeflection behavior, 2) crack pattem at fidure; 3 )
concrete strain distribution; and 4) reinforcing steel strain distribution. Results Eorn
srperimental tests on thirty-four fidi sale siabs are utifued to develop these four models,
hcorporating ail of the parameters that govern thek behavior. The rationale behind and
the details involved are explained for the setup, wrnputer implernentation and selecion of
each optimum neural network model. Resuits show that the neural network technique can
perfonn as a satkhctory altemative to experimental testing or detailed caldations to
provide speedy predictions of all four aspects of the smictural behavior of concrete slabs.
A comprehensive spreadsheet tool is next created to incorporate all four of the optimum
neural networks. The spreadsheet uses readily available software and can be used by
strucRual agineers for instantanmus access to the prediction of any or ail of the four
loading conditions. This tool, combineci with the resuIts for the four neural network
models, demonstrates the powerfiil capabilities and success of neural networks in the
realm of civil and structural engineering in generd and ranforced concrete design in
p d c d a r . This approach could readiIy be expandeci to include the same predictions for
ACKNOWLEDGMENTS
The author wishes to express her appreciation to Dr. H Marzouk, Professor of civil
engineering at Mernorial University of Newfouadland and Dr. T. Kegaq, Assistant
Professor of civ engineering at the University ofWaterloo for their support and guidance
tfiroughout the tam of this research effort. Th&
Finally, the author would like to express her deep gratitude to Tom, Tornmy and Katie for
Contents
Page
..
.. ..-.. - . . II
2.1
Introduction ................................................................
7
2.2
2.3
2.3 .1
History ofN
2.3-2
2.3.3
2.3.4
2.3.5
2.3.6
rks
2.3-7
3-
3 -2
3.3
Mode1 Design.................................~.........................................................39
3 -3-2
3 -3-3
3 .3.4
51
51
..
Training..................................~..............................................................
4.4.2
56
RESULTS........................................................................................................ 6 2
5.1
Introduction ............................................................................................ 6 2
5 -4
5.5
6.1
Introduction .............................................................................. 8
6.2
6.3
6.4
8 2
7.2
7.3
Conclusions............................................................................................. -97
7.2.1
NN 1: Load-Deflection Behavior............................................... 9 7
7.2.2
7.2.1
7.2.1
General Conclusions............~...................................................
IO 1
.................................. 102
List of Figures
2.3
3.1
3.2
3.3
3 -4
3.5
3 -6
4.1
NeurosheU2Soffware............
.......
5.1
5.2
........................... 24
4 6
........................................................ 57
72
86
List of Tables
Characteristics of Cormnody Used N e d Network Paradigms ......................... 19
Description of Input Categorks and Factors ...................................................... 40
Data Extrapolation Methods.............
.
.
.......................................................... 54
Results for Training Load-Deflection Curve Neural Network (MW) ................. 65
Results for Training Crack Pattern Neural Network (NN2) ................................ 69
Resuks for Training Concrete Strain Distribution Neural Network (NN3) ........-74
Resuits for Training Reinforcing Steel Strain Distribution
Neural Network (NN4) .....................................................................................
77
A l .a Input Data for Load-Deflection Neural Network (NNl ) Traimng Cases ......... 110
O
A 1.b Input Data for Lod-Defiection Neural Network (NNI) .Test Cases ............... 111
A2.a Output Data for Load-Deflection N d Network (NN1) .Training Cases --.--.
112
A2.b Output Data for Load-Deflection Neural Network m l ) Test Cases............. 113
O
A3.a Input Data for Crack Pattern Neural Network (NN2) .Training Cases ...--.-......
114
A3.b Input Data for Crack Pattern Neural Network m2).Test Cases ................... 115
Output Data for Crack Pattern Neural Network (NN2) .Training and
s........................................................................................................ 116
Input Data for Concrete Strain Neural Network (NN3a and NN3b) .Training and
Test Cases.. ................ .
.
.
............................................................................ 117
Output Data for Concrete Strain Neural Network (NN3) .Training and
Output Data for Steel Saain Neural Network (NN4) - Traimng and
Test Cases........................................................................................................ 120
B.3
Results for Optimum Concrete Strain Distributioa Neural Network (NN3a) ..... 1%
B.4
Face N d
Results for Optimum Steel Strain Distni'utionNeural Network (NN4) ............. 126
=.
b = perimeter of the slab d c a l section
K.
K,
K.
= siab span
In (SI units)
= Neural Network
= ultimate fIexural Ioad capacity
= actual uitimate punching load
(3
-
= sIab thickness
= column perirneter
= perimeter of c o I m
Chapter 1
Introduction
1.1
Reinforced concrete slabs are used today in a variety of applications including building
floor systems, bridge decks, and oehore oil platforrns. In these apptications, concrete is
selected over other building materiais primarily due to the supezior formability, durabiiity,
ire resistance and insulation capabilities of wncrete. Reinforced concrete, however, is
expensive setups and lengthy periods of time are involveci in the performance of these
tests.
prediction of the structural behavior of remforced concrete slabs. Such predictions could
be u&ed by a stmcturd engineer on a preliminary basis to d e t e r d e the initial suitability
proceed with f.urtfier, more traditional methods of design. This will serve to illustrate 1)
the simple mamer by which neural networks mode1 the impact of a set of parameters
(inputs) on a set of simultaneous conclusions (outputs); and 2) the powemil leam-byexample and generaiization mechanism that neural networks use to detect the hidden
relationships linking the inputs to their outputs Ofegazy et al., 1996).
Neural networks are computationai models that adopt a training mechanism to extract the
relationships that link a set of causal input parameters to th& resulting conclusions. Once
neural networks are train&, they cm predia the results for an unknown case (not used in
training) if provided with the input parameters alone. Some characteristics of neural
networks that make them potmtiafiy useful for many different types of applications are
(Moseihi et al., 1992):
0
They are
fault
tolerant, that is, they are still f'unctiond men after several
They have the ability to simulate the behavior of systems with limiteci modeling
effort; and
O
The main purpose of this thesis is to develop a set of neural networks to predict the
structural behavior of reinforceci concrete slabs. The research is applicable for no&
strength, hi@ strength and lightweight concrete slabs subjected to concentfated and
f i e d Ioads.
1. I d e the detailed fctors that govern the structurai behavior of concrete slabs.
2. Imrestigate the niitability of neural networks for application in the structural analysis
do-
srperimental testS.
5. Deveiop a comprehensive spreadsheet tool for the stnictud analysis of reidorced
concrete siabs,
13 Research Methodology
1. Evaluate the problem by reviewing the theory and current practices in both neural
d network software.
1.4
Thesis Content
discusd
Components of neural networks are defineci, and the various neural network paradigrns
are briefly descnbed. General uses of neural networks in civil and structural engineering
are examineci as weII as their specific uses in the design of reiaforced conmete.
Chapter 3 descrrcbes the devdopment of four neural network models to descnie the
structural behavior of reinforceci concrete slabs.
the partidm problem at hand is discussed as weU as the d o n a l e behind the seleetion of
the backpropagation paradigm
analysis and stniCh<rgig for each neurai network model are then discussed, dong with the
development of alternative models for optimal network seleaioe
Chapter 4 describes the cornputer implementation of the four neural network models.
S p d c s regarding data preparation and software selection are discussed. Details for
aaining and teshg each ofthe four neural network models are then descnbed.
Chapter 5 disnisses the results and observations for all neural network models as weU as
providuig aaalysis of the resuits.
Chapter 6 presents a compreheosive spreadsheet tool that indudes the four neural network
modules for the design and analysis of reinforcd concrete slabs. Development of the
spreadsheet is descnbed and two wnple problems are provided to illustrate the usefulness
ofthe tool.
Chapter 7 is the thesis conclusion and summary. Prospects for frther research as an
extension to the results obtained fiom this thesis are also discussed.
Chapter 2
Literature Review
2.1
Introduction
This chapter evaluates m e n t research efforts in the area of neural network applications
networks and their history is next presented, dong with the various neural network types
applicable to the structurai engineering domah General civil engineering applications of
neural networks are then briefly surveyed. State of the art research descn'bing the use of
neural networks for the sbvcnual behavior of reinforced concrete is then reviewed to
assist in the development of a specific neural network mode1 to predict the structural
behavior of wncrete slabs.
2.2
Traditional research efforts in concrete structural anaiysis that have evolved in the
erature during the past few decades generally aimeci at developing mathematical models
to predict concrete behavior under different loading conditions. These mathematid
models, however, focused generally upon deteminhg the behavior of individual struaural
elements which could not be generalized to describe the behavior of other elements. Also,
the models reque the caldation of severai equations to arrive at predictions for more
than one parameter. Modeling with neural networks is much simpler because, although a
neural network captures the mathematical relationships in its collection of interconnections
between its nodes, no formal mathematical d e s or formula are used or obseniable within
the model (Garrett et al., 1992).
Examples of some mathematical models which are in existence in the literature for
describing the structural behavior of coocrete are descn'bed below. These examples have
been chosen as neural networks have also been developed to model these same behaviors.
The shear behavior of deep beams subjected to point loads can be simuiated by the strutand-tie mode1 (Schliach, 1980), which applies a series of equahions to define the ultirnate
shear forces in the beam. When cornparecl to experirnental test resdts, however, this
mode1 is only accurate when the ratio of shear span to beam height is l e s than 1.04; at
higher values, the model resuits d e c h e rapidly because deep beam behavior no longer
applies (Schliacb 1980). This same m a t i o n applies to alternate models which e i s t in
the iiterature for d u i t h g the shear srength of deep beams (dePaiva and Siess, 1965;
Ramakrishnan and Ananthanarayanana, 1968; Smith and Vantsiotis, 1982; and Subedi,
1988). Other models must theref'iore be applied to predict the structural behavior of
shallower beams.
(given concrete and steel material properties) to predict the concrete shear strength, the
shear distortion, the steel strains and the concrete mains. Although this model and others
(Galletly, 1952; Benjamin and Williams, 1957; Hsu and Mo, 1985) do a reasonably
acwate job in prediaing the previously d e s c n i values, they, iike their cornterpart
models for deep bearns, are Iimited because thqr oniy appiy to low-rise shear walls.
Some research is descnied in the litenihue for mathematical rnodels which predict the
punching shear behavior of reinforceci concrete slabs.
developed that predict the
&kt
concrete slabs subjected to concentrated loads for normal strength (Elstner and
Hognestad, 1956 and Moe, 1961) and high strength (Marzouk and Hussein, 1991)
concrete. The Moe equation is as follows:
where: vu
= column width
= concrete compressive strength
= ultimate flexural Ioad capacity
Marzouk and Hussein (1991)propose that this equation be modifieci to include the cubic
root o f f , when high strength concrete is used. The Elstner and Hognestad equation is:
(3
Kinnunen and Nyiander (1960) also conducteci a theoreticai anaiysis for axiqmmetric
punching shear, by solving a series of eqdiirium and strain compatib%ty eqyations. This
mode1 requres cornputer programming to formulate a solution,and is time-coIlSuming to
&e.
Regan(1980) improved upon this by proposing the foffowing equation for the
where: K.
= KJUc ( p f ~ "ad(&
= constant = 0.13
+ 7.85d)
K,
= 1.15 .[4x
K.
= perimeter of column
(SI units)
The Canadian code (CSA A23.3-94) requires that the d e s t v, resulting fiom the
following three equations be used to determine the fctored shear resistance of a concrete
siab:
where:
a.
bo
= perimeter of critical
It is clear that there is still a wide range of uncertain@ for explaining the punching shear
pariicular slabs used in each experimental testhg prognun Neural networks couid be
used to detect the subtie merences betweea the different types of slabs, thus eliminating
the initial need for lengthy caldations for each model.
1972). These equations are appiied, in matrix form, to descriie a constitutive relationship
in terms of stresses and strains;
concrete. However, the equations are cornplex and are more easily wmputed with the
aid of time consuming serial cornputer dgorithms. Neural network models developed for
the same application (Wu and Ghaboussi, 1992) are much simpler and easier to use.
For all of the abow descncbed mathematical models, several iterations of the following
procedure were necessary (Garrett et ai., 1992):
The mathematical mode1 was then modifieci to account for behaviors observed
but unexplained by the model.
Such a process can be both tedious and time-consuming mtil a niccessfiil mode1 is
developed. Neural networks circumvent this process entireiy as the underlying rationale
for explainhg the behavior of the model is ignored. In addition, the ability of all of the
Extensive testhg on new cases is often not performed and some of the goveming
factors of the concrete behavior, pdcdarly subjective criteria, might be
ornitteci.
it is clear that mathematical models, whiie usually quite accurate for predicting concrete
stnictural behavior,
are limiteci to the extent of the specific application for which they are
developed and can not dways be generalized to apply to those untesteci conditions. In
addition, m a t h d c a l models can be cumbersome and time consuming. Neural network
2.3
Neural networks were first mtroduced as a concept in the early 1950's after Donald
Hebb, a psychologist who studied the effect of learning on the neurons in the brah,
imroduced a simplifieci training mechanism called Hebb's law (Hebb, 1949). This concept
was then extended by Rosenblatt (1958) with the introduction of the perceptron training
algorithm, this became the first mathematid mode1 suitable for cornputer simulation In
accordance with Hebb's law, this procedure viewed biological learning as a dynamic
sensory process which was r d y adapted to cornputer modeling (Hajela and Berke,
1991). Then, in 1969, with the iduential publication by Minsky and Pappert of the book
Perceotrons, all research in neural networks was essentidy haited; the book showed that
a single or double layer perceptron network was inadequate for real world problems
(Caudill and Butler, 1990). It wasn't untii the 1980's that new architectures, such as the
backpropagation training algonthm (see Section 2.3.1 for a description), were imroduced,
and the problems raised in Minsky and Papert's findings were addressed. This gave
engineers ( m g others) reason to explore neural networks as a fast, simple altemative to
Neural networks are types of information processing systerns whose architectures are
inspired by the structure of biologicai neural systems (Caudill and Butler, IWO). UnWce
traditional cornputer programming, which accepts and processes information in a digital
Each network is composed of three basic components as illustrateci in Figure 2.1: 1) input
neurons or processing elements, which represent the input for the problem, 2) connectig
"axons," which connect input and output neurons and represent the connection weights
that associate the input to the output, and 3) output neurons or processing elements,
wbich represent the output for the problem. Neural networks can be composed of a
single Iayer or many layers, accordhg to the complexi~of the architecture of the n e m k
Muiti-Iayer neural networks may contain one or more middie iayers. These middle or
nidden" layers (see Figure 2.1) consist of neuroas with no direct wnnection to either the
input or the output of the network; rather, they are used to fiirther rehe training by
adjusting the comection weights for the network These connection weights are appiied
Input
layer
Hidden
iayer
output
layer
Input
Neurons
output
Neurons
at the links comeaing the inputs to the outputs (axons in Figure 2.1) and they associate
provided dong with the problem description In the case of unsupervised training, on the
other hanci, correct solutions are not provided. Neural networks train& in this marner are
uswlly capable of self-organktion and independent classification of the input data; that
is, the network itseif must decide how t will classq or partition the input data (Caudill
One comrnody used neural network architecture is the Backpropagation neural network
(Rumelhart et al., 1968). Backpropagation networks are training algonthms in which
patterns reco-enized by the network are associated through the layers, and thus the
1)
The input data patem generates a forward flow of activation of the murons nom
the input iayer, through the hidden layers, and nnally to the output Iayer,and
2)
Errors in the output generate a flow of information nom the output layer
backward to the input layer. As the errors are propagated backward, the weights
on the comecting "axons" are adjusted, therefore aiiowing the network to leam.
1982), the BAM (Bidirectional Associative Memory) architecture (Kosko, 1987), and the
ART-2 (Adaptive Resonance Technique) (Carpenter and Grossberg, 1987). Table 2.1
(Moselhi et aI., 1992) Sunmarizes these architectures, dong with their advantages and
disadvantages.
Perqtron
Counterpropagation
Training
tirne*
Slow
Execut ion
titne*
Fast
In formetion
content*
1iigli
Mediiun
Fast
1,ow
Very simple
Mediurn
Fast
1iigh
Slow
Slow
1iigh
Fast
Medium
Kosko
1987
Fast
Carpenter
and
Grossherg
l'ast
Developer
Rumelhart
et al. 1986
Rosenblatt
1961
HechtNielson
1987
machine
Hopfield network
Hinton and
Sejnowski
1986
Hopfield
IAW
Fast
Law
Mcdiiiin
I ligh
1982
HAM
Advantrip
Powerful and accurate association
Siiitable for fitatic environment
(input does not change with time)
Disadvantagea (limitations)
\Jtilirmtion
1Iigh
Could be rappe in Iml minima or
parsMg
Not suitable for real-time applications
No incremerital learning
LQW
Limited rqreseniation capability
Nat as acciunte as brtckpropagation
Could be trapped in local minima or
paraly&
Nol silitable for real-tirne applications
No incrcmental l m i n g
High
Very slow
Not suitable for real-lime applications
No incremental l m i n g
Stlibility not guaranteed
Memary limitations (stored patterns)
Diflicult to Tormalize training methad
Caiild he trapped in lacal minima or
par8l yze
No incremental lecuiiing
Could be rapped in local minima or
paralyze
No incremental l m i n g
Memoty limitations (stord patterns)
iiigh
f Iigh
Low
1 ligh
the backpropagation network therefore reducig the amount of time it takes to train the
network However, the errors produced while testing the network are comparable to those
produced by the backpropagation network (Adeli and Park, 1995). Other applications of
the wmterpropagation neural network in structural engineering couid not be found,
thereby making this architecture an unexplored option for neural network users in
structurai engineering.
The fkq-ARTMAP technology has also been used to predict plain concrete matenal
swngth This form of neural network architecture is a hybnd network that performs
incremental, unsupervised learning of recognition categones and can perfonn a
limited.
In addition, because this type of network maps many dimensions into one
dimensios it wodd work more successniny with problems that containeci a great amount
of input variables and only one or two output variables (Kasperkiewicz et al., 1995).
engineering. This is primarily because backpropagation neural networks are still the most
simple form of neural network architecture. They also appear to be the most capable of
leaniing the association between input and output patterns under a static environment
given adequate training examples (MoseIhi et al., 1992). Furthemore, most problems in
civil and structurat engineering involve the sort of predictions for which backpropagation
networks are best suited.
During the past few years, the area of structural analysis has exhiibited an increasing use of
neural networks for a wide range of applications. Some of these uiclude the modeling of
initial design processes, the modeling of plain concrete materiai strength, and the modelng
of reidorced concrete structura! behavior.
Most civil engineering systems are wmplex and are subject to a wide variety of internai
and extenial forces (e-g., wave forces, weather conditions, seismic loadq and material
mechanics). Andyzing such systems has been a nifficult task and traditional tmls that
acairately predict and model the behavior of nich systems are limited in scope. This is the
main reason that Artificial Intelligence techniques have increasingiy been experimented
with in the civil engineering domain. Among these took, Neural Networks (NNs) have
been reported as efficient pattern recognition and classification tools that model the causeeffect relatioaships of a particula. systern or problem without exploring the underlying
rationale used to mode1 the behavion (Hegazy et al., 1996). Correspondingly, the
useflness of neural networks as tools for design and decision support in civil engineering
is wefl documented throughout the merature (e-g, Moselhi et al., 1992). Figure 2.3
sunmarizes examples found in the literature of applications of nemai networks within civil
engineering in the general realrq in the construction r e a h and in the structural analysk
reah
Examples found in the Literature of generai applications of neural networks withi civil
for equipment production estimation and construction trade productivity level estimation
(Moseihi et al., 1992), as well as the assessment of construction risks in the bidding
process (HegazyJ993). Aithough none of these examples are direcfjy related to the
STRUCTURAL
ENGINEERING
curent investigation for predictig the behavior of concrete slabs with neural networks,
Neural networks are most suitable for applications that have the foflowins features:
problem; and
O
Many examples o f the problem are avaiiable for accurate training of a neural
network-
Many problems, including those containeci within structurai analysis, meet the above
neural networks as an aid for both the design and analysis of structures. The first
prototype application of neural networks as a tool for stmcturai design was proposed by
Vanluchene and Sun in 1990. The study demonstrateci, through the use of three examples
realm of structural design. Since then, neural networks have been applied to nearly every
facet of stnictwal engineering.
categories: 1) models of the stmctural design process (e-g-, Hajela and Berke, 1991;
Messner et ai,
plain concrete materiai strength (e-g., Wlams et al., 1992, and Kasperkiewicz et ai.,
1995) and 3) models for predicting reinforceci concrete material behavior (e.g., Wu and
f'unctional-link net, which essentidy involves non-linear transformation of the input data
used in a flat, single-layered network The report shows only limited success with the
latter architecture, with errors reduced to the range of 3.9% to 5.2% and marginai
increase in leaming speed. The multilayered architecture performed better, with mors in
the range of 1% to 2%. Because this problem is similar in nahue to the problem at hand,
network architectures is useW for se~ectingthe appropriate architecture for the models of
concrete materid properties.
A second example within the nrst category describes the deveiopment of a cornputer
application to select the most effective structural system for a building @en
attributes
regarding the project Sue, budget, etc. (Messner et ai., 1994). This paper explores the
rationale for choosing a neural network model over a de-based expert system model
(another fonn of artincial intelligence). The uitimate reaMn for choosing a neural network
is because of the rnany interrelations between the different project parameters and the lack
of explicit causahies between these parameters (Messner et al., 1994). This cm be
directly related to the current research for determinhg the structurai behavior of concrete
slabs because the input data for these neural networks wnsists of many different
properties with respect to the concrete slabs whose interrelationships have not exactiy
been determined.
Another example involves the modeling of initiai design processes using neural networks.
This example (Muktierhjee and Deshpande, 1995) uses traditionally selected design
criteria as input and uses the neural network to determine the size (Le., depth and width),
r d o r c i n g steel area, c o s t h and moment capacity of a reinforcd wncrete beam. Unlike
the m e n t investigation, however, this mode1 uses mathematicaily generated data to train
the network, as initial design processes are readiiy rnodeled using more traditional
of damaging comectioa links on the desired output for the neural network; it is found that
as many as two nodes could be damageci with little &ect on the overd performance of
2.3.5
From a review of the fiteranire it is found that a iirnited number of studies have been
conducteci on the use of neural networks for predicting concrete strength. The fist
example of nich research (Wiams et al., 1992) utilizes the same neural network
deveiopment software (Neurosheii) as is proposed in the current investigation to
determine the structural behavior of concrete slabs. The model utilizes data regarding one
&y, three day and seven day compression strengths as inputs for the mode1 to determine
the twenty-eight day compressive strength as output. The midy observes reasonable
performance of the neural network as cornpareci to hear regression analysis. It also
determines that, with the M e d data used to train the model the performance of the
network appears to improve with the addition of input variables to the model; five
different models are trained with an increasbg number of input variables, and the accuracy
of the network irnproves with the addition of each input variable.
The data f?om this report is srtremeiy usefid for the current investigation as it
demonstrates the suitabiiity of NeurosheU software for a model sirnilar to the proposed
model.
However, the model desmied in the report only addresses plain concrete
A more recent shidy addresses the same problem, Le., predidon of concrete strength,
however a greater niimber of different variables are selected to model the input for the
problem. Moreover, a ciBiirent n e d network architecture with a different leanillig
train the
The tirst investigation regarding the feasibiliity of uMg neural networks to mode1
Networks in Civit Engineering), a neurai network software package which employs the
back-propagation training algorithm. This study used as input oniy a limiteci number of
variables to demie the concrete behavior (bending moment applied, reinforcing steel
strength, concrete compressive strength and reinforcing steel ratio) to anive at an ideal
depth for the beam (the only output for the network). Training and testing was conducted
using randody chosen patterns obtained from conventional mathematical formulas rather
than data obtained fiom experimentaf results. While limitai in its scope, this study was the
initial impetus for the use of neural networks in concrete design.
More recently, a study was conducted to analyze fiameci shearwail behavior ushg neural
networks (Mo and Lin, 1994). Again, only limited data was utilized as input to describe
the concrete materid behavior (concrete compressive strength, steel yield stress,
longitudinal steeI ratio and shear main); the only output parameter was shear stress. Two
study groups were used for aaining and testing the network; one study group included
results fiom srperimental tests whiie the other study group included results nom
calculations of the miss model theory (descn'bed in Section 2.2). Models for both study
groups performed well. The papa mggesteci that the methods used couid be apptied to
the behavior of other concrete structures. A h , it found that the effect of the transfer
fiuictions and leaniins niles on the network is sigdicant while the effect of the number of
processing ekments in the hidden layers on network learning is uisi&cant
1994).
The feasibility of using neural networks to evaluate the dtimrrte strength of deep
reinforceci concrete beams in shear has also been investigated (Goh, 1995). Again, both
experimentai data and data obtauied &orn mathematicai calcdations were avaiiable for
training and testing the network The study showed that, when compared to conventional
methods (the strut-and-tie mode1 discussed in Section 2.2) for predicting the ultimate
strength, the newd network approach was actuaily more reliable.
The major t h s t of their research has been aimed toward the deveiopment of proper
constitutive relationships for finite element modehg of the material (Ghaboussi et ai.,
199 1; Wu and Ghaboussi, 1992; SankaraSzTbramanian and Rajasekaran, 1996). Because
wncrete is a difficult material to modei f?om a finite element perspective, neural networks
Ghaboussi, Garrett and Wu originally studied the use of neural networks to predict the
stress and main behavior of plain concrete (Ghaboussi et al., 1991) and laer extended tins
study to include the stress and main behavior of reinforceci concrete (Wu and Ghaboussi,
1992). In the models for piah concrete, the input for the networks included stress and
saain increments, and the output included either stress or strain increments, dependig on
The rnodels for reinforcd concrete included aii pertinent data to descnbe the behavior of
the concrete. The data included the concrete compressive strength and strain and the
cracking strength of the concrete; this information was implicitly included in the stressstrain materiai variables through nonnalization on the principal compressive and tensile
stresses and p ~ c i p l ecompressive snauls (Wu and Ghaboussi, 1992). Also included in
the input data was information regarding the reinforcing steel such as yield stress and
reinforcement ratio. F i y , the stress and strain States for two stress increments were
included in the input data.
included current stress or shain increment. Ail neural network models were found to
perfonn S a t i ~ o d y ,i-e, they were able to predict stress and strain nates with
reasonable accuiacy.
Although the outwme of the Gluboussi, Garrett and Wu models was then utilized in a
&te
element model for concrete, hese rationale behind the formulation of the neural
network models for these studies was similar to the m e n t investigation. Therefore, the
content and methods of these studies wilI be very useful for the current investigation.
However, these studies do not directiy wnflict with the merit investigation as they were
compieted for a diEerent purpose.
experimenfal data f?om tests conduded on reinforced concrete panels subjected to inplane shear i contrast to the reinforced concrete siabs subjected to concentrateci and
flexurai loads used in the present investigation Therefore, the idonnation sought in the
current investigation is outside of the training domain for these studies.
materiai properties for input. This study is usefbl as it again shows the niccess of neural
Neural network research for concrete slabs has focused on the initial structural desip of
these slabs. The neural network developed by Arslan and Ince (1994), for example, takes
the moment and slab support conditions as inputs to produce ouiy the moment coefficients
2.3.7
Limited research has been conducteci in using neural networks to model cracks in general.
One application studied the detection and mapping of cracks in eggs (Patel et ai., 1994).
This investigation used cornputer vision to model the picture of a cracked egg usng a grid
of pixels.
predicted whether the egg was cracked or not; the study did not address the prediction of
the a
d pattern of the crack Like one of the crack pattern neural networks in the
current investigation (NNZa), however, the success or M u r e of this mode1 was based on a
2.4
Summary
In the present study, the use of neural networks in prediaing the structural behavior of
concrete in slabs is experimented with for severai reasons, includig:
Neural network approximations are equally as accurate as other complex mathematical
An adequate number of trainhg cases will be used to train the network as experimental
results on M-scaie slabs were monitored since 1990 (Marzouk and Hussein, 1991).
34
This chapter has reviewed the previous work related to neural networks in civil
engineering and stmchiral andysis. While it is apparent that a large amount and variety of
applications of neural networks exists in these fields, there is no single application that has
been used to d e t e d e the stmctural behavior of wncrete slabs in particular- However,
all of the previous works d e s c r provide signifiant insight h o the development and
modehg of a neural network for the curent investigation of the structural behavior of
concrete slabs,
Chapter 3
3.1
Introduction
2) final crack pattern formation; 3) reinforcing steel srrain distribution at slab failure; and
4) concrete strain dimibution at slab fidure. Each aspect lends &self to a neurai network,
t h d o r e , four separate neural network models have been developed to predict these
aspects.
phases of the mode1 development, whe chapter 4 wi focus on the nnal phase, that k, the
The concept stage, as shown in Figure 3.1, includes two steps that involve selecting the
application then the paradigrn for the neural network model. The fint step involves
choosing an application which is amenable to neural network modeiing. AU four of the
proposed n e d network modules encompassed by the current sudy were easily
converted to neural networks as the inputs and outputs for each module were readily
tactors that describe the slab, it wuid be said that the behavior of a concrete slab is
patterned according to its makeup. Therefore, the four neural network applications that
were selected for the curent study are primarily pattern recognition problems. For the
second step of the concept phase, the Backpropagation paradigm was selected as the
neural network type suitable for modehg the applications. The p ~ c i p a lreason this
*!
.
m
e
sm
.
.
8
s
sa
.
I
e
e
i
a
s
m
m
m
m
Ie
DESIGN
Figure 3.1. Neural Network Development Methodology
used in existing structural analysis applications due to its simplicity and its suitability for
pattern-recognition problems.
b
d
,is the identification and use of the independent (non-correlated) factors that M y
descnbe the slab. Problem structuring, on the other han& entails the representation of
sucb descriptive factors dong with th& associated r d t in the fom of inputs and
outputs, as required by the modehg of each individual neural network. Identification of
the input data was conducteci simuitaneously for all four neural n e ~ o r k s(NNs) as the
same data was used to desmie all of the reidorced concrete slabs in the study. The
Won which could descnie the physicai properties of a reinforced concrete slab were
f
h
t grouped into four main categories: 1) slab geometrical dimensions; 2) aggregate
properties; 3) concrete properties; and 4) reinforcement-steel properties. The inputs for
ail four NNs were then readily denned from these four categories; boundary and loading
conditions for each slab were also added to the input descriptions. These resuited in a
total of nineteen input &on
Roblem analysis was required on an individual basis for each of the four NNs in order to
determine their outputs. The outputs for al1 four models were obviously Meremf as each
SI.bGcomttric
PmpertScs
-
MFWb
Pmpertics
Conmete
Propertia
atmforcement
Std
Properties
Loadmg&
Conditio~ts
1.
2.
3.
4.
5.
6.
7.
mode1 was designed to produce separate yet complementary results. In addition, different
ways of problem stnictu~g(the second W o f model design) were experimented with for
each NN in order to achieve the optimum network to solve the problem The resulting
structure for each NN is desc~ibedin the following paragraphs, with the results and
corresponding optimum structure describeci in Chapter 5.
The loaddeflection neural network model was the first mode1 tested in the m e n t study;
the vdidity of the neural network technique was tested in this original neuraI network.
For this reason, two separate neural network models were experimented with in detail in
order to arrive at the opMial configuration of the outputs for this model. As show in
Figure 3 -2,the number of inputs and outputs in each network are as foilows:
Outputs
inputs
XI. Aggregate Type (I=Sandstone; 2-Granite)
Neural Network 1:
01, ikfl. ai lO%Uit. h d (mm)
0 2 . Defl. al 20% UIt, h d (mm)
03, Dcfl, at 30% Uii. Load (mm)
Neural Network 2:
01. Yicld Load (KN)
0 2 . Deflection at Yield (mm)
0 3 . Ultimatc h a d (KN)
W.Defiection at Ult. Load (mm)
The final configurations for these two models are as shown in Figure 3.2.
As previously descn'bed in Section 3.3, the inputs for this neural network model consisted
descnbe the final crack pattern for a reinforced concrete slab, howeveq was a diflicult
tasic; while recognizbg pattems of images is an intuitive and simple task for humans, it is a
cornplex task for cornputers and requires specialized AI-based rnodehg. Despite the
modehg dificulties, a concrete siab's crack pattern provides insight into the failure
mechanism of the slab and its rate of deterioration under loading. The outputs for this
neural network could be modeled in a variety of ways. These varied fkom exact ddailing
detailed model, however, was expected to involve a large size neural network, thus
requirllig a larger number of training cases than were a d a b l e nom the experimental
testing. Three xhematic models were then proposed as s h o w in Figure 3.3, and the
43
advantages and disadvantages of each rnodel were reviewed. mer thorough anaiysis and
initial experimentation with the three types of models, the third approach was seleaed due
to its simple representation and its appropriate proportion of outputs to inputs. A fourth
model, which was desgned with a less subjective approach to produce quantitative results
for the extent of radiai and tangential crackuig, was aiso selected for cornparison with the
r d t s for the schematic model chosen. The final outputs for the two neural network
distniution of the strains throughout the siab, that k, if the measured concrete strains are
Therefore,
representaive values for the concrete strain at the edge of the slab, at a midpoint of the
siab and at the cofumn fkce at M u r e of the siab would indicate the extent of the cracks
throughout the slab. These values were easiiy converted to outputs for the fkst neuraI
network model for predicting the distriiution of the maximum concrete strains throughout
the slab. Again, aii nineteen inputs describeci in Section 3.3 were used in this model.
Orighally, prediction of both tangential and radial strain distniution was proposeci,
however, ody tangential strains were measured dong a radius for most of the tested slabs.
45
02- Crack S
m (O=essymeical;-1
about horiz axiq
2=sym about vert a~& 3= M y syrnmecai)
Nend Nmork 2:
0 1 Radms &Tangentid Cracking (mm)
0 2 Extent of Radial Cracking ( h o n e ; 1 = inuer tbird; 2 =
outer third)
Thus, one neural network model was developed to predict only the tangentid strain
distriution
A good indicator of the concentration of stresses at the face of the stub column wouid be
the measured wncrete strains at various loads at the column face. This wodd provide
information regarding any M u r e that couid occur at the slab-column cornecfion, since
this point is the most stressed point on the concrete slab. So, another neural network
model was developed to predict concrete strains at various load increments at the column
face. The final outputs for both concrete strain rnodels are shown in Figure 3. 5.
A fourth group of neural network models was developed to predict the distriiiution of
designed to provide information regarding the extent of yielding of the reinforcing steel.
This information is usefiil because, d e n the reibforchg steel yields, the fii tende and
compressive loads are camed by the concrete alone; M u r e of the slab wodd probably
to predict the radius of yield for the reinforcing steel ody as weil as the distniution of the
maximum strains in the reinforcing steel dong a radius through the slab. The r d t i n g
Outputs
Nealtri Network 1:
01- Tangeatial Concrete Strain at Edge Gauge
02-Tangenthi Cuncretg Strain at Mid-Siab Gauge
03-Tangentid Concrete Strain at CoInmn Face
Neural Network 2:
01- Concrete Strain at CoIumn Face at 25% Uitimate Loai
02-Concree Strain at Cdrmui Face at 50% Ultimate Loac
0 3 - Conmete Strain at Column Face at 75% Ultimate L m
04- ConStrain at Column Face at Ultimate Load
05'UltimateLoad,kN
Outputs
O1 Yield radius
02-Steel Strain at edge main gauge
03- Steel strain at mid-slab strain gauge
04- Steel strain at column face strain muge
3.4
Summary
Developrnent of a reliable neural network model is essential for the proper predictioa of
results for a problem Therefore, the appropriate selection of inputs and outputs for each
NN module proposed for the prediction of the structural behavior of reinforceci concrete
slabs must be conducteci prior to implementation of the model. This chapter discusses the
selection of these inputs and outputs dong with the reasons for heir selection. Chapter 4
will then discuss the bnplementation of these models, dong with the assignment of actuai
Chapter 4
Cornputer Implementation of the Neural
Network Models
4.1
Introduction
The implementation phase of neural network model development is comprised of two main
tasks: 1) training data preparation; and 2) training and testing. Once the inputs and outputs
for the four neural network models were defineci in the concept and mode1 design phases,
the validity of the neural network concept was then tested by conducting training and
testing on the data for the concrete slabs. This chapter addresses both phases of model
implementation for the four neural networkr that were designed in Chapter 3.
4.2
Data Preparation
the structura1behavior of concrete slabs has been conducted at the Mernorial University of
199 1; Emam et ai., 1995, and Jiang, 1994). The experirnentai tests reported in Marzouk
and Hussein (1991) snidied the behavior of seventeen normal and high streagth concrete
slabs subjected to concentrated Ioads applied axiaiiy through a stub column FoIiowing
tha, additional tests (Emam et al., 1995) were conducted on fourteen reinforced concrete
slabs and column connections subjected to not only axkd load but also bending moment.
To funher study the effects of shear reinforcement on the slabs' behavior, Jiang (1994)
conducted supplemental tests on seven high strength concrete slabs. For each slab tested
in these studies, detaiied Information regarding the factors that d e m i e a concrete slab
A data acquisition system was connecteci to the test setup of the concrete slabs and was
used to autornatidy record severai data elernents during ail of the tests. Deflection at the
slab centers was mea~ufedat a series of loads using linear variable differential transformer
compression face of the concrete slab. Cracks were rnarked duruig Ioading and the final
The frst step in preparation of data involveci fomulating the load-test results for the
thirty-eight fidl scaie reinforceci concrete slabs in the appropriate input and output formats
for each neural network model. Details on how the data was actnipoiated for each neural
network are included in Table 4.1. F i crack patterns were not available for ail thirtythree slabs and Mme of the concrete and reinforcing steel strah gauges were damaged
during testing, so not aiI of the slabs could be used for training and testhg NNs 2 through
4. The following analysis of the straui gauge data @oui for the concrete and reinforcing
steel) was required. Strain gauge readings were reviewed for consistency. those that
remainesi at the extrernes (near O or 1) throughout the test or that fluctuated significamly
during the course of the test were considered to be unreiiabte. If main gauge readings did
not appear reliable, the entire case was removed fiom the pool of data available for
tFaining and testhg the neural networks. The reailting total nrunber of slabs used for
training cases and those that were reserved for later testing o f al1 of the aained neural
In order to validate the Wonnation content of the training cases use& a simple test was
fkst conducteci on training data for the seventeen slabs used in the Hussein study (1991);
this was completed prior to training NNl (loaddeflection m e ) . The test examinai the
LoiidDeflection
--
Data extrapolated
from plot of load
deflection cuve
(produced durig
testind
Crack Pattern
Photos of crack
patterns visuaiiy
intemreted
Concrete Strain Strain gauge
--
--
readings
4
Steel Strain
Strain gage
readirrs
27
28
output parameter (dtimate Ioad reached), depicted in all of the training cases. These
relationships (or general trends) were established through simple regression analysis and
then compared with common knowledge in this domain. Following this analysis, the
concrete compressive strength exhiitited a logical direct relatiomhip with the dtimate load
reached by the slab, and as such, it was concluded that the data was sufncient for initial
training of NN1.
4 3 NeuroShell2 Software
NeuroShell (1990) is an aristng neural network software package which contains alI of
the features that are necessary to train and test a neural network This original version has
been upgraded several times since the original issue. NeuroShell (1990) was used for the
initial training and testing of NNl (Load-defiection).
transferred to NeuroSheU 2, Release 3.0,and the upgraded software was then utilized for
finai modelig of the problem and training of all of the networks (NNs 1 through 4). The
Wmdows-based neural network software was chosen for t s ease-of-use, speed o f training,
and for its host of features that permit user optinrization of network training. Some
advantages of NeuroShell2 inchde: 1) the ability to import and export data files; 2) the
choice of several diffrent neural network architectures, which aiiows the user to select the
paradigm most suiteci to hidher partidar application; and 3) visual trainirig, which d o w s
55
the user to evduate when training is dEcient by viewing the trainhg gtaphicdy or by
viewing the network training statistics. Figure 4.1 demonstrates the user-fnendly aspect of
NeuroSheil 2.
4.4
Training
After the data for the training cases (Appendix A) was input to the software, training was
completed for ail of the neural networks. OnginaUy, for NNl ooad deflection), training
was conducted on the twelve slabs containeci in Mmouk and Hussein's study (1991).
This was done to c o h the suitability of the neural network technique for the problem at
hand. Once this was confirmecl, fifteen cases fiom the remaining two studies were later
added and retraining was conducteci. The addition of the results ?omthese tests widened
the domain of concrete slabs included in training, thereby augrnenting the ability of the
neural network to generake the model. The remaining three neural network models
(NN2 through 4) were then trained with all of the training cases available for each model.
The same iterative procedure was utilized for training each of the four neural networks.
Two separate f o m of neural network architecture were utilized to train ail four neural
4.4.1
Since training is essatiaily an iterative process, two simpEed approaches were carried
out to arrive at the optimal training level for each n e d network using BP. In the first
approach, the "NetPefiectY feature of the software was used where, at predetennined
intexvals during training, NeuroSheU 2 would test the network on an independent test data
set. If the emor on the test set was Iower than the previous opMial network, the new
network would be saved; this process would continue until no improvement in the
network occurred. An average error for al training cases would then be computed for
cornparison purposes. The TurboProp feature of NeuroShell2 was also used; this method
adjusts network connection weights only after the network encounters an entire set or
epoch of training patterns. By doing ttiis, the network can converge at a faster rate than
when weights are randomly updated without the network seeing an entire set of training
patterns.
4.4.2
Stepwise training was next conducted by successively increasing the number of training
epochs (Le., cycles through a cornplete training set) beyond which the error was
minimized on the training set. This was done because the srnaiier number of cycles at the
earlier stages exposed the network to less training h e and thus the network did not focus
on the training cases. This would result, theoreticdy, in good g e n e r h t i o n capabilities
of the neural network on any test cases for concrete slabs not previously applied to the
neural network. On the one hanci, if training t h e were not suflncient, this could mean
unacceptable network performance. On the other hanci, if the network were ovemained
(Le., higher number of cycles since average minimum error), this couid occur at the
orpense of its generabtion perfonnance. Steps of 50, 100,200 and 1000 cycles beyond
the minimum error were progressively applied as the training tirne was increased for each
neural network model. Once again, average errors for training cases were cornputeci for
cornparison purposes.
4.4.3
An alternate neurai network paradigm, the general regession neural network was also
applied to aU of the NN models. This type of neural network has been shown to perfonn
best on models for which there is only a minimal amount of data available to train and test
the model. This architecture was experirnented with in this study as there was a Limited
number of cases available to train and test some of the neural network models; this
paradigxn could theoretidy provide lower mors than the more wnventionally used
Backpropagation neural network.
4.5
Testing
Once the neural networks were trained, the predictive capability o f the neural networks
was then checked on an independent test set. In this case, a weighted average of the
mors for the sample and test cases was computed, ushg a 70% weight on the test cases,
and a 30% weight on the training cases. These weights were randomiy chosen to
emphasize a greater weight on the test cases because, it could be assumed that the neural
network would have a greater chance for learning the results for cases previousiy shown
to it ( m g cases) and a lower error on these cases would be expected; however, a
lower error for the test cases w d d indicate greater performance of the network. Once
the resdts were reviewed, the optimum neural network models were chosen for each
module (those with the lowest weighted =or), dong with the ideai training method to be
utilized.
4.6
Summary
Data preparaon and training/testing are processes which are used to implement the neural
network model. Several iterations of this process are usually required in order to achieve
optimum resuits for the model.
discussed in this chapter, dong with a review of the software utilized to accomplish the
training and testing for all of the networks. Chapter 5 addresses the results obtained &om
training and testing each of the neural network models, as weil as the selection of the ideal
neural network models and training methods utilized for each problem.
Chapter 5
Results
5.1 Introduction
Results for ail of the four neural network modules were produced by the software and
were reviewed on an ongoing basis as each NN was trained and tested. From these
results, the optimum network was selected for each problem in two stages. The first stage
consistecl of choosing the ideal eaining method and network architecture by mulimizing
the weighted error for each method and architecture. The ideal step for stepwise training
was nrst selected and cornpared with the results produced by the use of the NetPerfect
fecmire. Then, the resuIts produced ushg backpropagation were compareci with the
GRNN model and the optimum architezture (with the lowest weighted error)
was
selected- As previousiy descnied, the weghted error is an average of the errors for the
sample and test cases, with a 70% weight used on the test cases and a 300? weight on the
training cases. The second stage involved choosing the ideal model for the problem, again
by minimiPng the weighed mors produced for each individual mode1 and by also
evaiuating the ability of the network to produce redts (for example, load-deflection
m e s ) consistent with those produced by the experimentd load tests. The resuits for the
optimum model for each neural network are tabulated in Appendix B.
5.2
Load-Deflection Behavior
The r e d t s produced by stepwise training for NNla, which predicted the deflection of the
slab at ten Merent load increments and the dtknate load, and NNlb, which predicted the
yield and dtimate loads and ddections, are plotteci in Figure 5.1, showing the training
stages venus the weighted network performance error for eacn network As shown,
NNl a performed ideally at the last step (training iterations beyond minimum average error
= 1000), with a weighted error of 16.31%,
(training iterations beyond minimum average enor = 200), resuiting in a weighted error of
15.09%. This iltustrates the necessity for cunducting stepwise training separately for each
N w Table 5.1 compares the weighted mors r d t i n g ffom stepwise training with those
produced by use of the NetPerfect fumire. It can be seen from the results for both
networks that the Neteerfect fature trained the networks in the optimum manner.
Although the average error on the training cases was lower at the optimum aepwise
63
-.
50
100
200
lm
Model
NN la:
NN lb:
Network
Architecture
Backpropagation
Backpropagation
General
Regrcssion NN
Backpropagation
Backpropagation
General
Tiaining
Meehanism
Average
Average
~rror*on E m r on
Tirinhg
Test
Weighted
~rror*on aii
Cases
Cases
(%)
cases
W)
(%)
NetPerfect*
Stepwise**
N/A
8-36
1-91
1-72
19.72
22-49
16.3 1
16.3 1
13.85
10.22
NetPerfect
Stepwise
N/A
6-87
7-59
19-83
12.04
18.30
20.97
15.09
10.48
20.63
metwo*
training level for each networlg the average error for the independent test cases was much
higher (>22% for NNl a and >18% for NNl b), leadkg one to conclude that the networks
a d y ovemaineci on the sample cases and rnay have lost their ability to generalize for
Table 5.1 also compares the weighted erron produced by the backpropagation NN with
those produced by the GRNN model. While the GRNN version of NNla showed a Iower
average weighted error than the optimum network trained d
the ultimate load and deflection predicted by the network feu fr short of the actual
d a t e load defiedon reached in the experimental tests, as shown in the sample slabs in
Figures 5.2 and 5.3. And, the error produced by the GRNN network for NN1b was in
excess of 20%4 the worst for ail of the training conducted. Thedore, the usefulness of
the backpropagation architecture for this particular neural network was confirmed.
experimental tests.
determineci that this neural network could be used as a reliable alternative to the costly test
500-
4!!m-)O 4 -
350
-+BP Network
+GRNN Nework
+Amal
O
IO
15
20
2S
Deflection, m m
+BP Network
+GRNN
Network
+Acual
O
10
1s
.
20
2s
Deflection, m m
procedures for the prediction of the load-defidon curve values, both numencally and
glaphically.
The accuracy of the crack pattern modei was more ditncult to mterpret as the ab*
of the
of eight
These weights were then apptied to the error for that output (error computed using the
absotute value of
Actual - Network
ActuaI
was then computed for both training and test cases. As was done in the loaddeflection
c w e modei, a nnal weight was applied to this average* using a 70% weight on the test
cases, with a 30% weight on the training cases. The nnal mors are Summarized in Table
5.2 for both backpropagation and GRNN modek. Using this method for cornparhg
errors, the second stage of stepwise training (number of cycles since minimum average
error = 100) provided the optimum d t s 7 with an average weighted error of 24.8%. As
ModeY
Emr
Method
Network
Architecture
Training
Mechanism
Average
Average
~rror+on Error on
Training
Test
Cas=
(%)
NN 2.:
Badrpropagation NetPerfect*
B a ~ r o p i g a t i o n Stepwise**
Using
Wtd
Error
NN2a:
19.49
14-74
Genera
N/A
1 1.38
Regression NN
Baeropagation
NetPerfect
38.00
% of
Badqropagation
Stepwise
28.57
Wrong
Generai
N/A
19.05
pictu=
Regression NN
Backpropagation NetPerfect
16.33
NN 2b: Backpropagation
Stepwise
2.00
Generat
NIA
2.87
Regression NN
* - Auto-optimization feature ofNeuroshell2
** +
++
Cases
Weighted
~ r r o r o aii
n
Cases
('w
('w
55.82
30.39
29.08
57-22
24.78
43 -46
83.33
50.00
66.67
5 1.60
43.57
52.38
23.36
28.07
20-03
21.25
20.25
14.88
cm be seen nom Tabie 5.2, this method of training produced a weighted error which was
Iower than that produced by training conducteci with the use of the NetPerfct feature.
Also, the choice of the backpropagation architecture over the
again confinned, as the GRNN model produced weighted errors in excess of 40%.
In the second method for computing errors, the total number of incorrect pictures of the
crack pattern as a percentage of the overall number of crack pattern was calcdated to
detexmine the overd efFkctiveness of the model. These erron were then weighted for the
eaining and test cases a s was done in the first method. The second stage of stepwise
training again produced the optimum resdts, however, with this method of analyzhg
errors, 44% of the predicted crack patterns would be incorrect or 56% of the predicted
crack p a t t e ~ produced
l~
the correct pictures. Tabie 5.2 again nimmarizes these erron for
all levels of training.
Regardless of the method utilized to evaluate the overall error for the neural n ~ o r k a,
weighted error of either 24.8% or 44% couid not be considered accurate enough to
reIiabIy predict a crack pattern for a previoudy untested concrete slab. Because of these
inaccuracies, a new modei, NN2b7was proposed with fewer, more quantitative outputs in
an effort to fkther minimize the mors. This modei, as s h o w in Figure 3.4, predicted
only the tangentid cracking radius and the extent of radial cracking. Although Iess exact,
the nnal crack pattern formafion NeuraI networks generaiiy pedorm better with fewer
Table 5 2 , the weighted errors for the new NN mode1 p 2 b ) were substantially less than
those reflected in the results for NN2a for ail forms of training and architecture. W~tha
weighted error of 14.88%, the general regression neural network produced the optimum
Aithou& a dEcient amount of training cases appeared to be available for this network, a
high percentage o f the training cases (more than 80%) predicted crack patterns indicative
of fdure due to pure or d u d e punchkg shear. As a resuit, d of the networks appeared
to focus on this type of M u r e pattern and had difncdty generalizing to other crack
patterns. To illusnate this, Figure 5.4 coatains network-produced sample crack pattern
for punchig shear flailure and flexurd M u r e as compareci to the actuai crack patterns
encountered during experimental testing.
The GRNN mode1 is designed to predict outputs around the average for the results in the
training domain As a resut, this form of neural nmork was more successfu than the
more ttadjitiod backpropagation modei in predicting outputs for this particular group of
training and test cases. It is anticipateci thet, with the addition of fiuther test cases
wider training domain (ie., a wider V81Ciety of crack patterns), the backpropagation neural
network would provide better resuIts, thereby M e r improving the performance of this
rnodel.
5.4
Separate training and testing was conducted for each wncrete strain dismbution mode1
(NN 3% which predicted tangenthi concrete strain distribution at three points dong a
radius through the slab, and NN3b, which predicted the wncrete strains at the colurnn
face at four load increments). The optimum NN architecture and training method was
then chosen for each mode1 as each was designed to provide m d y exclusive results.
Unfominately, a minimal set of results (nine in total) were available for training and testing
both models as the majority of the strain gauges were damageci afler cracks started
forming in the concrete. Table 5.3 summrkes the errors produced during training and
testing for both models. As s h o w the m o r s were quite hi@ for aiI f o m of training and
testhg except that conducted using the NetPerfect and Turbohop features; the minimum
weighted errors produced by these networks were 22.1% for NN3a and 17.26% for
NN3b. When plotted against the resuits fiom
were both able to produce resdts that followed trends encuuntered by the actual tests as
is shown by Figwes 5.5 and 5.6. It is anticipaed that the r d t s for both networks c o d d
o d y improve with the addition of M e r experimental data for training and testing the
networks.
Table 5.3. Resdts for Training Concrete Striiin Distribution Neurai Network (NN3)
Modd
Error
Method
NN 31:
NN3b:
* ** +
* -
Network
Architecture
Backpropagation
B=kpropagation
General
Regression NN
Badrpropagation
Backpropagrtiw
General
Regression NN
Training
Mechanism
Average
Error'on
Training
Averrige
Error on
Test
Weighted
Cases
("/)
Cases
(W
(%)
NetPerfiect*
Stepwise**
5.21
7.8 1
29.35
48.02
22.10
44.09
N/A
NetPerfect
Stepwise
40.72
1222
18.90
3 5 -28
19.42
36.92
26.42
24.16
E m r " on ai
Cases
17.26
N/A
28-69
40.64
3 7.05
Auto-optimitation feature ofNeuroshell2
Trainmg method by which the number of epochs since minhum error is
sequentially increased until the optimum results are obtained.
Average error = Absolute value of INetworkOumut
Output)
A d Output
20
40
60
80
100
Figure 5.6. NN3b Concrete StrPin Distribution at Column Face for Slab # M l
5.5
As there was just one mode1 for steel saain distriiution, selection of the ideal model only
involved choosing the training rnethod and NN architecture which produced resuits with
the minimum error for this rnodel. Training and testing for this network again involved a
d e r number of cases tban those amilable for the Ioad-deflection and crack pattern
models (nine training cases and three test cases) as strain gauges were again damageci
during experimental testing, reducing the data available. The neural network, however,
pdonned well considering this limited amount of available data Table 5.4 shows the
weighted erron for the resuits for this model; as shown, the network aained with the
for the steel strain distriiution through a sample slab in the radial direction is also shown
in Figure 5.7. The network predictions foilow the actual results quite closely when plotted.
Whde an error withui this range can not be considered flly accurate, a reasonable
distribution can still be shown and these r d t s can be wnsidered reasonable with respect
to the complexity of the problem and the limiteci number oftraining cases availabie.
Table 5.4. Results for Training Reinforcing Steel Strain Distribution Neural
Network (NN4)
Network
Architecture
Mechanism
~ v e r a ~ e Average
Emr'on
E m r on
Training
TestCases
Clists
('w
weighteci
E r r o r on ai
Cases
('w
(W
Backpropagaton
Badrpropagation
Generai Regression NN
" -
NetPerfect
Stepwise
N/A
13.63
11-69
22-11
14-91
21-28
22-58
14-52
18-40
22.44
t+
5.6
Summary
Results for the four neural network models are disaissed in this chapter. The combineci
implementation of these models in a comprehensive spreadsheet tool is discussed in
Chapter 6,and conchisions reached fkom these results are discussed in Chapter 7.
Chapter 6
Spreadsheet for the Prediction of the
Structural Behavior of Reinforced
Concrete Slabs
6.1
Introduction
This chapter focuses on the development of a spreadsheet which combines the four neusal
network models imo a cornprehensive tool that can be used for the structural anaiysis of
reinforced concrete slabs. Given the fcton that dem'be the slab, a spreadsheet can be
Utizized to predict, through four separate modules, the loaddeflection m e , the filure
crack pan-
the concrete stmh distncbution and the reinfiorchg steel strain distribution
for the slab. A user-fiiendly %terface'' sheet guides the operator of the spreadsheet
through the four modules for simple and quick prediictions which cm then be printed for
M e r use.
6.2
The spreadsheet was deveioped as a Microsoft Exce15.0 Workbook which interfces with
the NeuroShell2 software for neural network predictions. The workbook is divided into
seven separate worksheets: an "interfke"or main mem sheet, an instructions sheet, an
input data sheet and four output sheets which display: the prediaed Ioad-defiection cuwe,
the predicted crack pattern at failUreythe predicted cuncrete tangentid strain distribution
and the predicted strain development at the c01um.n &cey and the distribution of the
maximum reinforcug steei strain distriion in a radial direction The user moves
through the workbook by clicking d i r d y on buttons on the main interface sheet, first by
inputhg the nineteen factors which descn'be the slab on the input data sheet, then by
Basic recorder
featwe of Microsofi Excei 5.0. This f e u e records the mouse movements of the
programmer to a macro which then sirnulates these movements whenever the bunon is
activateci. The neural network predictions for each module were accessed through a
The "CALL" fnction of Microsofi Excel5.0 was utilizd within ce& on the output sheets
to cal1 the p r d u r e in the DLL. A sepanite cell for each output item would then c d the
"Predict" fnctioa of Excei to open and execute the trained neural networic. nie
foilowing syntax was used in each output ceU:
where defqath is the file path for the trained NN, input-array is the array of ceUs which
contain the input data, and output is the output node number. The "Chart Wlzard" f e u e
of Excel was then utilized to create graphic representations for the NN predictions for
each module.
A reinforcd concrete slab 6om the Emam et al. study (1995) is chosen (Slab # Ml),
because the redts for this slab are hown and can be refmed to by the reader for
cornparison purposes. These results were used to train or test aii of the four NNs in the
current investigation
Figure 6.t shows the interfhce sheet that the user sees when the file is exearted. The
"InstnictiOnsy' bunon can be clicked upon at any t h e for help in using the spreadsheet, if
82
--
Instructions
necessaxy- The "Input Data" button is first clicked upon for input of the data; this input
data screen is show in Figure 6.2. Mer inputting a d o r editing the data, the user is
retumed to the interface sheet. The user can then choose any one of the four icons
representing each of the NN modules for prediction of the behavior of this example slab.
Figures 6.3 through 6.6 show each of the screens that are displayed when each of the
butions is ciicked,
The total t h e spent inputting the data and receiving the r d t s was approximately thee
minutes. As can be seen, this spreadsheet provides a very quick method by which one
could estimate several aspects of the stmctwd behavior of reinforced concrete slabs.
6.4
The spreadsheet mode1 was next vaiidated by comparing predictions for the ultimate
punching load wah actuai resuits for selected tests conducted by Elstner and Hognestad
(1956), Kinunen and Nylander (1960), Regan et ai. (1993) and Hallgren (1996).
Representative slabs for each series of tests were chosen and were compareci by using the
ratio of the spreadsheet predicted punching load divideci by actual punching load. The
results of this cornparison are compiied in Tables 6.1 and 6.2; results for the tests used to
train the neural network withi. the spreadsheet (Marzouk and Hussein, 1991; Emam et al.,
1995, and Jiang, 1994) are included in Table 6.3 for cornparison.
. -
LOAD-DEFLECTION CURVE
* O - none
460
4
0.00
2516.39
2377.92
1282.51
200
'l
Barn
lm
OO
300
480
500
480
2 w 4 o o e o o
'CakWtdinNNl -Losd--NcuralNdwwl<
~triht
(X 103
Figure 6.5 Predicted Coacrete Strain DMbution Sheet for Example Problem
88
Yield radius, mm
Max. steel m i n @ col face
Max steel strain @ mid slab
Max steel strain @ slab edge
382.51 21
3511363
2749.809
1552-465
+
s
c
z m
Qi
g Io00
Figure 6.6 Predicted Steel Strn Distribution Sheet for Eumple Problem
89
4,
SIab
No
mm
dd
min RC MPa
2,lY
2.19
2.19
2.19
2.19
2.23
1.78
1.78
1.78
1.78
1.78
1.78
2.23 1.78
2.23 1.78
3.12 1.78
O
O
A-l(c)
A-l(d)
A4
B-9
B-11
B-14
A-11
116
116
116
116
116
114
114
114
114
117
117
121
127
128
125
124
128
119
120
123
127
127
128
125
130
1.28
1.28
1.24
1,18
1,17
1.2
1.21
2.34
O
O
O
4, Lond tr,
Type mm Type mm
Elstner and Hognatad, 1956
A4a) 2
15 O 120
15 O 120
A-l(b) 2
2
15
15
15
15
15
15
15
2
2
2
2
2
O
O
O
O
O
O
O
15
5233
15
5269
3651
3467
3415
3465
5107
5251
5281
5089
5125
3390
3448
3408
2
2
2
15
15
15
15
15
15
15
15
15
15
15
15
15
15
3436
2
2
2
2
2
2
2
2
2
2
2
O
O
Q
O
O
O
O
O
O
O
O
O
O
O
O
O
120
120
120
120
120
120
120
1960:
149
150
153
153
153
150
150
158
151
151
155
152
152
153
150
153
2,5
2.44
234
2.34
2.34
2.4
2.31
1.66
1.66
1.66
1.66
156
1,66
1.66
166
1.66
1.66
1.66
1.66
1.66
1.66
1.66
1.66
f,,
Fa,
MPa
OPa
f00,
O
O
O
O
O
O
O
14.1 1.07
20.3 1.54
25,2 1.92
36,s 2.80
26-1 1.98
43.9 3.34
13.5
1.03
50,5 2.42
25.9 1.97
33.5 2.55
32.5 2.47
40.1 3.05
O 32,4 2.46
O 33.4 2.54
0 33,9 2S8
O 33.8 2S7
O
33 231
O
32.4
37.6
O
O
O
O
O
O
2.46
2.86
32.9 2.50
34,3 2.61
33,7 2.56
35,s 2.72
33 231
34.3 2.61 26.34 2.07
35
461 200
387 521
Nm!iaia: d, = a g p j p h diamda, t,
diiriicity, p fcinfwcing
--
2.52
l,,
1,35
10
10
10
10
12
O
O
O
O
O
10
20
20
20
6
10
20
26
27
20
Hallgren, 1996
HSCO 2
19
HSCI 2 19
HSC2 2
19
HSC4 2 19
HSC6 2 19
HSC9 2 19
Ndaiimi: 4 =-a
2
3
4
5
6
21
22
23
24
25
\O
2
2
2
2
2
2
2
2
2
2
2
2
2
O
O
O
O
0
O
O
O
O
120
120
120
120
120
98
98
98
98
c/d
mm DC MPa
1.38
1.38
1.38
1.38
133
98 1.38
120 98 1.38
120 98 1,38
120 98 138
120 100 1.35
12098 1,3
1,53
1.53
1,53
1.53
1.53
1.53
1.53
1.53
813
1.32
1.53
225
225
225
225
225
2.1
2.1
2.1
2.1
2.1
2.1
200
200
194
200
201
225 198
2.1
2,l
2.1
2.1
2.1
2.1
MPa
OPa
Size S h a p
kN
Pm
O
O
O
0
O
O
O
O
O
O
O
O
O
O
O
O
O
0
O
4%
200
565
451 0.80
i$= m i i m e modulw of
n d w d prdirtd uhimta
Jiang, 1994
HSl7
I
JS18
2
HS20
HS21
HS22
2
2
2
HS23
2
Emam et al,, 1995
Ml
2
19
19
M. 2
2
19
M3
M4
M5
M6
M7
19
19
19
19
19
19
2
2
M8
M9
As shown, the spreadsheet provided excellent predictions for the dtimate punching loads
for aii of the Elstner and Hopestad (1956) slabs except for slabs B-9 and B-14. Both of
these slabs, however contained a remforcing steel ratio of 3.0, which was higher than the
reinforcing steel ratio for any of the slabs (Martouk and Hussein, 1991; Jiang, 1994;and
Eman et al., 1995) used to train the newal network within the spreadsheet (NNlb). The
spreadsheet also provided satisfctory predictions for many of the Kinunen and Nylander
(1960) slabs as weii as the Regan et al. (1993) slabs. Again, in the cases for which
predictions exceeded 25% of the a
d loads
than .75), at least one input parameter was outside or at the boundaries of the range or
domain of the parameters for the slabs used to train the neural network within the
spreadsheet. For example, the ratio for slab # 3390 (Kinunen and Nylander, 1960) was
1.54; for this particular slab, the spacing of the reinforcing steel was 74 mm, which was
jus above the minimum spacing (71.4 mm for slab # HS3, Marzouk and Hussein, 199 1) of
the slabs used to train NNlb. F i y , rather poor predictions were provided for the
Wgren (1996) slabs; however, the conmete compressive strength for most of these slabs
was well in excess (at least 20 MPa) of the concrete compressive strength for the slabs
used to train the neural network containeci within the spreadsheet. Excellent predictions
were provided for aii of the slabs within the Mrzouk and Hussein (1991), Jiang (1994)
and Emam et al. (1995) slabs because these were the siabs which were used to train and
test the neural network embedded within the spreadsheet.
cornparisons therefore suggest that the spreadsheet mode1 perfoms well when presented
with siabs within the domain of the slabs used to train the neural networks within the
spreadsheet.
6.5
Summary
This chapter has describeci the development and impiementation of a spreadsheet mode1
which can be used by structurai engineers for prelimuiary predicton of the structural
behavior of a reinforced coacrete slab. The spreadsheet combines the four previously
developed neural network models into one simple-to-use tool which can provide
Chapter 7
Summary
The investigation under consideration was conducted to evaluate the feasibility of using a
branch of artificial intelligence hown as neural n ~ o r k to
s predict several aspects of the
structural behavior of reidorced concrete slabs. This technique was examinai because, in
previous studies, neural networks have been found to be a quick and reiiable alternative to
Iengthy experimentd testing or detded caldations Four separate neural networks
W l b : load-deflection; NN2b: crack panem; NN3a & NN3b: concrete strain distniution;
and NN4: reinforcing steel strain distriiution) were developed using a varety of models
and training techniques for each network in an attempt to seek the optimum neural
network that could be constructeci for the problem unda consideration. One neural
network software prograrn, NeuroShell2, was utilized for modehg, training and testmg of
ail of the'neural nemorks in order to achieve consisteacy of results for cornparison
purposes. Ail four neural networks were trained and tested using the r d t s fiom three
evaiuated the behavior of normal and high strength concrete slabs subjected to
concentrated, 0exura.l and cyclic loading conditions. In addition, ail four neural networks
considerd the same input data, which consisted of a d e r of variables (groupecl under
slab geometric dimensions, aggregate properties, concrete propemes, reinforcing steel
properties, and loading and boundary conditions) which could ffect the behavior of the
slabs. NN1 predicted the load-deflecion behavior of reinforcecf wncrete slabs in two
ways; one mode1 predicted the deflection at ten load increments while the other mode1
predicted the yiefd and ultimate Ioads and defldons. N e c NN2 predicted the crack
pattern at M u r e of the concrete slabs usmg two approaches. The first approach predicted
a schematic representation of the crack pattern whereas the second approach predicted
just two numerical aspects of the final crack pattern. The rhird n e d network, NN3,
predicted the distribution of concrete strains throughout the slab through two versions.
The first version predicted the maximum concrete strain distribution at three points dong
a radius of the slab while the second version predicted the wncrete strains at the column
face at various load increments. The final neural network NN4, predicted the distribution
of reuiforcing steel strains throughout the concrete slab.
ResuIts &om the four neural networks, either individually or combine& could provide
useful information to a structural engheer regardhg the prediction of the behavior of the
wncrete slab. To fiditate access to this information, a comprehensive spreadsheet tool
was developed which included ail four of the neural networks in one easy-to-understand
format- Resuits codd then be obtained for any or d of the four neural network models,
providiag vduable information for subsequent design or andysis of reinforcecl concrete
slabs.
7.2
Conclusions
Based on the neural network m o d e h g trainhg and testing conducted, the conclusions
reached regarding the use of neural networks to predict the stnicturd behavior of
reinforceci concrete slabs are swnmarized below.
7.2.1
The Ioad-defiection n
would determine the applicability of the neural network technique for the problem at hand.
1. The backpropagation technique was the most accurate Paining algorithm for this
be the architecture most suited to problems within the civil engineering realm, due to
the simpiicity of the architecture.
2. The mors produced were higher for the neural network (NNla) with the greater
number of outputs to predict, Ieading to the conclusion that neural networks are more
accurate when fmer outputs are predicted. The reason for this may be tfiat, with
fewer ouputs, the connections within the neurai network layers would be less
3. The optimum neurai network mode1 was able to produce resuits with an overaii error
those produced during arperimentai teSang, again confirming the suitability of the
neurai network technique as a reiiable aiternative to such testing.
7.2.2
NN2:
From the results produced by this neural network, the following conclusions cm be
dram
1. Mapping crack patterns using b i n q numbers to indicate the exact locations o f cracks
is a cornplex task which is not suitable for neural network modehg due to the high
number of outputs which are required.
problem, the neural network rnodels were more accurate for p r e d i h g those outputs
which could be quantified rather thanclassified.
resdts for this mode1 due to the "clustering" of data in the training domain around the
same values. This codhms the suitability of this neural network type for predihg
r d t s for problems where the results are sunilar for each training case presented to
the network. It is anticipated th& because the backpropagation neural network was
more able to provide better generalizatim abilities in the other neural network models
for this problem (NNI, NN2 and NN4), this fomi of architecture could provide
Eiproved results with the addition of training cases with a wider variety of crack
patterns.
7.23
Despite the minimum rmmber of cases avaiIable for training and testing of both NN3 and
NN4,the neural networks were stiU able to predict results thai differed by 17.26% and
14.52%, respective1y fiom those produced during experimental testing. SimiMy, NNs 3
and 4 were able to predict strain distriiution curves that almost matched those produced
during experimental testing. The foliowing conclusions can also be drawn nom these
redts:
1. Neural networks can, given sparse training data, predict redts for cases previousIy
2. The backpropagation technique again provided resdts with mors that were in the
m o a acceptable range.
7.2.4
General Conclusions
In summary,the fonowing conclusions can be drawn f?om the combineci results for all four
neural networks:
2.
The backpropagation
2. Neural neworks perform best when a minimal number of outputs are predicted by the
rnodel.
3. The neural network models predicted results with the minimum errors when they were
presented with test cases within the domain of the training cases, especially when a
minimal aumber of cases was used to train the neural network model,
5. Simple s p r d h e e t s are powaful tools that can be used to illustrate and summarize
While the effectiveness of the neural network technique has been wnclusively
demoastrated by the work contained withui this thesis, M e r research could serve to
enhance this effdveness. For example, the addition of resdts from a greater number of
reinforced concrete slabs with a wider v a r i q of propemes and loading conditions will
moa Wrely serve to M e r improve the accuracy of the neural network modeis already
devdoped. These neural networks could be srpandeci to predict r e d t s for a wider
varety of reinforceci concrete structurai elements such as b e ~ m ~columns
,
and shear walls.
From this, a g e n d cornprehensive tool for the prediction of the structurai behavior of
reinforced concrete could then be developed for general use by structural engineers as a
References
Adeli, H,and Park, H S., (1995). "CounterpropagationN d Networks in Structural
Engineering." Jounirzl of Struc~aZEngineering ASCE, 121(8), 1205-1212.
Ars&
A, and Ince, R,(1994). "The Neurai Network Based Design of Edge Supported
Canadian Standards Association, (1994). "Design of Concrete Structures (CSA A23 -394). " Canadjm Stanapds Asociti'on, RexdaIe7Om
Carpenter, G., and Grossberg, S. (1987). "ART-2: Self-organization of stable category
recognition codes for analog input patterns." AppZied Optics, 26(23).
Carpemer, W. C.,and Barthelemy, J.F., (1993). "CornmonMisconceptions about Neurai
Networks as Approlomaton." Proc, 5th Intl. Cor$ on CC.and Strrrct. EngrgCompun'ng & 3rd Inrl. Coj. in the Appl. Art$ if:nteII.CN. and Shuct. Engrg.,
Civil-comp Limited, Edinbwgh, Scotl., 1 1-17.
Caudill, M., and Butler, C., (1990).
Cambridge, Ma
Chauvin, Y., and RunieD., ed. (1995). BACKPROPAGA7ON:
Theory,
Archilecmes, tard Applcations- Lawrence Eribsiim Assoc., Inc., h b i , Hillsdaie,
N.J.
Chen, K M-, Tsai, K K, Qi,G. Z., Yang, l. C. S., and Amuii, F., (1995). " N e d
Network for Structure Contr01." J
i of Cornputhg in CN. Engg. ASCE,
9(2), 168- 176.
Darwin, D., and Pecknold, D. A, (1974). "InelasticModel for Cyclic Biaxial Loaduig of
Reinforceci Concrete." CM2 Engineering Studies. Stnrcl~rdResearch Senes, No.
409,University of Illinois,Urbana, Illinois, July, 1974.
dePaiy H. A R,and Siess, C.P.,(1965). "Strength and Behavior of Deep Beams in
Shear." J m I of the ShucturaI Division, ASCE, 9 1(ST5), 19-41.
F., Chang, K C., and Lee, G. C., (1994). "A Structural Damage Neural
Network Monitoring System." Microcornp. in CN.figrg., 9,83096.
Elkordy, M.
Elstner, RC., and Hognestad, E., (1956). "Shearing Strength of Reiaforced Concrete
Slabs." ACIJozrrtt~I,28(1), 29-57.
Emam, M.,Marzouk, H. M. and Hilai, M- (1995). "Effect of Concrete Strength on the
Behaviour of Slab-Column Connections Subjected to Monotonie and Cyclic
Loading." Engineering and Applied Science Research Report No. 95-002,
Mernorial University of Newfoundland, St. John's, Newfomdand, Canada
Garrett, J. H., (1992). "Neural Networks and their Applicabtlity within Civil
Engineering."Proc., 8th Cor$ of Comp.in Civ. Engrg.,ASCE, New York, N.Y..
1155-1 162,
Goh, A T. C., (1995). "Prediction of Utnate Shear Strength of Deep Beams Using
N d Networks-"AC1 Stnrchcral J-I,
92(1), 28-32.
Hanna, A S., and Senouci, A B., (1995). WEXJROSLAB Neural Network System for
Horizontal Fomwork S e l d o n " Cm J
i of CN.Eng~g-,
22,785-792.
Hinton, G-, and Sejnowski, T., (1986). "Leamhg and Relearning in Boltmiann
Machines." Pmalel Distributed Processiing. Vol. ., MIT Press, Cambridge,
Mass,,282-3 17.
Hopfield, J., (1982). "Neural Networks and Physicai Systans with Emergent Collective
Computational Abilities." Pm...
Nid. A c d m y of Sciences, 79,2554-2558.
Hni, T. T. C. and Mo, Y. L.,(1985). "Softenuigof Concrete in Low-Rise Shearwalls."
ACIJ. P~Yx...
82(6), 883-889.
Kan& H-T., and Yoon, C. J., (1994). "Neural Network Approaches to Aid Simple TNSS
Design Problm. " Microcomp- in Civ. Engrg., 9,211-2 18.
Kasperkiewicq J., Ra- J., and Dubrawski, A, (1995). "HPCStrength Prediction Uing
m c i a I Neural Networkrkn
JoMnal of Computing in CN.Engr,.,
ASCE, 9(4),
279-284.
Klarmnen, S., and Nylander, Ky(1960). "Punchhg of Concrete Slabs Without Shear
Reinforcement." Tt~nracn'omNo. 158, Royal Institute of Tecbnology,
Stockho@ Sweden
~tansactiomon
Kupfer, K B., and Gerstle, K H (1973). "Behaviorof Conaete Under Biaxiai Stresses."
J of figneeering Mechcs. ASCE, 99(EM4), 852-866.
Minsky, M-L. & Papperf S., (1 969). Percepbons MIT Press, Cambridge, Mass.
Mo, Y. L., and Lin, S. S., (1994). "Investigation of Framed ShearwaiI Behaviour with
Netirai Networks." M w n e of Conmete Resemch, 46(169), 289-299.
Moe, J. (1961). "Shearing Strength of Reiaforced Concrete Slabs aad Footings under
Concentrated Loads." LkveZopment Department BuIIetin No. D47, Portland
Cement Association, Skokie, IIl., ApriL, 1961.
Mo*
O., Hegazy, T., and Fazio, P., (1992). "Potential Applications of Neural
Networks in Coosmiction." Cmr J
d of CiY.Engrg, 19,52 1-529.
PateI, V. C.,McClendon, R W., and Goodnim, J. W., (1994). "Crack Detection in Eggs
Using Cornputer Vsion and Neriral Networks." AI Applcutiom, 8(2), 21-3 1.
Sadamsubramanian G., and Rajaselcaran S., (1 996). " C o & ~ i v eModehg of Concrete
Using a New Failure Criterion." Cornputers & Stmctwes, 58(5), 1003-1014.
Schliach, J.. Schafer, K and Jennewein, M-,(1980). "Towards a Consistent Design of
Reinforcd Concrete Structures. " J m I of the Prestressed Concrete Immte,
32(3), 74-150.
Siao, W. B., (1993). "Stmt-&Tie Mode1 for Shear Behavior in Deep Beams and Piie
Caps FaiLing in Diagomi Splitting." AC1 Struc~aI
J~trrnaI,
90(4), 3 56-363.
Snih,
wr(fiam't,T.P., Khjwh,
APPENDIX A
NEURAL NETWORK INPUT DATA
Table A.1.a. Input Data for Load-Deflection Neural Network (NNI) Training Cases
NN la NN lb Slab Agg d. Load i. d Slab L m
f,
fa
E.
Zase # Case # # Type m m Type mm mm CM rn BC MPa MPa GPa
16
2
3
S
6
7
8
3
4
6
8
9
10
11
12
10
10
11
12
liang, 1994:
13
13
14
14
15
15
16
16
12
13
1s
HS17
HS18
HS20
HS21
17
17 HS22
Zmam et al,, 1995:
18
18
M2
19
19
M3
20
21
22
23
20
M4
21
22
M6
24
25
26
M5
23
24
M7
25
26
M9
MI0
M8
2
19
74 3.54 35.42
170
490 200
Ndaiais: 4 = aggegak d i s r n a , t, = slab thickn-, 4 = slab efidive deph, L = slab span, f, 4 mudeamyressive m g t h , Ta = made taisilo sirmgth, E, = mudo
modulus of daaicily, p = rtinfwcingsteel d a , q = a d epacing f, = sli yield m@Fq
h,me1 &lus of dadicily
Table A.1.b. Input Data for Load-Deflertion Neural Network (NN1) Test Cases
L
NNla NN lb
Agg d, Load 4
4 Slab m
Case # Case # Slab # T m mm Type mm mm cld m
Manouk and Hussein, 1991 :
120 90 1.67 2
1
4
1
20 O
1
20 O
2
1
17
150 12 12.5 2
3
2
5
1
20 O
150 95 1.58 2
4
3
11
1
20 O
90 70 2.14 2
S
4
Jiang, 1994
14
5
HS23
Emam et al,, 1995
7
6
Ml
6
f' fa E,
BC MPa MPa GPa
O
O
O
O
2 O
20 O
120
19
150 1 1 1 2.25 2 O
19
IM 1 1 2.21
95 2.32
1 0
66 3.17####
2.4 2
30 2.28 #### 0.9 1
68 3.26 #### 0.6 1
703,36#### 1 1
72 3,46 #### 1 3 1
94
1
1
1
1
125
125
150
602,88#### 1,l
200
200
200
O
O
O
O
490
200
490
2 0
71
1
1
1
1
1
490
490
490
490
490
166
170
200
200
Ndaiau: d, = a
m diunda, t, = slab thibrncss, 4 = slab ~ ~ i depth,
v aL = slab qui,% = a m a d o oomprcssive atKngth, fa = a
m
W taiirile itraia,E, = m u d e modulusof
diaidy, p = reinfwcing a d d o , q = dad spacing f, steel yield slraifl, E,= deel modulus of dasicity
Table A.3.b. Input Data For Crack Pattern Neural Network (NNZ) Test Cases
Case# # Type mm Type mm mm c/d m BC MPa MPa GPa
Marzouk and Hussein, 1991:
2
1
20 O
120 95
1
20
O
120 95
3
Jiang, 1994
3 HS7 1
20 O
150 119
Emam et al,, 1995
19 O
150 119
4
M2 2
19 3
S Ml1 2
150 119
6 M12 2
19 3
150 119
1 7 Ml3 2
19 3
150 119
Ndnlianr: 4 = a
m diunda, 4 = alab thidaiaia. d, ix
1
2
138
1.58
2 0
2 O
70 3.3634.68 0.84
69 3.31 34,08 1,47
1
1
1
1
125
71
49020
490200
2.1
2 O
67 3.2234.08 \,O9
166
490200
2
2
I
1
170
170
170
170
1
1
1
1
O
O
4W200 O
2.1 2 O
2,l 2 O
2.1 2 O
2.1 2 O
1
0.5
0.5
49020
490 200
1
1
75.8 3.64 35.80
36.8 1.76 27.03
1
1
490 200 O
slab efkdiw deph, L = slab qan, f, = madeannprtssivo taraiglh, fa = m u d e taisile iami@h,E, = oma-de
Table A.4. Output Data Lr Cnck Pattern Neural Nchvork (NNZ) Training and Test CI~M
Punching Extent of
Radius,
radial
Case # Slab #
mm
cracking
Manouk and Hussein, N O 1 :
NN2
rraining
1
2
3
4
5
6
hnching Extent of
Radius,
radial
NN2
Test
Case #
Slab #
mm
cracking
16
1
7
4
5
6
8
9
10
11
12
13
14
15
7
8
9
1O
11
12
13
14
liang 1994:
1s
16
17
18
19
20
HSt8
HS19
HS20
HS21
HS22
HS23
3mam et al., 1995:
21
M3
22
M4
23
24
25
M5
M7
M9
26
Ml4
MI
27
M2
MI 1
MI2
M 13
550
475
575
238
3
2
4
4
Table A.6. Output Data for Concrete Stnin Neural Network (NN3) Training and Test Cases
TRAINING CASES
NN 3a
b ~&
r
he
Case# Slab# XI@ x W 3 x W 3
Marzouk and Hussein, 199 1
1
5
1750 650 350
Emam et al,, 1995
2
MI 2500 2400 2100
TEST CASES
I
Case# S M # x l ~ ~ x l xlo4
~ ~ x l ~ ~
Marzouk and Hussein, 1991
Table A.7. Input Data for Steel Strain Neural Network (NN4) Training and Test Cases
TRAINING CASES
--
2
2
M7
M8
8
9
Ml3
2
2
2
4
5
6
Ml4
19
19
19
19
19
19
19
2
2
2
O
3
3
1 4 Slab L
fa fa
Ea
mm m m c/d mm E
X MPa MPa GPa
--
1
2
1
1
71
100
32 1.54 25.73 1
36 1,7426.88 1
34 1.63 26,26 0,5
35 1.6926,61 1
67 3.2234,11 1
37 1,76 27.03 0.5
35 1,7026.64 1
2
2
1
1
1
1
1
1
1
1
170
170
170
170
170
170
170
1
1
1
1.9 O
1.9
1.9
1.9
1,9
O
O
O
0
1.9 O
1.9 O
2
2
1
1
1
1
1
490 200
490 200
490 200
490 200
490 200
4W 200
490 200
490 200
490 200
O
O
O
O
O
O
O
O
O
TEST CASES
NN4 Slab Agg 4, h d
,Cese# # Type mm
Ernam et al., 1995:
' 1 M2 2
2 MI1 2
3 Ml2 2
19
19
19
Qpe
O
3
3
4 d, Slrib L
fa
fa
Eo
m m mm Jd m m BC MPa MW GPa
150119 2.1 1.9 O
150125 2 1.9 0
150 125 2 1.9 O
37 1.7927.15 1
72 3,4735,13 1
76 364 35.80 0.5
f,
E, Shesi
S i m Shaw mm Layea MPa GPa rebat
2
2
1
1
1
1
170
170
170
1
1
1
490 200
490 200
490 2 0
Naationi: 4 = aggfcgpLa diamder, t, = stab titidoicss, 4 = alab e c f d v a depih, 1. = slab opan, fa = m u d o 00m~ramive~ron@h,
fa = amcdctcnsiladm#,
O
O
O
=
Table A.8. Output Daia for Steei Strn Neuril Ntwork (NN4) Training and Test Cases
TRAINING CASES
Y~dd
NN4
radius, a b b
-Case# Slabt mm x10-~ xlU3 XIO-~
Marzouk and Hussein, 1991
1
3
252
2
344
Emam et A,L995
3
Ml
350
7
8
9
10
14
15
M5
M6
M7
200
425
48
580
600
200
MS
Ml3
MI4
TEST CASES
NNQ
case# sm#
Yieid
radius,
mm
Emam et al., 1995
1
M2
385
2
Ml1 550
3
Ml2 550
k x d
Emarra G-
x ~ o -x~~ o -x~~ o - ~
3230 2030 1700
3730 286 1710
5590 3520 1240
Naea ~ = o m x ~ @ w h m n f i i c c ; ~ = m a x
Prain@mi-~=Rmxrtrain@slabadgt
APPENDIX B
TRAINING CASES
Y M
Uiilmate
Yleld
NNLoadNNtd. ~
Uilrnale
270.307
167.702
248092
258570
288.148
3% 164
274 748
430.442
570.857
153967
225.103
370.788
364414
201.075
348 313
380.327
427.801
453842
210462
TEST CASES
=O32
275988
62.07
330367
445562
370.662
5.0074
8.7837
8.1084
8.7092
10858
13.791
336.391
398.252
163.6
474.195
028.070
431 487
88
17.283
25.M
17.087
20335
17.332
087
-15.49
48 83
-37.87
-4303
-20 60
2.19
-0.74
0.40
4.44
-1.96
O 51
5961
-3325
32.40
2301
4.27
-0.38
1 40
O73
-38.98-0.74
44.01 4 62
15.05
32.87
0.11
2.27
18.53
5.18
4.78
4,12
661
3.75
0.26
21.04
Oversll Average Emn (TefAlng)
Wtd A b . Errw
Table 8.2. RESULTS FOR OPTIMUM CRACK PATTERN NEURAL NETWORK (NNZb)
0,M)
0.28
2.70
0.00
0.11
0,12
0,01
0.03
0,04
0.00
0.00
0.00
0.OB
0,M
0.02
0,oe
2.45
17.50
27.83
0.11
84.02
0.81
0.00
0.00
0.00
0.03
2.64
0,04
Omal Avg A h E m (Tmhhg)
TEST CASES
-28.83
213.00
0.07
88.74
-13.00
144.80
-358.28
12.H
m,12
0.01
11.9s
2.74
28.18
180.54
TRAINING CASES
801.76
j203.28
1718.98
342.49
253.18
286.22
510.25
B A
-2.53
-2.03
17.05
-23.73
4.03
26.15
-10.64
M
-0.98
-0.70
d F
-1.76
21,41
-187.57
0.80
173.06
-3.28
4-99
-02.49
-3.18
93.78
136
0.75
K
0,10
0.08
0.08
0.03
0.78
25.11
0,11
M c
0.22
0.27
0.53
31.73
1.27
28,m
057
1.58
0.28
2.m
17.68
0.57
0.14
0,14
OvenllAwrageAboduteEm(Tnlnlng)
w m
0.13
0.13
0.63
1QA7
0.55
16.28
0.20
5.21
TEST CASES
074.20
713.16
37.80
2.77
35.05
8.41
42.26
49.78
Owr~ll
A~er~g
Abdute
e
Error(Te6tin~)
WelgMed Average Abwluts Error
25.21
33.40
29.35
22.10
Table B.5. RESULTS FOR OPTIMUM STEEL STRAlN DISTRIBUTION NEURAL NETWORK (NN4)
TEST CASES
350591
16597
434434
644
47M54
288271
2714 40
1
2
1WBl
MgO
1
6808
4041
-WB1
45271
15775
d l 4 34
4
145d0
-11801
as000
APPLIED
IMAGE,lnc
fi 1653 East Main Street
-- -----*
- 0