You are on page 1of 95

ABSTRACT

This paper presents the development of Artificial Neural Network (ANN) model

for prediction of ultimate capacity of driven precast concrete piles based on Static and

Dynamic Load Test data. 50 sets of data, obtained from various published literature, were

selected and used for training the ANN models. Input considered in modeling are pile

geometrical characteristics (length, cross-sectional are, as well as shape) and soil

characteristics (SPT value and cohesion). The results indicate that the ANN model serves

as a reliable tool to predict ultimate capacities of piles with an overall coefficient of

correlation (R) values 0.99272, mean squared error (MSE) of 902.12 and average

percentage error of 4.81%. The weights and biases produced from the best model was

used to formulate an equation with the aid of Microsoft excel. In addition, sensitivity

analysis was conducted by using parametric testing in which shows that cohesion and

length are dominant factors in the proposed predictive model.

Keywords: Pile foundation, ultimate capacity, artificial neural network,

backpropagation, driven precast concrete piles

1
Chapter 1

INTRODUCTION

1.1 Background of the Study

When unsuitable soils are present in the near surface and soil conditions are not

capable to support shallow foundations, piles will serve as foundation which can transmit

loads to deeper and more competent strata that has a high bearing capacity. Piles and pile

foundations have been in use since prehistoric times. The Roman wooden piles are

classic examples for this [1]. Nowadays, piles are usually made of concrete or steel.

Piles are classified according to method of installation – driven piles, cast-in-situ

piles and driven cast-in-situ piles. Driven piles maybe of prefabricated concrete, steel or

timber and are driven vertically by using a pile hammer. Cast-in-situ piles are generally

concrete material constructed by drilling holes in the ground up to the necessary depth

and then concrete is filled in to the hole. Driven cast in-situ concrete piles are formed

through a close-ended steel shell or concrete casing that is driven into the ground and

then concrete is placed. The shell may be either left in position or withdrawn for reuse

while the concrete is being placed.

The stability and safety of buildings constructed by using pile foundations depend

on the capacity of piles. The piles’ capacity mainly depends on its size, shape, material,

and method of piles installation. A second factor on its capacity is the supporting

strength of the soil. When the load is transmitted from the piles to the soil, the soil

2
provides friction between the soil and the sides of the pile and/or the load can be directly

transmitted to the soil just below the piles’ tip.

The prediction of axial pile capacity is a complex problem in geotechnical

engineering. At this time, it is very difficult to recommend any single approach as being

the more applicable for estimating the axial bearing capacity of piles. While scientific

approach provides the framework within which they are to be considered, design

calculation still rely on empirical correlation. The most reasonable approach seems that

of going on a load test which is believed to be the only method than can give reliable

capacity of piles [1].

Load tests are conducted to determine the settlement of pile under working load

and its ultimate axial capacity. Conversely, the results of the load tests do not show

distinct plunging ultimate load, therefore the results need interpretation to estimate pile

capacity or ultimate load [2].

Artificial intelligence (AI) has provided a great development in prediction and

modeling of complex behavior of geotechnical engineering materials compared to

traditional empirical methods. Over the past few years, AI has been applied successfully

to various in geotechnical engineering problems including prediction of piles bearing

capacity. There are several available AI techniques that are used today and one example

of it is artificial neural network (ANN) which is mathematical models, inspired from the

brains of certain information-processing characteristics, producing meaningful solutions,

which fall beyond the reach of conventional digital computers.

Artificial Neural Networks (ANN) have different kinds of training or learning

algorithms which include Back Propagation Algorithm (BPA). Among these, Back

3
Propagation Network (BPN) is considered to be the quintessential neural networkbecause

it is commonly used to determine the weights and biases of the network which will

narrow down the difference between desired output and achieved system output in neural

network.

In geotechnical engineering, there are several methods and approaches that were

developed to overcome the uncertainty in determining the ultimate axial capacity of piles.

The methods include empirical approaches regarding soil stratigraphy, soil-pile structure

interaction and the distribution of soil resistance along the pile shaft and tip. Thus, the

quantitative values obtain from these approaches cannot be used directly in pile

foundation design [3]. According to Osman (2013), the best method to determine the

bearing capacity of piles is to conduct a pile load test. However, performing load tests are

costly and requires considerable time and effort. Due to rapid improvement of

instruments used in in-situ pile testing, the application of such techniques has increased

for geotechnical design.

Implementing feed forward back-propagation artificial neural network (ANN) for

solving geotechnical problems has recently gained attention mainly due to its ability in

finding complex nonlinear relationships among different parameters. In this study, an

ANN-based predictive model for estimating axial bearing capacity of piles and its

distribution is proposed.

4
1.2 Significance of the Study

To all of the inconveniences that these pile construction problems had given to the

community, the quality of pile installation reacts on the load distributed through it

without the circumstances that it can change the lateral behavior of pile. One of the main

reasons to have such poor quality of pile system is the fact that development models of

the driven pile system which are reliable to foundations tasked to be inconvenient within

its installation damages, were not locally settled up but rather adopted from other

countries that gave the process of construction the way it has to be.

With the help of this study, the researchers will be able to promote the

development and usage of a model that has parameters that came from the place where it

will be used. In this case, the government will be able to determine different pile

construction design that needs to be applied for future construction. It will help them to

lessen the additional expenses allotted for the construction of piles.

Accordingly, the community will be able to maximize the efficiency of each

individual because geotechnical engineers will be able to minimize the amount of time

wasted due to construction issues. The wasted time will now be allotted to pile

construction purposes for more process would be done.

Consequently, this study can serve as a guide for predicting ultimate capacity of

driven precast concrete piles. Civil engineers who are in line with this field can use the

technique provided by the researchers to lessen the number of unnecessary pile

construction design.

5
1.3 Objectives

This study aims to create an artificial neural network model in predicting ultimate

capacity in driven precast concrete piles in different cases satisfying the given parameters

of the study. Specifically, it aims to:

1. develop a model for predicting ultimate capacity of driven precast concrete piles

2. perform a sensitivity analysis using parametric testing to determine the

importance of each input parameter on the ultimate axial capacity of piles

3. validate the accuracy of the derived model using coefficient of determination, R-

squared value

4. develop a program with the final weights and biases that will easily predict the

ultimate capacity of piles in terms of kilonewtons with the input values such as

pile depth, pile cross section, standard penetration test, and cohesion of the

precast concrete pile with the aid of Microsoft Excel

6
1.4 Scope and Limitations

In accordance with its objectives, this study encompasses the development of a

model for predicting ultimate capacity using neural network on driven precast concrete

piles. Data, codes and restrictions, and pile load tests to be used for the model are

extracted from different published case studies and geotechnical investigations that will

be listed in the references. The study focused on reports regarding cohesive soils only.

Matrix Laboratory (MATLAB) version 2015a is used as a Neural Network Tool

in designing the model for predicting scaling distress on rigid pavements. Back

Propagation Algorithm is the system that is used in computing such gradient which

means hidden connections between parameters that lessen the percentage error of the

model. To initialize the design, the software is in need of input values such as pile depth,

pile cross section, shape, standard penetration test results, cohesion of the soil, and the

actual ultimate capacity of the driven piles. The transfer function used is hyperbolic

tangent sigmoid function (TANSIG), with 1 hidden layer. The number of neurons in the

hidden layer ranges from 5 to 15 nodes. The minimum acceptable R value of each

network is set to 0.9, according to a study conducted by Tarawneh which he used ANN

for geotechnical reports.

In designing a model in predicting ultimate capacity of piles, the study focuses

only with a certain type of pile type and pile material which is driven pile and concrete

respectively. It did not consider other type of pile and pile materials. The parameters that

were used in study is limited to the four input values which are pile depth, pile cross

section, shape, standard penetration test, and cohesion. The impact of the study is

determined in terms of the improvement of the determination of the ultimate capacity of

7
driven piles in terms of effort, and time efficiency of computing the ultimate capacity of

piles.

With the final weights and biases of the predicting model, the equation that was

formulated is used to develop a program which can easily predict the ultimate capacity in

driven piles by feeding the program with the input parameters. The software that is used

in creating the program is Microsoft Excel.

1.5 Conceptual Framework

This study is conducted under a particular order as shown in Figure 1.

Preliminary Research about Artificial Neural Network, along with the standards and

specifications provided by the published studies and geotechnical investigations is the

first major step in the run of the study. Gathering of Data for the input values as

parameters such as pile depth, pile cross section, standard penetration test, and cohesion

coming from different cases of the study.

After determining the parameters, the model is constructed using MATLAB and

input values are entered in the model. Once all inputs are done, the weights and biases are

generated. It is followed by the predicting model and the derivation of equation. Then the

equation is used in Microsoft Excel to create a program that easily predicts the ultimate

capacity of precast concrete piles.

8
INPUT PROCE OUTP
SS UT
DATA GATHERING MATRIX
WEIGHTS AND BIASES
LABORATORY

▪ Pile Length
▪ Pile Cross
Section
▪ Pile Shape DERIVATION OF
BACKPROPAGATION
▪ Standard EQUATION
ALGORITHM
Penetration
Test
▪ Cohesion
▪ Ultimate
PROGRAM
Axial
Capacity

Figure 1 Conceptual Framework

9
1.6 Definition of Terms

Artificial Neural Network - are a family of models inspired by biological neural

networks which are used to estimate or approximate functions that can depend on a large

number of inputs and are generally unknown. Artificial neural networks are generally

presented as systems of interconnected "neurons" which exchange messages between

each other. The connections have numeric weights that can be tuned based on experience,

making neural nets adaptive to inputs and capable of learning.

Axial Load - is a force administered along the lines of an axis.

Backpropagation - is a common method of training artificial neural networks used in

conjunction with an optimization method such as gradient descent. The method calculates

the gradient of a loss function with respect to all the weights in the network. The gradient

is fed to the optimization method which in turn uses it to update the weights, in an

attempt to minimize the loss function.

Foundation - is the element of an architectural structure which connects it to the ground,

and transfers loads from the structure to the ground. Foundations are generally considered

either shallow or deep. Foundation engineering is the application of soil mechanics and

rock mechanics (Geotechnical engineering) in the design of foundation elements of

structures.

Hidden Layer – is a part of the network next to the input layer where the transfer

function is applied. The hidden layer can be one up to many layers which have a number

of neurons each.

10
Input Layer – is a part of the network which shows the data that is fed to the network in

order to produce weights and biases that corresponds to the output layer. The summation

of the input multiplied by its corresponding weights and biases will proceed to the hidden

layer.

Matrix Laboratory - is a multi-paradigm numerical computing environment and fourth-

generation programming language. A proprietary developed by MathWorks, MATLAB

allows matrix manipulations, plotting of functions and data, implementation of

algorithms, creation of user interfaces, and interfacing with programs written in other

languages, including C, C++,Java, Fortran and Python.

Output Layer –is the summation of the hidden layer multiplied with another set of

weights and biases. The transfer function will be applied again in the matrix generated by

the hidden layer and the weights in order to proceed to the output layer.

Piles - is a long slender foundation member, made either of timber, structural steel or

concrete which might be cast-in-situ or driven and acts as a structural member to transfer

the load of the structure to a required depth in deep foundations carrying a load which

may be vertical or lateral or lateral plus vertical.

Parameters - is any characteristic that can help in defining or classifying a particular

system.

Regression - is a statistical process for estimating the relationships among variables. It

includes many techniques for modeling and analyzing several variables, when the focus

is on the relationship between a dependent variable and one or more independent

variables.

11
Sigmoid Function - is a bounded differentiable real function that is defined for all real

input values and has a positive derivative at each point.

Soil Cohesion - is the component of shear strength of a rock or soil that is independent of

interparticle friction.

Standard Penetration Test - is an in-situ dynamic penetration test designed to provide

information on the geotechnical engineering properties of soil.

Transfer Function - is the ratio of the output of a system to the input of a system, in the

Laplace domain considering its initial conditions and equilibrium point to be zero. It is

also used in the analysis of systems such as single-input single-output filters, typically

within the fields of signal processing, communication theory, and control theory.

Ultimate Capacity - is the capacity to support the loads applied to it. The capacity is the

maximum average contact pressure between the elements in which should not produce

shear failure.

12
13
Chapter 2

REVIEW OF RELATED LITERATURE

In this chapter, the definition, classification, and process of determining capacity

of piles is discussed. The methods that can be used in this study is briefly described.

Moreover, this chapter explains the Artificial Intelligence (AI) and its examples and the

software to be used in the study.

2.1 Piles

Pile foundation is necessary when the soil bearing capacity in the near surface is

not adequate enough to withstand the foundation loads of the structure for the design of

shallow foundations. The requirement of pile foundation depends to the soil condition,

site and operational conditions as well as the foundation loads. The need of piles

foundation, as well as the type of foundation to be used and the required dimensions will

be based on the site investigation report. The accumulation of cost statistics for

projections and comparison of pile foundation should be accomplished before selecting

the type of pile to be used.

2.2 Classification of Piles

The classification of piles is categorized into three major divisions. These

divisions Piles are classification according to functional behavior, type of material and

method of installation.

14
2.2.1 According to functional behavior

Piles can be classified according to load transfer and functional behavior

which fall into three broad categories – end bearing, skin friction piles and friction

cum end bearing piles.

2.2.1.1 End bearing piles

15
The end bearing pile transmit the foundation loads through the hard rock

strata or high density sand and gravel. Most of the carrying capacity of these piles

is resisted at the toe of the pile (pile tip) (see Figure 2). It is design as a column

since its behavior is the same as a column. Conversely, piles do not have a failure

due to buckling and since the shaft of the pile is surrounded and by that, supported

by soil. Buckling is only considered if part of a pile is exposed either in air or

water hence, the pile is unsupported.

Figure 2 End Bearing Capacity (Engineering Articles, 2014)

2.2.1.2 Skin friction piles

The load in this type of pile is transferred mainly along the side of the pile

(pile shaft) through skin friction resistance. as shown in Figure 3. Skin friction

16
piles must be in required length, in order to have an adequate load-carrying

capacity that is mainly dependent to the shaft of the pile in contact with the soil.

Figure 3 Skin Friction Piles (Engineering Articles, 2014)

2.2.1.3 Friction cum end bearing piles

17
In the majority of cases, piles are classified as friction cum end bearing

pile where the load-carrying capacity is dependent on both end bearing and shaft

friction as shown in Figure 4.

Figure 4 Friction cum End Bearing Piled (Engineering Articles, 2014)

2.2.2 According to type of material

Piles are usually made of timber, concrete or steel. Timber as material for

piles is used for temporary construction subsequently it has the most economical

price. Concrete is used in the manufacturing of cast in-situ piles as well as pre-

cast and pre-stressed concrete piles. Steel piles, on the other hand, are used for

permanent structures hence it is the most expensive among the two materials.

Table 1 shows the classification of piles based on type of material.

Table 1 Classification of Pile Based on Material Type (Engineering Articles, 2014)

18
Material Characteristics

▪ Timber piles are made of tree trunks driven with small end as
Timber
a point
▪ Maximum length: 35 m; optimum length: 9 to 20m
▪ Max load for usual conditions: 450 kN; optimum load range =
80 to 240 kN
▪ Disadvantages: difficult to splice, vulnerable to damage in
hard driving, vulnerable to decay unless treated with
preservatives (If timber is below permanent W.T. it will
apparently last for ever), if subjected to alternate wetting &
drying, the useful life will be short, partly embedded piles or
piles above W.T. are susceptible to damage from wood borers
and other insects unless treated.
▪ Advantages: comparatively low initial cost, permanently
submerged piles are resistant to decay, easy to handle, best
suited for friction piles in granular material.

▪ Max length: practically unlimited,


Steel
▪ optimum length: 12 to 50 m
▪ load for usual conditions = maximum allowable stress ´ x-
section area,
▪ Optimum load range = 350 to 1050 kN
▪ The members are usually rolled HP shapes/pipe piles. Wide
flange beams & I beams proportioned to withstand the hard
driving stress to which the pile may be subjected, In HP pile
the flange thickness= web thickness, pipe piles are either
welded or seamless steel pipes, which may be driven either
open ended or closed end. Closed end piles are usually filled
with concrete after driving. Open end piles may be filled but
this is not often necessary.
▪ Advantages: easy to splice, high capacity, small
displacement, able to penetrate through light obstructions,
best suited for end bearing on rock, reduce allowable capacity
for corrosive locations or provide corrosion protection.
▪ Disadvantages: Vulnerable to corrosion, HP section may be
damaged/deflected by major obstruction

19
▪ Concrete piles may be precast, prestressed, cast in place, or of
Concrete
composite construction.
▪ Precast concrete piles may be made using ordinary
reinforcement or they may be prestressed.
▪ Precast piles using ordinary reinforcement are designed to
resist bending stresses during picking up & transport to the
site & bending moments from lateral loads and to provide
sufficient resistance to vertical loads and any tension forces
developed during driving.
▪ Prestressed piles are formed by tensioning high strength steel
prestress cables, and casting the concrete about the cable.
When the concrete hardens, the prestress cables are cut, with
the tension force in the cables now producing compressive
stress in the concrete pile. It is common to higher-strength
concrete (35 to 55 MPa) in prestressed piles because of the
large initial compressive stresses from prestressing.
Prestressing the pile tends to counteract any tension stresses
during either handling or driving.
▪ Max length: 10 to 15 m for precast, 20 to 30 m for prestressed
▪ Optimum length: 10 to 12 m for precast, 18 to 25m
prestressed
▪ Loads for usual conditions 900 for precast, 8500 kN for
prestressed
▪ Optimum load range: 350 to 3500 kN
▪ Disadvantages: difficult to handle unless prestressed, high
initial cost, considerable displacement, prestressed piles are
difficult to splice.
▪ Advantages: high load capacities, corrosion resistance can be
attained, hard driving possible
▪ Remarks: cylinder piles in particular are suited for bending
resistance.
▪ Cast in place concrete piles are formed by drilling a hole in
the ground & filling it with concrete. The hole may be drilled
or formed by driving a shell or casing into the ground.
▪ Disadvantages of Concrete piles: Concrete piles are
considered permanent, however, certain soil (usually organic)
contain materials that may form acids that can damage the
concrete. Salt water may also adversely react with the

20
concrete unless special precautions are taken when the mix
proportions are designed. Additionally, concrete piles used
for marine structures may undergo abrasion from wave action
and floating debris in the water. Alternate freezing & thawing
can cause concrete damage in any exposed situation.

▪ In general, a composite pile is made up of two or more


Composite
sections of different materials or different pile types. The
upper portion could be cased cast-in-place concrete combined
with a lower portion of timber, steel H or concrete filled steel
pipe pile. These piles have limited application and are
employed under special conditions.

2.2.3 According to installation method

A simplified division of driven and bored is often employed to piles.

2.2.3.1 Driven Piles

Driven piles are considered to be displacement piles. These piles are

constructed by using a pile driving machine in which soil is removed radially as

the pile is driven into the ground (refer to Figure 5).

21
Figure 5 Installation of Driven Pile (Grimtech India, 2015)

2.2.3.2 Bored Piles

Bored piles are generally considered to be non-displacement piles. These

piles are produced after a void is formed in the ground by using bored piling

machine. Reinforcement is then placed into the void and concrete will be filled

(see Figure 6). The casing or borehole walls is only temporary and maybe

removed immediately when the soil is stable.

22
Figure 6 Installation of Bored Pile (Arabian Boring General Contracting, 2015)

2.2.3.3 Comparison of Driven and Bored Piles

Comparison of driven and bored piles are shown in Table 2.

Table 2 Comparison of Driven and Bored Pile (Pile Driving Construction


Association, 2013)

Driven Pile Bored Pile

➢ Length revisions are ➢ Length revisions are

relatively easy feasible but take

Constructability Issues (splicing or cut-off time (cage

of steel piles, cut-off modifications)

of PSC piles). ➢ Minimum

23
➢ Minimum penetration

penetration requirements are

requirements (e.g., typically not an issue

consideration of

scour and

liquefaction

potential, fixity,

uplift) may be hard

to meet

➢ Auger cuttings
➢ Driving resistance
➢ Observation of
➢ Transferred hammer
bottom cleanliness
energy/hammer
Construction Feedback ➢ Concrete volume
performance
➢ Shaft
➢ Driving stresses
profile/geometry
➢ Pile integrity
(possible but not
➢ Capacity
widely done)

➢ Smaller elements ➢ Bigger elements

➢ Lower capacity ➢ Higher capacity


Redundancy and Loading
Issues ➢ Lower cost ➢ Higher cost

➢ More elements used ➢ Fewer elements used

24
➢ Highly redundant ➢ Little to no

redundancy

➢ Delivered free of

defects ➢ Defects may occur

➢ May be damaged during construction

during or ➢ Damage after

infrequently, after construction is


Integrity and Reliability installation possible but rare
Issues
➢ Damage is typically ➢ Defects are not

easily detectable – A easily quantified

Driven Pile is a ➢ Repair is generally

Tested Pile very complex

➢ Repair generally

2.3 Pile Load Test

Performance testing of pile foundation is necessary since piles are used generally

to support critical foundation loads. According to Osman (2013), the best method to

determine the bearing capacity and behavior of piles is by performing a pile load test. Pile

load tests are generally performed to either prove that piles are capable of sustaining the

design load or to gain more detailed information that will enable a more efficient design.

This method is applied to either evaluate whether the piles are suitable to sustain the

design load or to gain an information that will lead to a more efficient and improved

25
design. The two main types of pile testing which are commonly used to measure the load

capacity of piles are the static and the dynamic load test.

Performing pile load tests are typically done to prove that piles are capable of

sustaining the ultimate design load ("proof test") or to to gain an information that will

lead to a more efficient and improved design ("load-deformation test"). In conducting a

proof test, the deflection is determined at the head of the test pile after the ultimate design

load hence (allowable design load multiplied by factor of safety) is loaded. The design of

the test pile is considered to be accepted if the deflection is within the required levels.

Proof tests are generally performed during construction as the piles are installed.

When the test procedure is conducted during the period of designing the piles

prior to actual construction, it is called load-deformation tests. In load deformation

testing, the deformation is measured every time a load is applied along the pile shaft, as

well as at the pile head and tip until the pile reaches its failure. Conducting this method

will give a detailed load-deformation data which reflects specific properties of the

construction site that enables a reduced value of factor of safety and provide more

efficient design.

26
2.3.1 Static Load Test

The most common method in determining pile foundation capacity is by

static loading. ASTM Standard D1143 have provided several procedures in

setting up and performing a static loading test. The Osterberg Cell (O-CellTM)

testing, which was recently developed form of static loading test, has been

frequently used for testing drilled piles.

2.3.2 Dynamic Load Test

Since static loading test is expensive and not time efficient, dynamic

testing has been an alternative method for evaluating the shaft resistance and end

bearing capacity of piles. ASTM Standard has also provided procedures for

dynamic testing.

2.4 Artificial Intelligence

It refers to computer the ability of a digital computer or computer-controlled robot

to perform tasks commonly associated with intelligent beings. The term "intelligence" is

difficult to define, and has been the subject of heated debate by philosophers, educators,

and psychologists for ages. Nevertheless, it is possible to enumerate many important

characteristics of intelligent behavior. Intelligence includes the capacity to learn, maintain

a large storehouse of knowledge, utilize commonsense reasoning, apply analytical

abilities, discern relationships between facts, communicate ideas to others and understand

communications from others, and perceive and make sense of the world around us. Thus,

artificial intelligence systems are computer programs that exhibit one or more of these

27
behaviors (David et. al, 1993). Still, despite continuing advances in computer processing

speed and memory capacity, there are as yet no programs that can match human

flexibility over wider domains or in tasks requiring much everyday knowledge. On the

other hand, some programs have attained the performance levels of human experts and

professionals in performing certain specific tasks, so that artificial intelligence in this

limited sense is found in applications as diverse as medical diagnosis, computer search

engines, and voice or handwriting recognition. (Copeland, 2000) There are a lot of

examples of AI which are being used nowadays. One of which is the Artificial Neural

Network (ANN) that is known to be the most commonly used AI in the world.

2.4.1 Artificial neural network (ANN)

An Artificial Neural Network (ANN) is an information processing

paradigm that is inspired by the way biological nervous systems, such as the

brain, process information. The key element of this paradigm is the novel

structure of the information processing system. It is composed of a large number

of highly interconnected processing elements (neurons) working in unison to

solve specific problems. ANNs, like people, learn by example.

There are three layers in ANN that are interconnected. The first layer

consists of input neurons. Those neurons send data on to the second layer

(hidden), which subsequently relay the output neurons to the third layer.

During processing, each neuron performs a weighted sum of inputs from

the neurons connecting to it; this is called activation. The neuron chooses to fire if

28
the sum of inputs exceeds some previously set threshold value; this is called

transfer (Davis et. al., 1993).

Inputs with high weights tend to give greater activation to a neuron than

inputs with low weights. The weight of an input is analogous to the strength of a

synapse in a biological system. In biological systems, learning occurs by

strengthening or weakening the synaptic connections between nerve cells. An

artificial neural network simulates synaptic connection strength by increasing or

decreasing the weight of input lines into neurons (Davis et. al., 1993).

Neural networks are trained with a series of data points. The networks

guess which response should be given, and the guess is compared against the

correct answer for each data point. If errors occur, the weights into the neurons

are adjusted and the process repeats itself. This learning approach is called back

propagation, and is similar to statistical regression.

Current applications of neural networks include: oil exploration data

analysis, weather prediction, the interpretation of nucleotide sequences in biology

labs, and the exploration of models of thinking and consciousness. An example of

a network architecture is shown in Figure 7.

29
Figure 7 Network Architecture of Artificial Neural Network (ANN)

2.5 Matrix Laboratory (MATLAB) as a Neural Network Tool

Neural Network Toolbox provides algorithms, functions, and apps to create, train,

visualize, and simulate neural networks. You can perform classification, regression,

clustering, dimensionality reduction, time-series forecasting, and dynamic system

modeling and control. Neural Network Toolbox supports supervised learning with feed

forward, radial basis, and dynamic networks. It provides Preprocessing, post-processing,

and network visualization for improving training efficiency and assessing network

performance. It also supports unsupervised learning with self-organizing maps and

competitive layers (The MathWorks Inc., 1994). Anyone can use Neural Network

Toolbox for applications such as data fitting, pattern recognition, clustering, time-series

30
prediction, and dynamic system modeling and control. Figure 8 shows the Network

Architecture of the Neural Network Toolbox in MATLAB.

Figuer 8 Network Architecture of the Neural Network Toolbox in MATLAB, (The


MathWorks Inc., 1994)

2.6 Microsoft Excel


As one of the specific objectives of the study which is to develop a program on

which ultimate bearing capacity of precast concrete driven piles will be easily predicted,

Microsoft Excel will be used as a programming environment. Figure 9 shows the network

architecture of Microsoft Excel.

31
Figure 9Microsoft Excel’s Network Architecture

32
2.7 Related Studies

These studies aimed to present a feedback mechanism about the process that the

project will undergo. It also shows how feasible the different things can be done through

any type of methods that had been portrait at the beginning. Different complications were

shown below including the application of Artificial Neural Network, and ANN using

different sample sizes.

2.7.1 Application of Artificial Neural Network for Predicting Shaft and Tip
Resistances of Concrete Piles (Momeni, et al., 2015)

In this study, an ANN-based predictive model for estimating ABC of piles

and its distribution is proposed. Axial bearing capacity (ABC) of piles is usually

determined by static load test (SLT). However, conducting SLT is costly and

time-consuming. High strain dynamic pile testing (HSDPT) which is provided by

pile driving analyzer (PDA) is a more recent approach for predicting the ABC of

piles. In comparison to SLT, PDA test is quick and economical. Implementing

feed forward back-propagation artificial neural network (ANN) for solving

geotechnical problems has recently gained attention mainly due to its ability in

finding complex nonlinear relationships among different parameters.

The main objective of this study is to propose an ANN-based predictive

model of bearing capacity using real PDA and site investigation data. The

predictive model is built for predicting shaft, tip and ultimate resistances (Qs, Qp

and Qu) of piles. Nevertheless, it is worth mentioning that this study uses the Case

33
Pile Wave Analysis Program (CAPWAP) predicted pile bearing capacity rather

than the determined bearing capacity of piles through SLT. The tested piles were

reinforced and pre-stressed concrete piles with different diameters, lengths, sets

and SPT (N) values around the pile shaft and tip. Most of the tests were conducted

in cohesion less soils. The PDA test results include ultimate bearing capacity of

piles, pile shaft and tip resistances.

Among researchers who have addressed ANNs for predicting bearing

capacity of pile foundations, Goh (1995; 1996) developed an ANN-based

predictive model to estimate the ultimate load capacity of driven piles in sandy

soils. His findings suggest that compared to conventional methods of pile bearing

capacity estimation, ANN-based predictive model works better.

In another study, Lee and Lee (1995) employed ANN for estimation of

pile bearing capacity. Their study focused on small scale laboratory tests where

the horizontal and vertical chamber pressure, the number of hammer blows, pile

penetration depth ratio, and mean normal stress of the soil were set as inputs of

the network model while the ultimate bearing capacity was selected as the model

output. According to their conclusion, ANN can provide good prediction

performance in bearing capacity problems. Teh et al. (1997) also addressed the

workability of neural network for predicting the pile bearing capacity.

Abu-Kiefa (1998) implemented ANN to predict the ABC of driven piles in

cohesion less soils. For network construction purpose, he compiled the data of 59

recorded cases of good-quality pile load tests. In his study, friction angles of the

soil, the effective overburden pressure around the tip of the pile, the length of the

34
pile and its equivalent cross-sectional area were considered as input layers of the

ANN model. His conclusion showed the feasibility of ANN for predicting shaft

and tip resistances of piles.

35
At the end of this study, the reliability of the ANN-based predictive model

of bearing capacity can be seen in Figures 10 to 12. These figures show the

predicted Qs, Qp and Qu of piles versus their measured values. Figure 2.9 shows

a comparison between predicted and measured Qs for training and testing data.

The obtained R² values equal to 0.999 and 0.941 suggest the reliability of the

model in predicting Qs. Similarly Figure 2.10 suggests that the predicted Qp is in

good agreement with the measured Qp. As displayed in Figure 11-b, the

coefficient of determination equals to 0.936 for testing data recommends the

feasibility of the ANN-based predictive model of bearing capacity. In Figure 12, a

comparison is made between the measured and predicted Qu of piles for both

training and testing data. Coefficient of determination equals to 0.951 for testing

data suggests that the ANN-based predictive model is good enough in capturing

mate bearing capacity of piles.

Figure 10 Shaft Bearing Capacity (Qs) of Piles Predicted by ANN Model Versus
their Measured Values (Momeni, et al., 2015)

36
Figure 11 End-Bearing Capacity (Qp) of Piles Predicted by ANN Model Versus
their Measured Values (Momeni, et al., 2015)

Figure 12 Ultimate Capacity (Qu) of Piles Predicted by ANN Model Versus their
Measured Values (Momeni, et al., 2015)

2.7.2 Training an Artificial Neural Network (Davis et al., 1993)

In the training phase, the correct class for each record is known, and the

output nodes can therefore be assigned "correct" values -- "1" for the node

corresponding to the correct class, and "0" for the others. It is thus possible to

37
compare the network's calculated values for the output nodes to these "correct"

values, and calculate an error term for each node. These error terms are then used

to adjust the weights in the hidden layers so that, hopefully, the next time around

the output values will be closer to the "correct" values.

A key feature of neural networks is an iterative learning process in which

data cases (rows) are presented to the network one at a time, and the weights

associated with the input values are adjusted each time. After all cases are

presented, the process often starts over again. During this learning phase, the

network learns by adjusting the weights so as to be able to predict the correct

class label of input samples.

Sample size for training can range from 5 training samples up to hundreds

of data. Table 3 shows previous research studies which includes application of

artificial neural network and their respective samples used for training.

Table 3 Samples Size of Artificial Neural Network Studies

38
2.7.3Pipe Pile Setup: Data base and prediction model using artificial neural

network (Tarawneh, 2013)

Over the last few years, artificial neural networks (ANNs) have been

applied to many geotechnical engineering problems with some degree of success.

With respect on the design of pile foundations, the ability to accurately predict

pile setup may lead to more economical pile design, resulting in a reduction in

pile length, pile section, and size of driving equipment. In this paper, an ANN

model was developed for predicting pipe pile setup using 104 data points,

obtained from the published literature and the author's own files. In addition, the

paper discusses the choice of input and internal network parameters which were

examined to obtain the optimum ANN model. Finally, the paper compares the

predictions obtained by the ANN with those given by a number of empirical

formulas. It is demonstrated that the ANN model outperforms the examined

empirical formulas.

For the one hidden layer models, Table 4 shows the results of the best

performing models. The RMSE values ranged from 548.72 to 1647.89, while the

coefficient of correlation (R) values were between 0.69 and 0.96 for the testing

data set. Model 4 which has four processing elements in the hidden layer,

using hyperbolic tangent as transfer function for both hidden and output layers,

was the best performing among the single hidden layer models.

39
Table 4 Single Hidden Layer ANN Models

ANN models with a single hidden layer


Hidden Layer-1 Output Layer Testing Data
Model Input
Process Transfer Processing Transfer
No. Nodes RMSE R
Elements Function Elements Function
1 5 1 Tanh 1 tanh 962.91 0.77
2 5 2 Tanh 1 tanh 670.75 0.90
3 5 3 Tanh 1 tanh 787.34 0.89
4 5 4 Tanh 1 tanh 584.72 0.96
5 5 5 Tanh 1 tanh 1099.78 0.76
6 5 6 Tanh 1 tanh 858.60 0.85
7 5 7 Tanh 1 tanh 1204.58 0.68
8 5 10 Tanh 1 tanh 1647.89 0.49
9 5 1 Sig. 1 Sig. 835.47 0.83
10 5 2 Sig. 1 Sig. 731.82 0.88
11 5 3 Sig. 1 Sig. 1211.43 0.69
12 5 4 Sig. 1 Sig. 708.51 0.90
13 5 5 Sig. 1 Sig. 845.96 0.84

40
41
Chapter 3

METHODOLOGY

The six purposes of this chapter are to (1) describe the research methodology of

this study, (2) discuss how the variables are determined, (3) explain how the network

works, (4) provide an explanation in selecting the best model that will provide the final

weights and biases, and (5) define how the equation will be formulated.

3.1 Research Design

A correlation research methodology was used in the study. All of the data that

were used in the study were obtained from existing geotechnical reports, case studies and

journals. The six parameters that were in the study are pile length, pile cross section, pile

shape, Standard Penetration Test (SPT), soil cohesion and ultimate capacity of the pile.

The software used is MATLAB. Data were trained several trials until the most favorable

result which will give the highest R – value with the lowest mean square error.

3.2 Data Collection

The data that were gathered by the researchers are the following:

● Pile Length ● Soil Type

● Pile Cross-section ● Ultimate Pile Capacity

● Pile Shape

● SPT

42
3.2.1 Standard Penetration Test (SPT)

The SPT values from the published journals are determined by driving a

sample tube 150 mm into the ground and then the number of blows needed for the

tube to penetrate each 150 mm (6 in) up to a depth of 450 mm (18 in) is recorded.

The sum of the number of blows required for the second and third 6 in. of

penetration is reported as SPT blow count value, commonly termed "standard

penetration resistance" or the "N-value". This data served as one of the parameters

that was fed in the neural network. Figure 13 shows a sample data of soil profile

indicating SPT N-value. However, in this study, a single value of SPT for a

certain pile served as an input to the neural network, thus the weighted average of

SPT was computed.

Figure 13 Soil Profile Sample Data (Design and Construction of Driven Pile
Foundations—Lessons Learned on the Central Artery/Tunnel Project, 2009)
43
3.2.2 Soil Cohesion

The value of soil cohesion depends on the type of soil in the test site as

shown in Table 5. The values given below were derived from Swiss Standard SN

670 010b - Characteristic Coefficients of soils, Minnesota Department of

Transportation - Pavement Design and NAVFAC Design Manual 7.2 -

Foundations and Earth Structures. In this study, the weighted average of the

cohesion from the soil profile was considered. In cases in which the soil profile is

not provided in the paper, the dominant soil was used. This parameter was also

fed in the network.

Table 5 Typical Values of Soil Cohesion for Different Soil Types

Description USCS Cohesion (kPa)

Clayey gravels, clayey sandy gravels GC 20


Silty sands SM 22
Silty sands - Saturated compacted SM 50
Silty sands – Compacted SM 20
Clayey sands SC 5
Clayey sands – Compacted SC 74
Clayey sands -Saturated compacted SC 11
Loamy sand, sandy clay Loam - SM, SC 50-75
compacted
Loamy sand, sandy clay Loam - SM, SC 10-20
saturated
Sand silt clay with slightly plastic SM, SC 50
fines – compacted
Sand silt clay with slightly plastic SM, SC 14
fines - saturated compacted

44
Inorganic silts, silty or clayey fine ML 7
sands, with slight plasticity
Inorganic silts and clayey silts – ML 67
compacted
Inorganic silts and clayey silts - ML 9
saturated compacted
Inorganic clays, silty clays, sandy CL 4
clays of low plasticity
Inorganic clays, silty clays, sandy CL 86
clays of low plasticity – compacted
Inorganic clays, silty clays, sandy
clays of low plasticity - saturated CL 13
compacted
Mixture if inorganic silt and clay – ML-CL 65
compacted
Mixture if inorganic silt and clay - ML-CL 22
saturated compacted
Organic silts and organic silty clays OL 5
of low plasticity
Inorganic silts of high plasticity - MH 10
compacted
Inorganic silts of high plasticity - MH 72
saturated compacted
Inorganic silts of high plasticity MH 20
Inorganic clays of high plasticity CH 25
Inorganic clays of high plasticity – CH 103
compacted
Inorganic clays of high plasticity - CH 11
saturated compacted
Organic clays of high plasticity OH 10
Loam – Compacted ML, OL, MH, OH 60-90
Loam - Saturated ML, OL, MH, OH 10-20
Silt Loam – Compacted ML, OL, MH, OH 60-90
Silt Loam – Saturated ML, OL, MH, OH 10-20
Clay Loam, Silty Clay Loam – ML, OL, CL, MH, OH, CH 60-105
Compacted
Clay Loam, Silty Clay Loam - ML, OL, CL, MH, OH, CH 10-20
Saturated

45
Silty clay, clay – compacted OL, CL, OH, CH 90-105
Silty clay, clay – saturated OL, CL, OH, CH 10-20

3.3 Process

After collecting all of the data, they were imported to the database of MATLAB.

In an excel file format, it was transposed in such a way that rows and columns of the data

were interchanged which were shown in Figure 15 and Figure 16. The Pile length, Pile

cross section, Standard penetration test (SPT), and Soil cohesion were the input

parameters while the output was the ultimate capacity of piles that were shown in Figure

14.

Figure 14Data Reflected in an Excel File (Excel 2007)


46
Figure 15 Creating a Folder for Input and Target Data

Figure 16Data Transposed from Excel to MATLAB (MATLAB 2015a, 2016)

47
Figure 17Command to Start with the Neural Network Toolbar

(MATLAB 2015a, 2016)

Figure 18Network/Data Manager (MATLAB 2015a, 2016)

48
After entering the command function “nntool”, click Import in the Data Manager

tab and select the wanted variable in the list that were shown in Figure 17 and Figure 18.

Also, the user must enter the type of data. Data for input and simulation must be input

data type. Data for target must be output data types which were shown in Figure 19 and

Figure 20.

Figure 19 Importing of Input Data Figure 20 Importing of Target Data

(MATLAB 2015a, 2016) (MATLAB 2015a, 2016)

49
Figure 21Creating a Network (MATLAB 2015a, 2016)

Figure 22Setting Network Parameters (MATLAB 2015a, 2016)

50
Figure 23 Showing the Successful Added Network (MATLAB 2015a, 2016)

Figure 24 Visual Model of the Training Network (MATLAB 2015a, 2016)

And then, click new and choose network. In creating network, there are several

functions to be selected by the user. All input and target data stored/created, can be

selected. The transfer function to be used must be selected also. There are three types of

transfer function to choose from (logsig, tansig and purelin). Adaptation learning function

51
can be selected also (learngd and learngdm). The user can alter the number of hidden

layer and hidden node. After all functions and data are entered, click create. Wherein the

process were shown from Figure 21 to Figure 24.

Figure 25Training Network (MATLAB 2015a, 2016)

Figure 26Neural Network Training (MATLAB 2015a, 2016)

52
When the input and target data were already classified, then proceed to training

the network repeatedly for getting the more precise and efficient output of the product

where Figure 25 and Figure 26 shows training output. The weights and biases of the

desired regression can be viewed at the View/Edit Weights tab which was shown in

Figure 27.

Figure 27 Sample Weights and Biases (MATLAB 2015a, 2016)

3.4 Output

Neural Network training was continuously repeated. It was simulated into

different number of hidden layers. Each simulation was solved for transfer functions

which was the TANSIG functions. After that, the network generated the best predicting

model which satisfies the following conditions (See Figure 28 and Figure 29)

● Lowest Mean Square Error

● Highest overall R – Value

If the model does not satisfy the desired value for overall Regression which is equal to

1.0, the best model that would be selected must have a regression value of at least 0.90. In

53
case that the regression value and mean square error does not satisfy the two given

conditions, model that has lowest MSE value will be considered.

This best predicting model will give the final weights and biases. The final

weights and biases will be used in creating the equation that will quantify the occurrence

of ultimate capacity of precast driven piles. The following figures show the process in

selecting the best predicting model.

Figure 28Performance (MATLAB 2015a, 2016)

54
Figure 29Regression (MATLAB 2015a, 2016)

3.5 Architecture of Neural Network

The researchers will set Cohesion, SPT, Length, Area and Shape as input

parameters while axial capacity as target output (see Figure 30)

Figure 30 Architecture of Neural Network

55
3.6 Parametric Testing

The model that is created was tested using parametric testing in order to verify the

effect in the output with respect to the individual parameters. Parametric testing is

conducted by assigning a constant value at all parameters except the variable that is being

tested. The relationship of each variable to the expected output can be observed by

adjusting the value of the parameter and checking its trend.

3.7 Analysis of Results

Having the final weights and biases from the best predicting model, such weights

and biases were used to formulate an equation aligned with the fourth specific objective

of the study which is to formulate an equation with the final weights and biases that will

easily predict the ultimate capacity of precast driven piles in terms of percentage with the

input values such as Pile length, Pile cross section, Standard penetration test (SPT), and

Soil cohesion of the pile. It will be introduced to Microsoft Excel and later on, a program

will be created that that will easily predict the ultimate capacity of precast driven piles in

terms of percentage.

56
57
Chapter 4

RESULTS AND DISCUSSIONS

This chapter explains how the researchers came up with the final weights and

biases of the best model developed as well as the results of parametric testing. The

process of developing the program that will easily predict the ultimate axial capacity of

precast concrete piles is also discussed in this chapter.

4.1 Summary of Data

The data used in the study were extracted from various published geotechnical

journals and researchesand were reflected in Table 6. 70% of data were used for training,

15% for validation and another 15% were for testing. Data for training, testing and

validation will be randomly picked by MATLab upon training a network. Training and

Validation data have the big effect in training a network while testing data will just attest

the network whether it give same results as training and validation data.

58
Table 6 Training, Test and Validation Data
Length Ultimate Capacity
Pile No. Shape Area (mm2) SPT Cohesion
(mm) (kN)
1 Circle 24600 196349.54 18 17.31 6000

2 Circle 18000 656692.9 30 24.74 5000

3 Circle 56400 159043.13 24 36.58 4150

4 Square 48500 160000 22 51.6 4100

5 Circle 39600 132025.4313 14 18.84 3959

6 Square 46600 168100 10 20.39 3781

7 Square 43300 168100 7 19.11 3612

8 Square 41100 168100 14 20.07 3558

9 Square 37200 168100 7 18.64 3558

10 Square 16800 168100 39 19.05 3447

11 Square 32900 168100 7 25.19 3447

12 Circle 47000 159043.13 16 50.44 3250

13 Circle 38100 132025.4313 11 18.72 3167

14 Square 47500 168100 11 20.25 3122

15 Square 45400 168100 16 20.4 3105

16 Square 46800 160000 12.5 46.62 3100

17 Square 25000 202500 5.74 62.96 2860

18 Circle 22600 80424.77193 5 20.67 2669

19 Square 12000 90000 25 42.86 2600

20 Square 17400 122500 15 65 2600

21 Square 18600 122500 15 65 2550

22 Square 13400 168100 16 20.51 2535

23 Circle 19500 132025.4313 5 20.41 2384

24 Circle 14200 96211.28 14.8 65.08 2200

25 Circle 34800 73061.7 12 67 2150

59
26 Circle 28500 125663.71 22.6 57.11 2100

27 Circle 16500 73061.7 6 67 2090

28 Square 18200 90000 15 65 1850

29 Square 14700 90000 15 65 1800

30 Circle 12900 70685.83 16.5 20 1800

31 Square 10000 90000 25 42.86 1760

32 Circle 15100 73061.7 7 67 1710

33 Square 16800 62500 15 65 1690

34 Circle 12500 73061.7 48 67 1690

35 Square 10000 126736 23 7.2 1637.6

36 Circle 14500 70685.83 20.5 45.97 1600

37 Square 41800 96100 4 19.01 1512

38 Circle 14700 73061.7 6 67 1510

39 Square 11250 202500 3 20 1500

40 Circle 11400 73061.7 6 67 1450

41 Circle 8400 73061.7 4 67 1340

42 Circle 11200 73061.7 50 67 1300

43 Circle 18700 49087.39 16.4 72.14 1200

44 Square 11250 202500 3 20 1100

45 Square 38700 96100 4 18.77 1014

46 Square 11250 202500 5 20 1000

47 Circle 8500 73061.7 12 67 920

48 Square 9250 108900 15 20 700

49 Circle 19820 73061.7 36 14 600.8

50 Square 9250 108900 11 20 600

60
4.2Modeling Process

On operating the ANN on MATLab, the researchers must set substances that

affect the validation of the gathered data to come up with the least average error value or

the mean squared errors. These substances are the network itself, hidden layer size,

number of nodes and the learning rate. The parameters give the network a different time

to produce an efficient model especially if the parametric data gathered are not

comprehensible that is why it is very important to carefully extract the data to be used.

Eleven networks made of different hidden nodes ranging from 5-15 were trained and

tested using TANSIG as a function. In order to determine the best created model, the

performance of the network was assessed based on the obtained correlation coefficient,

R, as well as themean squared error, MSE.

A lot of time is taken in training a network which also depends on the number of

node that will be used. In order to secure the ability of the network to generalize the

number of nodes has to be kept as low as possible. In this study, the goal of the modelling

process is the value of mean squared error and coefficient of correlation. In which, the

lowest MSE value and highest R value gives the most precise predicted outputs.

4.3 Architectural Network

Artificial Neural Network model for predicting the ultimate capacity of precast

concrete piles thru the use of software MATLab is the concentration of the study. The

input parameters used in the study were shape, length, area, SPT and cohesion, while the

axial capacity of pile will be the target output.

61
Figure 31 Overview of the Architecture of a Two-Layer Neural Network Model

Figure 32 Architectural Network – 13 Nodes

Figure 31 and 32 shows the architecture of the created network. It has a five node

for the input layer, 1 hidden layer with thirteen nodes, a one node for the output layer

which is the Ultimate Axial Capacity of the piles.

62
4.4 Training Output

The performance of the produced models was tabulated as shown in Table 7. The

mean squared error values range from 902.12 to 20990.27 while the values of correlation

coefficientwere between 0.93451 and 0.99272. The final weights and biases were

determined based on the model that produced the highest overall regression value and

lowest MSE value.As noticed,the model with 13 hidden nodes has the best performance.

The final results for each node are shown in the appendices for the study of the overall

neural network training as well as the average percentage of the nodes.

Table 7 Coefficient of Correlation and Mean Squared Error Values

NODE R VALUES
MSE
S Training Validation Test All

5 0.95463 0.99349 0.97888 0.96432 2454.09

6 0.97508 0.99555 0.95304 0.97373 9157.07

7 0.96808 0.99258 0.88354 0.96737 19054.45

8 0.95803 0.99952 0.96224 0.95883 1323.88

9 0.95737 0.99475 0.96839 0.95605 20990.27

10 0.98350 0.99739 0.99091 0.98565 4750.24

11 0.95219 0.99901 0.96701 0.96471 7529.56

12 0.97731 0.99472 0.96331 0.97915 13335.60

13 0.99306 0.99932 0.98398 0.99272 902.12

14 0.92540 0.99938 0.95788 0.93451 3388.87

15 0.96297 0.99964 0.92454 0.95509 1088.02

63
Figure 33 and Figure 34shows the reliability of the ANN model for predicting the

ultimate capacity of the piles. In Figure 4.3, the regression of the best network model

were the obtained R values are 0.99306, 0.98398 and 0.99932 for training, test and

validation, respectively with an overall R value of 0.99272 while Figure 4.4 shows the

performance graph of the model.These values satisfy the minimum R value that was set

for this study. While the other regression and performance got from other nodes are seen

in Appendix B.

Figure 33 Neural Network Training Regression of Network – 13 Nodes


64
The neural network training regression at 13 nodes shows that at Training R =

0.99306, Validation R = 0.99932, Test R = 0.98398, and Overall R = 0.99272. The graph

produces accurate results because the trained input data are closer to its output target line.

Figure 34 Neural Network Training Performance of Network – 13 Nodes

The best validation performance is 902.1223 which is the square of the average

error of the network. This means that the average error is also the square root of the MSE

value which is ±30 kilonewtons. With a target value ranging from 600-6000

kilonewtons, a ±30 average error is relatively minimal, which determines that the

network is a precise model.

65
4.5 Simulated Model

After choosing the best network, the weights and biases of the five input values are

extracted from the network and used to come up with a simulated output, the axial

capacity of the pile. These weights and biases are shown in Table 8. While the other

weights and biases acquired from other nodes are seen in Appendix C.

Table 8Weights and Biases of Best Developed Model – 13 Neurons


Weight to
Bias to Bias to
Weight to Layer1 from Output
Layer1: Output:
Input iw{1,1} from Layer
B{1} B{2}
1 lw{2,1}
-0.95028 -0.99293 0.30391
2.8857 1.2543
0.25328 -2.305

-0.73156 -0.97404 1.3998


2.1712 -0.23792
0.67582 1.1325

1.5465 1.2912 1.7774 -


-1.2343 -1.0627
0.90518 -0.29098

-0.71646 1.7799 -1.5789 -


0.9768 -1.7282
0.31851 -1.8105

1.9329 1.308 -0.81196 -


Ultimate -0.44947 2.243
1.5992 0.84504
Capacity 0.43351
Model 1.2386 -0.31258 4.4366
0.32101 1.6832
0.69673 -1.2118

-1.665 1.5552 1.33 -1.3307


0.73004 1.7552
1.5501

1.2899 0.2464 1.8084 -


0.19448 1.6087
1.7047 -0.48867

-0.28305 -2.4842 1.7832 -


-1.6976 -1.6643
1.6063 1.0461

66
2.9983 1.2611 -2.0443 -
0.95501 -2.952
0.9834 -0.068417

0.25072 -1.4037 0.84381 1.761


0.18865
1.5946 -0.15037

-0.26333 1.1005 1.385


1.8352 0.72965
3.0435 -1.9131
-0.0039411 1.926 1.3403
0.19435 0.0027454 -2.2179 -0.57503

The first step in producing a simulated output is to normalize the input that ranges

from -1 to +1, -1 being lowest accepted input and +1 as the highest accepted input. In the

case of the pile shape denoted as (V1 ) , -1 is set for square piles and +1 is for circular piles.

Equation 4.1 shows the normalization formula equation that is used.

( ymax  ymin )( x  xmin )


y  ymin
( xmax  xmin ) (Eqn 1)

Where:

y  normalized value

ymin  minimum value for normalized value, -1

ymax  maximum value for normalized value, 1

x  value to be normalized

67
xmin  minimum value for the input parameter

xmax  maximum value for the input parameter

Using Equation 1, the normalized value for the pile length denoted as (V2 ) , having

the minimum accepted input of 8400mm and highest accepted input of 56400mm, would

be

[1  (1)]( x  8400)
  (1)
Normalized Pile Length (V2 ) (56400  8400)

Simplifying,

x  32400

Normalized Pile Length (V2 ) 24000 (Eqn 2)

Using Equation 1, the normalized value for the pile area denoted as (V3 ) , having the

minimum accepted input of 49087.39mm² and highest accepted input of 656692.9mm²,

would be

68
[1  (1)]( x  49087.39)
  (1)
Normalized Pile Area (V3 ) (656692.9  49087.39)

Simplifying,

2 x  705780.29

Normalized Pile Area (V3 ) 607605.51 (Eqn 3)

Using Equation 1, the normalized value for the SPT denoted as (V4 ) , having the

minimum accepted input of 3 and highest accepted input of 50, would be

[1  (1)]( x  3)
  (1)
Normalized SPT-Value (V4 ) (50  3)

Simplifying,

2 x  53

Normalized SPT-Value (V4 ) 47 (Eqn 4)

69
Using Equation 1, the normalized value for the cohesion denoted as (V5 ) , having the

minimum accepted input of 7.2 and highest accepted input of 72.14, would be

[1  (1)]( x  7.2)
  (1)
Normalized Cohesion (V5 ) (72.14  7.2)

Simplifying,

2 x  79.34

Normalized Cohesion (V5 ) 64.94 (Eqn 5)

Using the normalized data for the input variables V1, V2, V3, V4 as elements of a

1x5 Matrix α

α = [𝑉1 𝑉2 𝑉3 𝑉4 ]

The weights of layer1 from input1 and the biases to layer1 can be represented as

Matrix β with dimension 13x5 and Matrix γ with dimensions 1x13 respectively:

70
β =[−0.95028 − 0.99293 0.30391 0.25328 − 2.305 − 0.73156 −

0.97404 1.3998 0.67582 1.1325 1.5465 1.2912 1.7774 − 0.90518 − 0.29098 −

0.71646 1.7799 − 1.5789 − 0.31851 − 1.8105 1.9329 1.308 − 0.81196 −

1.5992 0.84504 1.2386 − 0.31258 4.4366 0.69673 − 1.2118 −

1.665 1.5552 1.33 − 1.3307 1.5501 1.2899 0.2464 1.8084 − 1.7047 −

0.48867 − 0.28305 − 2.4842 1.7832 − 1.6063 1.0461 2.9983 1.2611 −

2.0443 − 0.9834 − 0.068417 0.25072 − 1.4037 0.84381 1.5946 − 0.15037 −

0.26333 1.1005 1.385 3.0435 − 1.9131 −

0.0039411 1.926 1.3403 0.19435 0.0027454 ]

γ= [2.8857 2.1712 − 1.2343 0.9768 − 0.44947 0.32101 0.73004 0.19448 − 1.6976 0.95501 1.761 1.8352 − 2.2179 ]

In order to calculate the values for the nodes in the hidden layer using TANSIG

function, Equation 6 can be used:

𝐻𝑛 = 𝑡𝑎𝑛ℎ(𝛼1,1 ∗ 𝛽𝑛,1 + 𝛼1,2 ∗ 𝛽𝑛,2 + 𝛼1,3 ∗ 𝛽𝑛,3 + 𝛼1,4 ∗ 𝛽𝑛,4 + 𝛼1,5 ∗ 𝛽𝑛,5 + 𝛾1,𝑛 )

(Eqn 6)

Substituting the values of the weights and biases from n=1 up to n=13 into

Equation 6. Series of equations can be derived:

𝐻1 = 𝑡𝑎𝑛ℎ(−0.95028 𝑉1 − 0.99293 𝑉2 + 0.30391 𝑉3 + 0.25328 𝑉4 − 2.305𝑉5


+ 2.8857)

71
𝐻2 = 𝑡𝑎𝑛ℎ(−0.73156 𝑉1 − 0.97404 𝑉2 + 1.3998 𝑉3 + 0.67582 𝑉4 + 1.1325𝑉5 +

2.1712)

𝐻3 = 𝑡𝑎𝑛ℎ(1.5465 𝑉1 + 1.2912 𝑉2 + 1.7774 𝑉3 − 0.90518 𝑉4 − 0.29098𝑉5 −

1.2343)

𝐻4 = 𝑡𝑎𝑛ℎ(−0.71646 𝑉1 + 1.7799 𝑉2 − 1.5789 𝑉3 − 0.31851 𝑉4 − 1.8105𝑉5 +

0.9768)

𝐻5 = 𝑡𝑎𝑛ℎ(1.9329 𝑉1 + 1.308 𝑉2 − 0.81196 𝑉3 − 1.5992 𝑉4 + 0.84504𝑉5 +

−0.44947)

𝐻6 = 𝑡𝑎𝑛ℎ(1.2386 𝑉1 − 0.31258 𝑉2 + 4.4366 𝑉3 + 0.69673 𝑉4 + −1.2118𝑉5


+ 0.32101)
𝐻7 = 𝑡𝑎𝑛ℎ(−1.665 𝑉1 + 1.5552 𝑉2 + 1.33 𝑉3 − 1.3307 𝑉4 + 1.5501𝑉5 +

0.73004)

𝐻8 = 𝑡𝑎𝑛ℎ(1.2899 𝑉1 + 0.2464 𝑉2 + 1.8084 𝑉3 − 1.7047 𝑉4 − 0.48867𝑉5 +

0.19448)

𝐻9 = 𝑡𝑎𝑛ℎ(−0.28305 𝑉1 − 2.4842 𝑉2 + 1.7832 𝑉3 + −1.6063 𝑉4 + 1.0461𝑉5 −

1.6976)

𝐻10 = 𝑡𝑎𝑛ℎ( 2.9983 𝑉1 + 1.2611 𝑉2 − 2.0443 𝑉3 − 0.9834 𝑉4 − 0.068417𝑉5 +

0.95501)

𝐻11 = 𝑡𝑎𝑛ℎ(0.25072 𝑉1 − 1.4037 𝑉2 + 0.84381 𝑉3 + 1.5946 𝑉4 + −0.15037𝑉5 +

1.761)

72
𝐻12 = 𝑡𝑎𝑛ℎ(−0.26333 𝑉1 + 1.1005 𝑉2 + 1.385 𝑉3 + 3.0435 𝑉4 − 1.9131𝑉5 +

1.8352)

𝐻13 = 𝑡𝑎𝑛ℎ(−0.0039411 𝑉1 + 1.926 𝑉2 + 1.3403 𝑉3 + 0.19435 𝑉4 +

0.0027454𝑉5 − 2.2179 )

The values of each node in the hidden layer can be represented as Matrix 𝛿 with

dimension 1x13

𝛿 = [𝐻1 𝐻2 𝐻3 𝐻4 𝐻5 𝐻6 𝐻7 𝐻8 𝐻9 𝐻10 𝐻11 𝐻12 𝐻13 ]


The weights of the output from layer1 can be represented as Matrix ε with

dimension 1x13

𝜀 = [1.2543 − 0.23792 − 1.0627 − 1.7282 2.243 1.6832 1.7552 1.6087 − 1.6643 − 2.952 0.18865 0.72965 − 0.57503] ]

After calculating the values of each nodes of the hidden layer, Equation 7 can be

used to compute the normalized value of the simulated ultimate capacity which will range

from -1 to +1. The ultimate capacity is denoted by τ while Bias2 which is equal to

0.43351 is denoted by F.

𝜏 = 𝑡𝑎𝑛ℎ( 𝛿1,1 ∗ 𝜀1,1 + 𝛿1,2 ∗ 𝜀1,2 … . . + 𝛿1,13 ∗ 𝜀1,13 + 𝐹)

By substituting the values of Matrix ε, the equation will be simplified to

𝜏 = 𝑡𝑎𝑛ℎ( 1.2543𝐻1 − 0.23792 𝐻2 − 1.0627𝐻3 − 1.7282𝐻4 + 2.243𝐻5 +

1.6832 𝐻6 + 1.7552 𝐻7 + 1.6087 𝐻8 − 1.6643 𝐻9 − 2.952 𝐻10 + 0.18865 𝐻11 +

0.72965 𝐻12 + −0.57503𝐻13 + 0.43351 ) (Eqn 7)

73
The actual value of the ultimate capacity denoted by U can be calculated by

denormalizing the value τ using Equation 8

(𝜏+1)(𝑥𝑚𝑎𝑥 −𝑥𝑚𝑖𝑛 )
𝑈= + 𝑥𝑚𝑖𝑛
2

(Eqn 8)

By substituting the values for xmaxandxmin as 6000 and 600 respectively, the final

equation for predicting the ultimate capacity in terms of kN is:

𝑈 = 2700𝜏 + 3300 (Eqn 9)

To have a better understanding of the prediction performance of ANN, the

simulated ultimate capacity values obtained from the derived equation (Eqn 9) is

compared with the actual target values (see Table 9). The percentage error per output was

calculatedusing (Eqn. 10). The model was able to predict the ultimate capacity of the

piles with an average error of 4.81%. Average error for others nodes were shown in

Appendix D.

∣𝐴𝑐𝑡𝑢𝑎𝑙 𝑐𝑎𝑝𝑎𝑐𝑖𝑡𝑦−𝑆𝑖𝑚𝑢𝑙𝑎𝑡𝑒𝑑 𝑐𝑎𝑝𝑎𝑐𝑖𝑡𝑦∣


𝑃𝑒𝑟𝑐𝑒𝑛𝑡 𝐸𝑟𝑟𝑜𝑟 = ∗ 100 (Eqn 10)
𝐴𝑐𝑡𝑢𝑎𝑙 𝑐𝑎𝑝𝑎𝑐𝑖𝑡𝑦

Table 9 13 Neurons Network Model Average Error

74
13 Neurons

Input Pile
Target Output Error %Error
Parameters No.
1 6000 5909.54 90.46 1.51%

Shape 2 5000 5005.02 5.02 0.10%


3 4150 4360.68 210.68 5.08%
Length
4 4100 4092.18 7.82 0.19%
Area
5 3959 3999.44 40.44 1.02%
SPT 6 3781 3300.78 480.22 12.70%

Cohesion 7 3612 3611.74 0.26 0.01%


8 3558 3534.19 23.81 0.67%
9 3558 3523.26 34.74 0.98%
10 3447 3030.18 416.82 12.09%
11 3447 3320.40 126.60 3.67%
12 3250 3246.28 3.72 0.11%
13 3167 3184.57 17.57 0.55%
14 3122 3126.09 4.09 0.13%
15 3105 2780.48 324.52 10.45%
16 3100 3116.63 16.63 0.54%
17 2860 2909.47 49.47 1.73%
18 2669 2637.44 31.56 1.18%
19 2600 2541.32 58.68 2.26%
20 2600 2548.60 51.40 1.98%
21 2550 2601.04 51.04 2.00%
22 2535 2448.05 86.95 3.43%
23 2384 2404.47 20.47 0.86%
24 2200 2179.43 20.57 0.93%
25 2150 2154.55 4.55 0.21%
26 2100 2101.32 1.32 0.06%

75
27 2090 1949.48 140.52 6.72%
28 1850 1978.18 128.18 6.93%
29 1800 1900.96 100.96 5.61%
30 1800 2041.49 241.49 13.42%
31 1760 1677.74 82.26 4.67%
32 1710 1717.98 7.98 0.47%
33 1690 1461.84 228.16 13.50%
34 1690 1678.93 11.07 0.65%
35 1637.6 1651.74 14.14 0.86%
36 1600 1896.61 296.61 18.54%
37 1512 1485.06 26.94 1.78%
38 1510 1655.09 145.09 9.61%
39 1500 1308.97 191.03 12.74%
40 1450 1341.51 108.49 7.48%
41 1340 1329.28 10.72 0.80%
42 1300 1308.33 8.33 0.64%
43 1200 1217.21 17.21 1.43%
44 1100 1308.97 208.97 19.00%
45 1014 1088.42 74.42 7.34%
46 1000 1207.34 207.34 20.73%
47 920 976.57 56.57 6.15%
48 700 691.66 8.34 1.19%
49 600.8 658.42 57.62 9.59%
50 600 637.15 37.15 6.19%
Average Error 4.810%

76
4.6 Sensitivity Analysis

The researchers conducted a sensitivity analysis in which relationship between the

input and output parameter was evaluated using parametric testing based on the simulated

model. In parametric testing, the input parameters will be set constant except for one in

77
order to determine how the varying parameter affects the target output. Three (3) sets of

data were used in the testing namely, the pile with highest and least ultimate capacity,

and the average of all the parameters. The results of parametric testing of each parameter

i.e. pile length, cross-sectional area, SPT and cohesion, are reflected in the following

figures.

4.6.1 Parametric Testing for Varying Length

78
Figure 35 Parametric Testing for Varying Length (Maximum)

Figure 36 Parametric Testing for Varying Length (Minimum)

79
Figure 37 Parametric Testing for Varying Length (Mean)

As observed in the figures 35 to 37, the graph shows that the ultimate capacity

from the 3 sets of data increased as the length of the pile increases. With the values of the

shape, area, SPT, and cohesion set as constants, the increase in length resulted in a

directly proportional relationship with the ultimate capacity as shown by the sloping

upward trend line of the graph.

80
4.6.2 Parametric Testing for Varying Area

Figure 38 Parametric Testing for Varying Area (Maximum)

Figure 39 Parametric Testing for Varying Area (Minimum)

81
Figure 40 Parametric Testing for Varying Area (Mean)

As observed in the figures 38 to 40, the graph shows that the ultimate capacity

from the 3 sets of data increased as the cross-sectional area of the pile increases. With the

values of the shape, length, SPT, and cohesion set as constants, the increase in area

resulted in a directly proportional relationship with the ultimate capacity as shown by the

sloping upward trend line of the graph.

82
4.6.3 Parametric Testing for Varying SPT

Figure 41 Parametric Testing for Varying SPT (Maximum)

Figure 42 Parametric Testing for Varying SPT (Minimum)

83
Figure 43 Parametric Testing for Varying SPT (Mean)

As observed in the figures 41 to 43, the graph shows that the ultimate capacity

from the 3 sets of data increased as the SPT increases. With the values of the shape, area,

length, and cohesion set as constants, the increase in SPT resulted in a directly

proportional relationship with the ultimate capacity as shown by the sloping upward trend

line of the graph.

84
4.6.4 Parametric Testing for Varying Cohesion

Figure 44 Parametric Testing for Varying Cohesion (Maximum)

Figure 45 Parametric Testing for Varying Cohesion (Minimum)

85
Figure 46 Parametric Testing for Varying Cohesion (Mean)

As observed in the figures 44 to 46, the graph shows that the ultimate capacity

from the 3 sets of data increased as the cohesion of the soil increases. With the values of

the shape, area, length, and SPT set as constants, the increase in cohesion resulted in a

directly proportional relationship with the ultimate capacity as shown by the sloping

upward trend line of the graph.

86
4.6.5 Parametric Testing for Varying Shape

Figure 47 Parametric Testing for Varying Shape (Maximum)

Figure 48 Parametric Testing for Varying Shape (Minimum)

87
Figure 49 Parametric Testing for Varying Shape (Mean)

As observed in the figures 47 to 49, the graph shows that the ultimate capacity

from the 3 sets of data is higher with the circular pile than with a square pile. With the

values of the area, length, cohesion and SPT set as constants, the graph showed that

circular piles result to higher capacity compared to a square.

88
4.6.6 R-Squared Values of each Parameters

Figure 50Strength of Relation between Ultimate Capacity and Input Parameters

The R-squared values are graphically presented in Figure 50 for better

comparison. The R-squared values pertain to the influence of different input parameters

on the output parameter. An R-squared value closer to 1 indicates a higher effect on the

ultimate capacity of pile. Based from Figure 4.20, the cohesion is the most influential

parameter in the set of data used in the study followed by the length of the pile, cross-

sectional area and SPT respectively.

89
4.7 Model for Predicting Ultimate Capacity of Driven Precast Concrete Piles

After the derivation of the equation has been done, it was then transported to

Microsoft Excel to easily predict the ultimate capacity of precast concrete piles which

requires four input values i.e. pile length in millimeter, shape, cross-sectional area in

square millimeter, SPT value and cohesion in kPa. Figure 51 shows a sample predicted

value of ultimate capacity of precast concrete piles.

Figure 51Sample Model in Predicting Ultimate Capacity of Precast Concrete Piles

90
Thismodel developed from Microsoft Excel can predict the ultimate capacity of precast

concrete piles. Since the model was trained from specific set of data, the input values that

can be used have specific limitations. Table 10 shows the range of the input values that

can be used in order to reduce errors in simulating the output. The predicted output is also

limited to minimum and maximum values of 600 kN and 6000 kN respectively.

Table 10 Range of Parameters

Parameters Min. Value Max. Value


Length (mm) 8400 56400
Area (mm2) 49087.39 656692.9
SPT 3 50
Cohesion 7.2 72.14

91
Chapter 5

CONCLUSION

This chapter presents the conclusion drawn from the results of extensive

experimentation conducted during the whole study. Each conclusion answers the

corresponding objectives mentioned.

An artificial neural network (ANN) model is developed in this study to predict the

ultimate capacity of driven precast concrete piles. Different types of network were

created and trained to determine which model will give the closest simulated output

compared to actual value. It was found that the network model of 13 hidden nodes in one

hidden layer yields the best performance with coefficients of correlation (R) closest to 1

and the lowest mean squared error (MSE)with 4% average error. The results suggest the

reliability of the neural network model in predicting the ultimate axial capacities of piles.

Additionally, sensitivity analysis is conducted using parametric testing to

determine the individual effect of the input parameters with respect to the behavior of the

output. Results of the sensitivity analysis show that the relationship of each input

parameter is directly proportional to the output.

Furthermore, the regression line of each parameter produces an upward slope

which gives the input and output a definite connection. It shows in the graph of

coefficient of determination (R-squared) that the cohesion and pile length are the two

92
parameters that have the highest value. This represents that cohesion and pile length are

the most influential parameters in predicting the ultimate capacity of driven precast

concrete piles.

A Microsoft Excel program was also developed by forming an equation for

predicting ultimate capacity of piles based on the weights and biases obtained from the

network model. The range of parameters were also set in the program in which will be

the basis for future application of the study. Thus, the program can be used as an accurate

and quick tool for estimating the ultimate axial capacity of driven precast concrete piles.

93
Chapter 6

RECOMMENDATION

The following are the recommendation that can contribute to future researchers

based on the test conducted and data gathered.

The group suggests performing the neural network process in different software

application such as the MemBrain andNeuroSolutions Neural Network.

Artificial Neural Network has a wide range of capability that can help predict

solutions to a variety of topics, trying other method of software application can help other

researchers assure that the results given to them by one software application is true and

correct. In this case, using other method or software application will only be used for

checking as an external validation.

The researchers also suggest that if possible, future users of this paper are

recommended to include additional parameters that can affect the ultimate capacity of the

pile such as the perimeter, and compressive strength of concrete in order to increase the

accuracy of the model.Data gathering is a vital part of this research, we suggest to gather

more data to increase the range of input and output values of the database.

Lastly, to the future users of this paper, the researchers suggest to try to make

other sensitivity analysis such as Cosine-altitude method, P-value method or other

statistical tools for determining the behavior of the parameters to see the accuracy of the

simulated model.

94
BIBLIOGRAPHY

[1] E. H. Mijena, "A comparison of friction piles bearing capacity based on theoretical
and empirical mathematical methods," 2012.

[2] A. Hasnat, A. R. M. F. Uddin, E. Haque, P. Saha and W. Rahman, "Ultimate load


capacity of axially loaded vertical piles from full scale load test results interpertations-
applied to 20 case histories," 2012.

[3] F. B. Eslami A, "Pile capacity by direct CPT and CPTu methods applied to 102 case
histories," Canadian Geotechnical Journal vol 34, pp. 886-904, 1997.

[4] H. Maizir, N. Gofar and K. A. Kassim , "Artificial Neural Network Model for
Prediction of Bearing Capacity of," Jurnal Teknik Sipil, 2015.

[5] F. Baligh and G. Abdelrahman, "Modification of Davisson’s method".

[6] M. A. Shahin, "Load-settlement moddeling of axially loaded steel," Soils and


Foundations, 2014.

[7] W. N. Abd Elsamee , "Evaluation of the ultimate capacity of friction Piles," Scientific
Research, 2012.

[8] B. Tarawneh, "Pipe pile setup: Database and prediction model using artificial neural
network," Soils and Foundations, 2013.

[9] A. E. Osman M, "Comparison Between Dynamic and Static Pile Load Testing,"
Electronic Journal of Geotechnical Engineering, 2013.

95

You might also like