•**4»>fr' ''t. i''" 'JSro.-' '-•« .

i^ifii^' <

PIPE ROUTING FRAMEWORK

FOR DETAILED SHIP DESIGN

ANDI ASMARA
X43^ö/5-|c>é \ii^

Pipe Routing Framework

for Detailed Ship Design

PROEFSCHRIFT

ter verkrijging van de graad van doctor

aan de Technische Universiteit Delft,
op gezag van de Rector Magnificus prof. ir. K.C.A.M. Luyben,
voorzitter van het College voor Promoties,
in het openbaar te verdedigen op woensdag 10 juli 2013 om 15.00 uur

door

Andi A S M A R A

master of science in automation and robotics, Universitat

Dortmund, Duitsland

geboren te Jakarta, Indonesië

Dit proefschriR is goedgelceurd door de promotor:
Prof.dr.ir. U. Nieniiiüs MBA

Samenstelling promotiecommissie:
Rector Magnificus, voorzitter
Prof.dr.ir. U. Nienhuis MBA, Technische Universiteit Delft, promotor
Prof.Dr.-Ing. S. Krueger, TU Hamburg-Harburg
Prof.ir. J.J. Hopman, Technische Universiteit Delft
Prof.ir. D. Stapersma, Technische Universiteit Delft
Prof.dr. I. Horvath, T'echnische Universiteit Delft
Dr.ir. J.M.G. Coenen, Technische Universiteit Delft

Published by

VSSD, Delft, The Netherlands

internet: http;//www.vssd.nl/hlf
e-mail: hlf@vssd.nl

ISBN 97890-6562-326-3

Copyright © 2013 by Audi Asmara

All rights reserved. No part of the material protected by this copyright notice
may be reproduced or utilized in any form or by any means, electronic or
mechanical, including photocopying, recording or by any information storage
and retrieval system, without the prior permission of the author.
Contents

1 Introduction l
1.1 Process Innovations in Shipbuilding 1
1.2 Innovative Ways to Route Pipes in Ships 2
1.3 Pipe Routing Process in Ships 4
1.3.1 Piping Design Phases 4
1.3.2 Pipe Routing Knowhow 6
1.3.3 Quality of the Pipe Routed 7
1.3.4 Design Requirements 8
1.4 Research Approach and Laboratory 9
1.5 Organization of the Work 11

2 Pipe Routing in Practice 13

2.1 Introduction 13
2.2 Stakeholders 14
2.3 Tools and Information 17
2.3.1 Tools 18
2.3.2 Information 18
2.4 Pipe Routing Process 19
2.4.1 Organizational Process 19
2.4.2 Routing Process 20
2.4.3 Working Area 21
2.4.4 Object Constraints 22
2.4.5 Collaboration between Pipe Engineers 23
2.4.6 Modification Possibility 23
2.5 Criteria for Pipe Routing in Ships 24
2.5.1 Rules of Pijjc l^outing 24

i
ii CONTENTS

2.5.2 Minimization of Pipe Cost 26

2.6 Pipe Routing Common Knowledge 28
2.6.1 General Guidance of Pipe Routing 28
2.6.2 Dividing a Pipe to Pipe Spools 28
2.6.3 Bended Pipes 29
2.6.4 Welding Elbow 29
2.7 Some Mistakes in Practice 29
2.8 Summary 32

3 Pipe Routing Frameworl( 35

3.1 Introduction 35
3.2 Required Data 36
3.2.1 Piping and Instrument Drawing 36
3.2.2 General Arrangement 38
3.2.3 Component Details 39
3.2.4 3D Steel Construction 39
3.2.5 Tank Plan Drawing 39
3.2.6 NoGoArea 40
3.3 Pipe Routing 41
3.4 Literature Review of Automatic Pipe Routing 44
3.4.1 Early Years 44
3.4.2 Zhu and Latombe 45
3.4.3 Kang's Expert System 46
3.4.4 Sandurkar and Ito 47
3.4.5 Zuurmond 47
3.4.6 Park and Storch 48
3.4.7 Commercial Automatic Pipe Router 49
3.5 Summary 50

4 Related Worl( 51
4.1 Shortest Path Problem 51
4.2 Deterministic Approaches 53
4.2.1 Graph 'iVaversal Algorithms 53
4.2.2 Maze Algorithm 63
4.2.3 Line search algorithm 67
4.3 Heuristic Algorithm 71
4.3.1 Genetic Algorithm 71
4.3.2 Particle Swarm Optimization 74
4.3.3 Ant Colony Optimization 77
4.4 Comparison of Algorithms 79
4.5 Beyond the Shortest Path Problem 80
4.5.1 Multiple Nodes 80
 CONTENTS iii

4.5.2 Mutually Intervening Case 81

4.6 Combinatorial Optimization 82
4.7 Model Simplification 83
4.8 Summary 84

5 The Methodology Architecture and its Implementation 85

5.1 Intro(l\ictiou 85
5.2 Interlace Module 86
5.2.1 Interface to CAD Packages 86
5.2.2 Smart P&I Diagram Tool 87
5.2.3 Simplification of the 3D Model 89
5.2.4 Summary of the Interface Module 97
5.3 Pipe Router Module 98
5.3.1 Routing Criteria 99
5.3.2 Pathfinder Module 102
5.3.3 Quality Measurement 110
5.3.4 Optimizing the Solution Ill
5.3.5 Routing Parameters Behavior Ill
5.4 Implementation Ill
5.5 Summary 114

6 Pipe Routing Methodology Validation 117

6.1 Introduction 117
6.2 Structural Validation 117
6.2.1 Internal Consistency of each Parent Construct 118
6.2.2 Internal Consistency of The Method 118
6.2.3 Appropriateness of The Example Problem 118
6.3 Performance Validation 118
6.3.1 The Usefulness of The Method for The Chosen Example
Problem 119
6.3.2 The Demonstrated Usefulness is Linked to Applying The
Methodology 128
6.3.3 The Methodology is Useful for Domains that are Broader . 128
6.4 Sensitivity Analysis 130
6.5 Sensitivity Analysis 131
6.6 Summary 135

7 Conclusions and Recommendations 137

7.1 Research Questions 137
7.2 Conclusion 138
7.3 Recommendations 140
 CONTENTS

Bibliography
149
Summary
151
Samenvattmg
1 ^*\

Acitnowiedgments

Curriculum Vitae
Chapter 1
Introduction

1.1 Process innovations in Siiipbuilding

The shipbuilding industry is a mature market. lYue to its nature, the dominance
within the shipbuilding industry has shifted from the European markets, penal-
ized by their high cost of production, to the (at least sometime) low-cost Asian
destinations such as South Korea, Japan and China. This happened mostly for re-
latively non-complex ships like oil tankers, dry bulk carriers, and container vessels.
To maintain their market competitiveness, the European shipyards and co-makers
need to dig into niche markets by concentrating on building specialized vessels,
because the cost of production for this kind of ships is relatively less important.
To be more competitive, European shipyards must also compensate their disad-
vantage of higher production cost with their skill to design and build complex
ships. However, the "low-cost" Asian shipyards are continuously expanding their
knowledge to build complex ships. Therefore the "more-expensive" European
shipyards are forced to reduce cost.
This situation forces the European shipyards to be more efficient, in design
and production processes. Product and process innovation have become main
issues. Projects such as LeaderSHIP 2015 at the European Union level, Con-
current Engineering, Planning, Pricing, and Production (CE3P) and Integraal
Samenwerken in The Netherlands, are aimed at improving the current level of
design and production processes in shipbuilding.
One of the main subjects in those projects is to conduct research and develop-
ment to utilize information and communication technology to improve the design
and engineering processes, and to implement that in the real process. Improving
process control; simulating the shop-floor and on-board production processes; de-
veloping a multi-party communication framework; and innovative ways to route
pipes and other conduits are some of the research subjects.

1
2 INTRODUCTION 1.2

1.2 Innovative Ways to Route Pipes in Ships

A ship has many systems and subsystems that consist of a large number o\ pipes.
Even a small vessel might have more than one thousand pipe and in a larger
vessel this number can reach three to four thousands pipes and more, and all
pipes must be routed. Pipe routing is one of the most important activities during
detailed design because many other detailed design activities depend on it. Also
it is important because this activity consumes a significant part of the detailed
design man-hours as the pipe systems in a ship typically consist of thousands
of pipe elements. Park and Storch [2002] mention that the time that is needed
can reach 50% of the total design process time. That number is significant,
because according to American Bureau of Shipping, the labor cost is around 60%
of total cost of ship. Based on our interview with the engineering department of a
shipyard, the process of pipe routing for a middle size complex ship can consume
30-40 thousand man hours.
There are basically four main characteristics of ship building that encourage
people to utilize computers to route pipes in the design process. The first char-
acteristic is that many of the middle to large size special-purpose ships normally
have a unique specification which leads to the necessity to redo the design process
for each vessel. The high sensitivity of pipe design to changes in the specification is
the second one. The third one is the time needed to route the pipes is substantial
and sometimes can be the bottle neck in the design and production process. Last
but not least is the fact that the pipe routing is largely done manually by pipe
engineers who need many years of experience to do this properly and efficiently.
The design of the piping systems consumes a large part of the engineering
effort for a modern ship. The fact that makes it so significant is that the nature of
ship production is different from other vehicles e.g. cars. Since the 1920&, nearly
all cars have been mass-produced; they were designed once and produced in a
large quantity. In ship production, especiaUy for the building of specialized ships
like dredgers, offshore and naval vessels, every ship that is built has a different
specification that requires the design process to be done for each ship. Even ships
with the same functionality often have a different specification.
Pipes are needed for distribution of fluids and gases between pieces of equip-
ment in the ship. Due to the complexity of the systems that are needed to support
the operability of the ship, the total amount of pipes needed is large. Also the
variability is large; only few pipe pieces are the same in different ships. The com-
plexity of pipe systems is compounded by the fact that the space available for
pipe systems in a ship is very limited. Normally, in the end, most of the avail-
able space will be occupied by the pipes. Tf a small modification is subsequently
needed, e.g. if there is a specification change for the size of a component and/or
the component needs to be moved to a different place, all the pipes that arc con-
nected to that component need to be rerouted or at least need to be modified.
Moreover and more awkwardly, when the compartment is already crowded with
pipes, the changes might affect other pipes that are not immediately connected
1.2 INNOVATIVE WAYS TO ROUTE PIPES IN SHIPS 3

to that component, but lie in the vicinity.

In the day to day practice oi ship design, the iterative nature of the entire
design process is such that structural changes, or changes of the equipment spe-
cifications or relocation of elements of the ship are prone to occur, requiring
time-consuming rework for any pipe affected.
In order to be more cost effective in the production stage, most of the pipes
should be installed during the pre-outfitting stage. It means that the pipes that
belong to one section of the ship must be installed immediately after the steel
construction of that section is ready, so when that section is assembled to the hull
on the slipway, all pipes are already in their place.
Nowadays in practice, the pipe routing process is largely done manually by
a pipe designer using CAD software which assists in checking for collisions and
defining the details of pipes e.g. the pipe bend radius and its specifications.
Therefore the experience of the designers is the most important ingredient for
this process. The assessment of the design results is a subjective matter due to
heterogeneous design preferences and conventions of designers, Kang et al. [1996].
Much research has been done to develop innovative ways to route pipes and
most are aiming to have a capability to route pipes automatically. Automatic
pipe-routing has been a research topic for a long time resulting in various ap-
proaches, not only in the ship production application, Wangdahl, Pollock, and
Woodward [1974J, but also in process plant design. Matsui et al. [1979] and Guir-
ardello and Swaney [2005]. The research started with 2D workspace and simple
obstacles, Wangdahl et al. [1974], and gradually extended to the stage of 3D work-
space with multiple constraints and multiple objectives. Park and Storch [2002].
In terms of the optimization techniques, either deterministic, Newell [1972], non-
deterministic. Fan et al. [2006] or a combination of both methods have been used
to improve the results, Asmara and Nienhuis [2008].
Unfortunately most of the research has focused on how to route the pipe itself,
without having due regard to the data that is needed to be prepared beforehand,
nor what to do with the result afterward, e.g. Newell [1972], Sandurkar and Chen
[1998] and Zhu and Latombe [1991]. Furthermore, most of the research on the
automatic pipe-routing problem is based on a simple environment that consists of
a small number of pipes, obstacles and conditions, such as Sandurkar and Chen
[1998] and Zuurmond [2004]. Only a few of them consider the practical aspects
of pipe routing in ships, Park and Storch [2002].
Until the date of this thesis (2013), it was questionable whether it is possible
to develop the complete methodology to perform automatic pipe-routing in the
real ship design process and satisfy all the rules and standards in ship design,
engineering and production.
In our attempt to improve the pipe routing process, we made a thorough
investigation of the current pipe routing process in practice. This has been done by
carrying out extensive interviews and discussions with experienced pipe engineers,
and also investigating their work results. Eventually, the quality of the manual
routing by experienced pipe engineers is excellent and does not require many
4 INTRODUCTION 1.3

improvements. What is more important is to shorten the time needed for pipe
routing processes while maintaining the same quality compared to the current
process. This can be achieved without fundamental changes to the commonly
used manual pipe routing method but takes advantage of computer power by
translating manual processes into computer procedures.
Those facts motivate the main research objectives of this thesis:
1. to what extent can the expertise of a pipe engineer be identified and trans-
lated into procedures that lend themselves to bo computerized?
2. how should we put together a pipe routing methodology bsised on the results
of the first question and combine it with advanced optimization techniques
in a practically applicable method?
In the next sections, we describe several issues related to the development of
the pipe routing methodology. These consist of the current situation of the pipe
routing process, pipe routing know-how and design requirements. For each issue,
we introduce a corresponding research question.

1.3 Pipe Routing Process in Ships

1.3.1 Piping Design Phases
In the design of a ship, there are two main skill areas involved, the naval architect
and the marine engineer. The naval architect is concerned with the hull, its
construction, form, habitability and ability to endure its environment. The marine
engineer is responsible for the various systems inside the ship necessary to support
the ship's functionality. More specifically, this refers to the machinery required
for propulsion, steering, anchoring and ship securing, cargo handling or other
mission-relevant functions, air conditioning, power generation and its distribution,
see Klein Woud and Stapersma [2003] ,Harrington [1992], Taylor [1996].
If we look in more detail to the various main systems that support the func-
tionality of the ship, each of these needs some separate sub-systems to support
its function. For example the propulsion system needs some other systems to
support its main components. The sub-systems that are needed by the main en-
gine normally consist of a fuel oil system, an oil lubrication system, a starting air
system, cooling water system (sea and fresh water) and an exhaust gas system.
Every system in a ship, regardless if it is a main or auxiliary system, needs
pipes or ducts to transfer liquid or gas between its component parts (includ-
ing tanks), and/or needs cables to deliver electrical power to the equipment and
receive electronic signals. Therefore the piping, IIVAC, and cabling is very im-
portant in a ship and can be said to be as important aa the blood vessel system
in a human body. Indeed, also the walkway and stairs for a human being may
be seen as a distribution channel. The complexity of arranging the distribution
channels is high.
Although our methodology is aimed to bo also suited for distribution of electric
1.3 PIPE ROUTING PROCESS IN SHIPS 5

power, air and even people, we, for now, concentrate on the piping system. The
design of the piping system is done in four different phases; conceptual design,
preliminary design, contract design, and detailed design, Harrington [1992]. Dur-
ing concept design, a tentative list of requirements is developed based on the
available ship characteristics. If sufficient detail arrangement is available, a pre-
liminary check can be periormed to make sure that the major pipe systems can
be accommodated. However, normally in this phase the available data is insuffi-
cient to develop independent cost estimates for each system. In practice, the cost
estimation is usually extrapolated from data for existing ships of similar design.
The major piping system components are selected and arranged in the ship
during the preliminary design. Preliminary estimates of system flows, pressure
and temperature are made to support component selection. Piping system com-
ponent selection needs to meet the piping system performance requirements with
due consideration of the weight, cost, and reliability. The performance require-
ments of the piping systems are determined on the basis of the ship mission, size,
operating profile, main machinery, and other factors.
In this design phase, the schematics that depict the interconnection between
the components in piping systems, called piping system diagrams, are started
to be developed with a preliminary level ol detail. The approximate locations of
major components and the largest pieces of piping arc determined and the general
arrangement is prepared and reviewed to ensure that enough space is available
for piping and other distributive systems.
In the contract design, the additional details and specifications of each system
are developed based on the outhnes that are defined in the prelimmary design
phase. Contract guidance drawings are developed to illustrate relationships and
interconnections between systems that may not be understood easily from a writ-
ten specification. These drawings together with specifications define the system
sufficiently to ensure that the owner's requirements for performance and quality
are mutually understood and agreed, and to permit the shipbuilder to prepare a
bid.
The last phase of the piping system design is the detailed design, and it pro-
duces a full definition of every pipe system element in a drawing format that is
used to manufacture all parts ol the systems, and also to install them in the ship.
While the first three design phases are primarily focused on system performance,
this phase is focused on construction.
The detailed design is begun with a completion of the piping system diagrams
as more detailed and final data become available. The piping system diagrams are
used to ensure that the systems will meet the specification requirements. They
also help to ensure compatibility of all elements in the systems with each other
and also with other elements like machinery interfaces. Normally, the information
of the system arrangement is included in the piping system diagram with varying
level of detail. The piping system diagrams contain the foundation of every piping
system e.g. the component symbols, pipe size and specification, valve description.
How direction, and other useful information. The quality and clarity of the piping
6 INTRODUCTION i.3

diagrams are important because tlicy serve as the baseline for all processes in the
detailed design phase.
After the piping system diagrams have a sufficient level of detail, the pipe
routing process can be started. The basic approach to pipe routing is develop-
ing a collision free route of a pipe between two or more connection points in a
3D, obstacle-scattered environment, according to the rules and standards. Pipe
routing is difficult because, among other reasons, the pipe is generally subject
to multiple design constraints. The most important constraint is that the pipe
routing solution must comply with the marine classification and regulations. In
addition, the space available in a ship and allowed to be used for pipes is usually
limited. This condition is especially true for certain areas of the ship, such as the
engine room of a ship. One more unavoidable aspect that makes this activity even
harder is the fact that the specifications of the ship are changing quite frequently
during the design process often requiring re-work.
The piping system diagrams are not the only information that is needed to
perform a pipe routing process. It also needs the 3D model of each system com-
ponent to know the exact location of the connection point of the pipe and to
prevent the collision between pipes and the system components. The 3D model
of the hull and superstructure is also needed to ensure that pipes can be routed
optimally without collision or unnecessary penetration of the hull construction.
Besides that, additional information of the system components might be needed,
such as the information about pipe systems.
When a pipe has been routed, it is divided into several parts, called pipe spools
and the pipe spool drawing is generated to be used in production and installation
of the pipe in the ship.
If we look at those four design phases, the pipes are actually routed during the
detail design phase and as mentioned in the previous sections, the pipe routing
process requires many working hours. Therefore, it is logical to focus on this
stage.
However, we also draw attention to the benefits to be able to route pipes auto-
matically in the pre-contractual phase. During this phase, the cost estimation of
pipes is needed. Currently, it is merely estimated using statistical methods based
on data from previous ships. The implementation of automatic pipe routing in
this phase will give more confidence on the estimated cost. Therefore we need to
investigate to answer the first research question below:

R e s e a r c h q u e s t i o n # 1 : On which design phases should we focus and what is

the reason for that?

1.3.2 Pipe Routing Knowhow

As we mentioned above, the pipe routing process is largely done manually by
a pipe designer using CAD software which assists in checking for collisions and
1.3 PIPE ROUTING PROCESS IN SHIPS 7

defining the details of pipes. Tlie assistance that can be provided by current
versions of CAD software is limited to the specification that immediately relates
to the pipe and the standard components that are attached to that pipe. It does
not take into account the type of the system to which the pipe belongs and/or
the category of area where the pipe is routed. It will not give a warning if a pipe
is routed in violation of the marine classification, for example if a pipe from a fuel
oil system was routed above the combustion engine.
As described in Subsection 1.3.1, a pipe engineer needs complete information
before he can start to route a pipe. Since our goal is to incorporate the manual
routing process into the automatic one, we need to identily what data is available
and needed.
Routing is a difficult task for pipe designers and the expertise of the pipe
designer is very important to assure that all pipes are routed in a proper way.
Currently, formal guidance on how to route a pipe in a ship does not exist. We only
can find specification requirements of pipes and systems in the standardization
books such as in a marine classification guide, or in scattered parts of a few books
that describe piping systems. Nevertheless, there are some informal rules that
have evolved into a common body of knowledge among pipe designers on how to
route a pipe that belongs to a certain system, inside a particular area in the ship.
Most of that common knowledge is based on the logical way of routing that is
acquired by experience. An example is that pipes that run in the same direction
and lie close together should be routed in parallel. Also pipes should be routed in
a certain order based on the diameter of each pipe and the system that particular
pipe belongs to.
In this thesis, we investigate the common knowledge of how to route pipe, and
the adoption of that knowledge to be implemented in the methodology. Thus, we
seek answers to the second, third and fourth research questions below:

R e s e a r c h q u e s t i o n # 2 : What information is needed to perform the pipe rout-

ing process, who is responsible to provide it and how can one get it?

R e s e a r c h q u e s t i o n # 3 : What is the common knowledge to route pipes that is

used as guidance by a pipe engineer?

R e s e a r c h q u e s t i o n # 4 : Which knowledge from Research Question # 3 should

be adopted in our proposed methodology and to what extent can the current
practical knowledge be absorbed into a programmable methodology?

1.3.3 Quality of the Pipe Routed

In the current practice, pipes that already routed are claimed to have a good
quality if the following conditions are fulfilled:
8 INTRODUCTION 1.3

1. The functional requirements are satisfied; pipes are connecting pieces of

equipment perfectly without having excessive length, unnecessary bends
and without collisions,
2. It complies with maritime rules and regulations as imposed by classification
societies,
3. It fulfills the subjective values of the pipe designer who has routed those
pipes; this judgment is very subjective but still in line with the common
knowledge of how to route a pipe.
Judging if the above conditions are met, can only be performed by manually
examining the routed pipes. In addition to the subjectivity involved, analyzing
the quality of a result can be very exhausting. Moreover, due to the one-off nature
of many ships makers, it is almost impossible to make an objective comparison
between ships. To overcome this inconvenience, it is important to transform the
qualitative subjective judgment to quantitative objective analysis.
The quantitative way of scoring the quality of a result set can be useful, not
merely to figure out the quality of the final result, but also to assist the pipe
engineer and the pipe router algorithm to find the best solution, or at least a very
good solution.
For that reason, we strive to answer our fifth research question:

R e s e a r c h q u e s t i o n # 5 : How to evaluate the quality of a set of routed pipes

quantitatively?

1.3.4 Design Requirements

Before we actually design the automatic routing system, we need to follow some
practical requirements. The first requirement is to make sure that the system
should be efficient and effective and allow the user to find the pipe route with
less effort than by manual routing using existing CAD software. This includes
the ease with which the methodology can be prepared and all the needed data
gathered. In addition to that, the solution that is found should be optimal and
comply with the marine classification rules and regulations.
When pipe routing is done manually by pipe designers, the quality of the
solution highly depends on the expertise of the pipe designers. One of the goals
of the first requirement is to reduce the differences in the solution quality between
pipe designers with different expertise.
The methodology needs some data before it can be used to perform the auto-
matic pipe routing process. As described in the procedure of the detailed design
process, this methodology needs the following data as its input: the piping sys-
tem diagram, 3D model of the component arrangement, 3D model of the hull
and superstructure, and additional data of the system components such as the
equipment's connection points and the type of a component.
1.4 RESEARCH APPROACH AND LABORATORY 9

In the current pipe design process all information that is needed is available,
either already in the CAD software or in some other format. Normally, during the
design process the ship is modeled using CAD software, including the 3D volume
of the hull and components, and all the underlying data. Because of this, the
automatic routing system that is developed should be able to exchange data with
the existing CAD software. By this connectivity, the user can use the existing
CAD software and the automatic rotiting system simultaneously.
Currently there are several CAD software packages that are widely used by
shipyards as design tools, and in some shipyards more than one kind of CAD
software package is used. This requires the automatic routing methodology to be
developed as a non-proprietary set of instruments. The automatic routing system
should be able to be used together with any kind of CAD software.
The marine classifications and the common rules for pipes are constantly im-
proving. New rules and knowledge emerge that must be adopted by the automatic
routing system. As one of the requirements, the automatic routing must be ex-
pandable to allow for such changes.
One of the requirements is to reduce the time needed during the detailed
design. The proposed methodology should be implemented carefully, require smal-
ler amounts of operator time and its computation time needs to be optimized.
In short, there are four main requirements that have to be fulfilled by a proper
pipe routing methodology: to minimize user input and reduce user routine tasks;
to be integrated easily with other systems, such as existing CAD software; non
proprietary and expandable; and able to produce good results within an accept-
able time frame.
This research mainly focuses on methodology development. Its implementa-
tion in the form of an automatic routing system package is mainly targeted as
a laboratory to investigate and prove the effectiveness of the methodology itself.
The development of the tools ready to be used in production is of course not
within the scope of this thesis.
These requirements are tightly related to the development of the automatic
routing methodology and to the implementation of this concept into actual tools.
While pursuing these tasks, we seek answers to the sixth research question below:

R e s e a r c h q u e s t i o n # 6 : How can a routing algorithm be efficiently implemented

that satisfies the four key requirements and works in any given 3D environment?

1.4 Research Approach and Laboratory

This thesis focuses on the creation oi the methodology and its application for the
pipe routing process in ship design. It starts with the development of the standard
procedures for pipe routing. These are based on the common sense and experience
of the pipe engineers when they route pipes manually in a ship. In practice, this
10 INTRODUCTION 1.5

can be divided in two categories. The first category is related to the functionahty
of the pipes. For example, pipes that run together in the same direction should
be aligned in parallel, or the hot water pipe and cold water pipe should be routed
together for ease of maintenance. Such standard procedures should be possible
in the proposed methodology. The second category is the common sense of pipe
engineers that can be categorized as their personal preferences, and this category
can be neglected.
T'he behavior of how the pipe engineers do the pipe routing in different parts of
a ship is also investigated, because in practice pipe engineers might use a different
approach for different pipe locations.
The rules and standards of marine classifications must also be adopted to
ensure that the results of the proposed methodology are allowed to be used in
practice. However in this thesis not all of the marine classifications are included,
but the method for including and implementing those classifications is investig-
ated.
Within this context, suitable routing algorithms are investigated and de-
veloped. Since much research on automatic routing has been done for many years,
there are many different algorithms that have been proposed. Those algorithms
are developed not only for routing of pipes, but often for other subjects such
as microchip design; to route the cables in airplanes; ground traffic navigation
systems; and finding the route for character movement in computer games.
Those existing algorithms are analyzed, tested and compared. Based on those
test results a combination of more than one algorithm is developed, and included
in the proposed methodology.
After defining the suitable routing algorithm for the methodology, the data
that is needed by the algorithm can be identified. There are two main categories
of data identification; identification of the type of data and the source of data.
Because there are many types of data and also many sources of data, the generic
type of data is defined.
In the pipe routing process we are dealing with a 3D environment. All com-
putations in the automatic routing algorithms included in the methodology are
done in the 3D environment. For that reason, to allow building our laboratory,
we had to develop the specific 3D library.
In order to prove that the proposed methodology can fulfill the research ob-
jectives stated in section 1.2, a laboratory is developed in the form of a computer
application package that contains the test implementation of the proposed meth-
odology. The laboratory consists of three applications; the interface module for
data exchange with the outside world, the router module that contains the auto-
matic pipe routing algorithm and the knowledge-based module that contains the
standard rules and routing criteria.
Using that laboratory, the proposed methodology has been thoroughly val-
idated and verified through experimental work in a realistic pipe routing design
process in a complex area of a ship. The experiments presented were carried out
to find the solution in the detail design phase of the ship.
1.5 ORGANIZATION OF THE WORK 11

1.5 Organization of the Work

This book consists of 7 chapters, starting with this chapter, the Introduction.
The next chapter discusses the current practical pipe routing process in a ship.
We start this chapter by describing the stakeholders and their functions in the
pipe routing process. Then, we explain the tools and information data that are
needed by the pipe engineer. We also describe the essential steps of the pipe
routing process. In this chapter, the common rules for pipe routing in a ship
are discussed and the common mistakes that are made by an inexperienced pipe
engineer are shown. Based on that, we conclude this chapter with the list of rules
that must be included in the proposed methodology.
Chapter 3 starts with the description of the outline of the functional framework
of the proposed methodology. It translates the common pipe routing rules found
in practice that in principle lends itself to practical, real-life application, to the
functional requirements. Since we intend to create a methodology, the procedure
of data retrieval and user interfaces are included in the functional requirements as
well. Then we review the state of the art of the automatic pipe routing method.
In this chapter, several well-known automatic pipe routing researches are reviewed
and discussed.
We start Chapter 4 by describing in detail shortest path algorithms. It starts
with the deterministic method and is followed by the heuristic method to solve
the shortest path problem. Then, the most popular algorithms are compared.
Based on this result the suitable algorithms are selected. In this chapter we also
discuss the other research subjects that are mentioned in Chapter 3.
Chapter 5 extends the outline of the functional framework that have been
discussed at Chapter 2 and 3. In this chapter, we investigate and implement our
proposed methodology into a pipe routing tools to perform the automatic pipe
routing process.
Chapter 6 tests the performance of our methodology. It starts by showing
the capability of the methodology to route pipes in the machinery room. Then,
we continue to discuss the quality of the result. The sensitivity analysis of the
proposed methodology is tested and described.
Chapter 7 contains the conclusion of this research and some suggestions for
useful future research.
Chapter

Pipe Routing in Practice

2.1 Introduction

Chapter 1 explained the need to have an innovative approach to route pipes

in a vessel to remain competitive in the global ship manufacturing market and
indicated the considerable efforts that have been made in this thesis to achieve
significant improvements to the pipe routing process. In Section 1.2 we described
our main goal to research a methodology for an innovative approach to route pipes
in a ship in such a way that tools that will be created based on that methodology
can be used in a real ship design process.

The proposed methodology must be built based on the proven pipe routing
process i.e. the actual pipe routing process in practice. This also helps to make
sure that future users of this methodology feel at home. This is essential because
it will increase the level of acceptance of automatic (pipe) routing.
Section 1.3.2 provided an overview of the current situation in the practical
design process. In this chapter, the current practical process will be described in
more detail and in a structured way and the common knowledge of pipe routing
will be summarized. This chapter is concluded with the list of aspects that must
be met to realize the targeted methodology.

The starting point of this chapter is by looking at the pipe routing as a project
that should be managed, executed, and monitored efficiently. There arc three
core areas relevant for our purposes: people (or stakeholders), tools (including
information), and process, as can be seen in Fig. 2.1. The next sections of this
chapter explain those three core areas for the pipe routing process.

13
14 PIPE ROUTING IN PRACTICE 2.2

Marine engineer

Equipment spec W Functional diagram

General arrangements
Tank plans
Compartment plans
etc

Section designer,

Sections

Pipe Routing _^i

Pipe routing
Spool splitting

Pipe engineer

Figure 2.1: Pipe Routing Processes

2.2 Stakeholders

By a definition, a stakeholder is a party that affects or is affected by the actions

of the business as a whole. In other words, a stakeholder can be defined as " those
groups without whose support the organization would cease to exist.". Freeman
and Reed [1983].
As described in Section 1.2, pipe routing is one of the most significant activities
during the detailed design stage because all other activities in that stage depend
on it. Therefore, many departments of the shipyard and its co-makers have a con-
nection with the pipe routing process. 'I'he client, the procurement department,
the basic design group, right through to the production and service departments
and also the classification society; all of them can be considered as stakeholders
of the pipe routing process. However, in this chapter we limit ourselves to the
particular stakeholders that have a direct effect on the task of individual pipe
engineers.
2.2 STAKEHOLDERS 15

Project Engineering IVIanager

The project, engineering manager^ is ihe person responsible for all engineering
activities and all disciplines in the project: process, naval architecture, structural,
mechanical, electrical and instrumentation. In the pipe routing scope, the main
task is to assemble and supervise pipe routing groups, and together with the pipe
group leaders determine the following items:
1. Inventory of the required specifications, such as piping classes, equipment
spacing requirement, the dimension of the walkways
2. Manpower planning
3. Defining what CAD software package will be used in the project
4. Dividing the ship into working areas and assign each area to (a group) of
pipe engineers
5. Align the work with production planning

Pipe Group Leader

The pipe group leader has a responsibility to lead a group of pipe engineers that
route pipes in one or more working areas. The group leader must have a hands-on
experience to route pipes and an excellent capability to judge routing results both
in terms of technical expertise and management expertise.

Pipe Engineer
The pipe engineer is the person who performs the task to route pipes in his
working area. Every pipe engineer must have an excellent knowledge of piping
systems in a ship and must be lamiliar with using the 3D software package. I'he
quality of the routed pipes depends on the expertise of the pipe engineer. Most
of the time, a pipe group leader also acts as one of the pipe engineers.

CAD Software Administrator

Currently, pipe engineers create piping systems using CAD software as their tools.
In the last decade, the 3D CAD software packages have evolved into complete
suites that allow many users to work together to perform multiple tasks using the
same model. The complexity of the software package requires it to be maintained
by an administrator.
In the beginning of the ship design project, the administrator configures the
CAD software according to the standard that is required by the contract. An
example is to define the minimum bending radius of pipes.

^Note that this may be organized difTcrently in companies but the essence of tlie process
remains the same. The same applies to the other functions in this section; they express the
various roles that need to be fulfilled in the process.
16 PIPE ROUTING IN PRACTICE 2.2

Marine Engineer
As described in Subsection 1.3.1, to support tlie functionality of a ship, various
systems such as propulsion and steering are needed. In its operation, those sys-
tems need some other subsystems in order to run properly. For example, a diesel
engine needs fuel to run, needs oil for its moving parts, its temperature must be
maintained in its operation range, and so on. For those purposes it needs to have
a fuel system, lubrication system, cooling water system, and many other systems.
The marine engineer is the person responsible to translate the specification
and the systems previously mentioned into the functional requirements and dia-
grams. The marine engineer must have an extensive knowledge of the equipment,
instrumentation and functionality of the process.

Pipe and Instrument Diagram Engineer

Before the functional diagrams are used by pipe engineers, they are translated into
Piping and Instrument Diagrams (P&ID). P&ID is basically a functional diagram
that is enriched with topological data of the equipment, piping and instrument.
The Piping and Instrument Diagram engineer is the person responsible to
translate the functional diagrams previously mentioned into the P&ID. Just as
the marine engineer, the P&ID engineer must also have an extensive knowledge
of the equipment, instrumentation and functionality of the process.

Naval Arcliitect
The naval architect is responsible for five elements, Lewis [1988]; Hydrostatics,
Hydrodynamics, Structure, Arrangement and Constructions. In short, a naval
architect engineer is responsible to make sure that the ship supports its main
functions efficiently and safely. This includes floating, moving and carrying.

Section Designer
The Section^ designer is responsible to render the construction plan into a 3D
section model while maintaining the requirements imposed by the naval architect.
It includes defining all brackets, profiles, plates and holes in a section.
During the detail design phase, if it was needed, a pipe engineer makes a
request to the section designer to make small changes in the section in order to
be able to route a pipe optimally, such as moving a location of a hole in a plate.
In this case, before the section designer makes any changes, he must get approval
from the naval architect to ensure that the changes do not have a negative effect
on the ship's structural integrity.

•^A section is a production specific part of a complete ship's structure. The subdivision into
section is made e.g. to take account of crane capacity and to increase production efficiency.
2.3 TOOLS AND INFORMATION 17

3D Model Builder
3D CAD software uses extensive libraries including the 3D model library, This
consists of the common components and normally can be re-used for other pro-
jects.
However, since there are many customized components, it is not possible to
cover everything from the library. The 3D model builder is the person responsible
to create new models of components and add them to the library for future use.
The quality of the 3D model is important in the pipe routing process, a com-
plete 3D model of one component will help a pipe engineer to get all the necessary
data of that component.

Ship Owner
The ship owner is the (future) owner of the ship that is currently being built in
the course of a project. This is the stakeholder who determines the type and the
capabiüties of that ship.
During the whole phase of ship building, progress must be reported to and ap-
proved by the ship owner. What makes things more complicated for the shipyard
is the fact that during that process some specifications might be changed by the
ship owner. One of the parts that is very sensitive to changes is the piping system.
A very small ship specification change can cause to a lot of pipes to be re-routed.

Marine Classification Society

A classification society is a non-governmental organization that establishes and
maintains technical standards for the construction and operation of ships and
offshore structures. Classification societies set technical rules, confirm that designs
and calculations meet these rules, survey ships and structures during the process
of construction and commissioning, and periodically survey vessels to ensure that
they continue to meet the rules.
Today there are a number of classification societies, the largest of which are
Det Norske Veritas, Lloyd's Register, Bureau Veritas and the American Bureau
of Shipping. In particular, classification societies may be authorized to inspect
ships, oil rigs, submarines, and other marine structures and issue certificates on
behalf of the state under whose flag the ships are registered.
Every classification society defines its own rules and standards. However many
rules are similar from one society to another. Normally each releases a set of books
of Rules for Classification of Ships twice a year.

2.3 Tools and Information

Tools have two main functions, to support or amplify a person's efforts allowing
them to be more efficient doing their tasks, or to replace a human operator when
18 PIPE ROUTING IN PRACTICE 2.3

a tool can do Ihc job more effectively. Some tools are focusing on one of these
things, but most tools combine elements of both.

2.3.1 Tools
3D CAD Software - Outfitting Tools
Nowadays, the outfitting tools in 3D CAD software is the most essential piece
of equipment to help pipe engineers to route pipes. To route pipes in a ship
there are many aspects that need to be handled correctly by a pipe engineer;
an example is to choose the pipe correctly according to the specification in the
P&I Diagram, or to make sure that there is no collision between pipes and/or
with other components in the ship. With the help of 3D CAD software, the pipe
engineer can solve those tasks using the function in the software.

3D CAD Software - Diagram Tools

The diagram tools are needed by the pipe and instrument engineer to create
the piping and instrument diagram. By using advanced diagram tools, the P&I
diagram can be easily created with the help of the system library. Also, the P&I
diagrams that are created have a connection with the 3D model that will be used
by a pipe engineer. In that way, the consistency between the diagram and the 3D
model is ensured.

2.3.2 Inforination
Functional and P&I Diagram
The functional diagram consists of graphical symbols and lines which illustrate
the process and its flow. It identifies the functions of its instruments such as
sensors, valves, indicators and instrument interconnections. The P&I diagram
basically almost the same with the functional diagram. The main difference is
that P&I diagram also contains the location of the pipes and instruments. It acts
as the primary guidance for a pipe engineer to accomplish his task.

Section and Construction Plan

The construction plan is a drawing that was generated during the basic design
phase. It is the basis for the section drawing that will be made in the detailed
design phase. A section is a 3D steel construction that represents a construction
plan in detail. The construction plan is not the only guidance to create a sec-
tion. The section designer must always ensure that the section complies with the
specification that was issued by the naval architect.
Beside the P&ID, the section is also necessary before pipe engineers route the
pipes. In some cases, when section details are not available, pipe engineers may
have to use the construction plan.
2.4 PIPE ROUTING PROCESS 19

Piping Specification

The required specifications can be separated into two categories, based on their
sources; the marine classification society and the demand from the ship owner.
The specification from tiie marine classification is absolute; it means that no
matter how hard it is to be implemented, it must be followed. Otherwise, the
ship will not comply with the marine class and it will not be certified by the class
organization.
The request from the ship owner, however, can be negotiated. It means that
some conditions can be violated as long as the ship owner agrees with it. The
ship owner may be compensated by other benefits, for example, if the changes
will cause the pipes to be shorter and thus cheaper to maintain.

Basic Design Information

Pipe engineers also need other information like the general arrangement and the
specification of the equipment, the compartment plan and the tank plan. All of
this information is available from the basic design phase.
Beside this information, the complete 3D models of components such as pumps,
engines and other equipments are also needed.

Component 3D IVIodel

The 3D model of a standard component usually is available in the CAD software.

For customized components, a 3D model must be created. Ideally, the component
3D model should come from the component supplier. For example, if the shipyard
buys a main engine from a supplier, next to the main engine the supplier must
also provide its 3D model.
However, in practice the component 3D models are often created by the
shipyard. There are two main reasons for this. The first is the fact that the model
from the supplier has too much detail. Most of the time all the bolts and small
holes are included in the model, which is not feasible to be used by the shipyard
because it reduces the performance of the CAD software. The second reason is
that sometimes the component supplier charges additionally to the shipyard for
supplying the 3D model.

2.4 Pipe Routing Process

2.4.1 Organizational Process
In order to execute a project efiiciently, before the process of pipe routing in ship
can be executed, it is necessary to plan the project. The first thing that must be
done is to determine the manpower that is available. The manpower planning is
extremely important. In practice, more than one ship is build at the same time
20 PIPE ROUTING IN PRACTICE 2.4

in a shipyard. This means that the available manpower is noL always Lhe same,
and if needed, more pipe engineers will be hired temporarily for a certain project.
I'he expertise and skill of the available pipe engineers also need to be known
beforehand. In a shipyard that uses more than one type of CAD software, the
choice of which CAD software will be used depends on the available manpower.
In some projects, the pipe system design process of a single ship is done in two
different CAD software packages. The reason is that some of the pipe engineers
are only able to use a certain type of CAD software and the others are only able
to use the other type of CAD software.
The allocation of working area and the pipe engineer group assignment also
depends on the available manpower at that time. Normally there are 3 to 7 pipe
engineers in one group, depending on the size and the complexity of the working
area.
Inside each pipe group, the pipe group leader divides the working area into
smaller areas and assigns each member of the group to route pipes in those areas.
However, there can also be the case where more pipe engineers work together on
the same area, and they each have to route pipes for a different system.
In order to be more cost effective in the production stage, most of the pipes
should be installed during the pre-outfitting stage. This means that the pipes
that belong to one section of the ship must be installed immediately after the steel
construction of that section is completed, so that when that section is assembled to
the hull on the slipway, all pipes that belong to that section are already installed.
This situation only can be reached if the pipe spool drawings of those pipes are
available a few weeks before section assembly starts. This allows that the pipe
spools can be manufactured in time. For that reason, it is especially beneficial
for the pipe routing process to be synchronized with the production plan.
Every pipe group leader then makes a comprehensive planning for their group.
Pipes that require more time to manufacture compared to the average get a higher
priority. This task demands a high level of expertise from the pipe group leader.

2.4.2 Routing Process

After a pipe engineer received an assignment from the group leader, he starts
with collecting all the data needed. He collects the information with respect to
the section, P&ID, piping specification and other drawings from the basic design
phase.
Then the pipe engineer examines the section to evaluate the spaces that are
usable for pipes. In most cases, the section 3D model is available, but not always.
If this is the case, the pipe engineer must resort to the construction plan. It adds
difficulty to routing pipes in that area.
At the same time, the pipe engineer also needs to examine the P&I Diagram,
and in combination with the usable space in the section, he can start to determine
how the pipes should be routed.
2.4 PIPE ROUTING PROCESS 21

Most of the time, the components of the ship such as main engines, pumps,
or other equipment are not placed in the 3D environment yet. In that case, the
pipe engineer must place those components based on the general arrangement
drawing. However, the main components such as the main engine and the other
system that related with the mechanical drive system are already defined during
the basic design process, and the information is available for the pipe engineer.
In the ideal situation, starting from that point, the pipe engineer can begin
to route the pipes according to the criteria of pipe routing. This is discussed in
detail in Section 2.5.
Then the pipe engineer routes pipes one by one. After one or more pipes in
the area are routed, all valves will be placed. Then those pipes are split into
pipe spools for the purpose of production, handling and assembly. Thanks to the
current CAD software functionality, the process to place the valve and to split
a pipe into pipe spools have been improved. However, even though most of the
latest CAD software package already include the automatic routing functionality,
it still only focus of finding a path between two nozzles^, and basically the total
layout of the pipes is still defined highly depends on the pipe engineer.
As we described in Subsection 1.3.1, a ship has many systems and subsystems
that consist of a large number of pipes. Even a small vessel might have more
than one thousand pipe and in a larger vessel this number can reach three to four
thousands pipes and more. In short, many pipes must be routed, while the space
that is available is limited. In order to accomplished the task to route all pipes
efficiently and comply to the rules, pipe engineers must employ a smart routing
strategy. The summary of this strategy is explained in Section 2.6.

2.4.3 Working Area

During the routing process, a pipe engineer needs to make an assessment of his
working area. It is needed because the way pipes are routed depends on some
aspects. First, it depends on the pipe specification itself; to which system that
pipe belongs, the pipe diameter, the pipe material, the surface treatment of the
pipe, and specific requirements for the pipe such EIS if it must be installed sloped
down or not. The second aspect is the number of pipes that must be routed in a
certain area. The last one is the technical aspect of the system itself. Things like
the placement of pipes and the sprinklers for the system that intend to be used
to wash the cargo area is part of this.
Based on those aspects, the areas of a ship can be categorized into three
different types:
i. Machinery type area
2. Accommodation type area
3. Technical type area

^Nozzle is the term that widely used in the 3D CAD application to represent the pipe
connector or pipe end. In this thesis, we adopt this term and use it in all chapters.
22 PIPE ROUTING IN PRACTICE 2.4

The compartments in the ship that are categorized as the machinery type area
are the main engine room, auxiUary engine room and pump room. This type of
area can be considered to have the highest density in terms of the number of
pipes per cubic meter. Also because there are many pipes there, more than one
pipe engineer will be assigned to route pipes in that area. As a consequence, they
must have a good coordination to make sure that the pipes are routed in the same
manner and do not violate each other.
The greatest difficulty of routing pipes in this type of area beside the large
number of pipes is the fact that one pipe can follow many different paths. There-
fore, a pipe engineer must find the optimal combination of pipe path to ensure
that all pipes can be routed properly.
The accommodation type area consists of the accommodation spaces and the
control room. In this type of area, the number of pipes is not as large as in the
machinery type area, but the space that is available is extremely limited. In the
accommodation area, all pipes must be hidden, either below the floor or above
the ceiling. Beside the limited space, the task to route pipes in this area is further
complicated by the rule that requires the black and gray water system pipes to
have a slope.
In accommodation areas, normally a pipe engineer is already able to figure out
the rough path of each pipe. However, since the space is very limited, he must be
able to make a good arrangement of pipes, and avoid conflicts.
The technical type area is the area of the ship where other systems are placed.
For example in a dredging ship, a technical area is where the large dredge pipes
are located.
In terms of finding the path of the pipe, this area normally has the lowest
difficulty level compared to other area types. However, routing pipes in this type
of area is still difficult since one needs a deep knowledge about the technical
system itself.

2.4.4 Object Constraints

During the routing process, a pipe engineer considers every object (e.g. a com-
ponent, another pipe or a piece of the steel structure) and area in a ship according
to the possibility to be penetrated or used by pipes. Thus, an object or an area
is no longer seen as its functionality but translated into an object constraint that
can be categorized into four types;
1. Absolute Constraint
2. Soft Constraint
3. Negotiable Constraint
4. Rules Constraint
The absolute constraint refers to an object or an area that is absolutely not
allowed to be passed by any pipe. An example of this type of constraint is a piece
of equipment, or an area outside the ship.
2.4 PIPE ROUTING PROCESS 23

The soft constrain!, is an object or an area that is preferably not passed by a

pipe. Basically this constraint should be satisfied, but if there is no other way, it
is allowed to be violated. In another word, pipes are allowed to be routed through
this area, but there will be a penalty for that. An example is the edge part of the
walking corridor in the machinery room.
The negotiable constramt is an absolute constraint that in a certain situation
can be negotiated to be passed through by a pipe, by changing or moving the
object or by modifying the pipe itself. An example is a stiifener of the steel
construction. In a normal situation, a pipe engineer must avoid routing a pipe
through a stiffener. However, if there is no other way, or if there is a big advant-
age to route a pipe through that area, it is possible to negotiate with the naval
architect to change or move that particular stiffener.
The rules constraint is an object or an area that becomes an absolute or a
negotiable constraint for a certain pipe due to the pipe routing rules. For example,
the area above the combustion engine becomes a negotiable constraint for an oil
pipe.
Normally, a modern CAD software package can detect a pipe violating an
absolute constraint and some of the negotiable constraints. It then can generate
a warning to the pipe engineer. However, the soft and rules constraints are not
actually available in a CAD package, thus they can only be detected by the
expertise of the pipe engineer.

2.4.5 Collaboration between Pipe Engineers

Every pipe engineer is responsible for his own task to route pipes in his or her
own working area and/or for the system assigned to him. They can work on their
own for some pipes that connect two nozzles in their own area. However, most
pipes are crossing other people's areas.
For example, in the bilge and ballast system pipes are running from aft ship to
fore ship. In practice, even for those pipes, every pipe engineer is only responsible
to route pipes in his own area. However, he must think about the whole route of
those pipes, so that when he sets the location where they leave his working area,
it will not cause a problem for the responsible pipe engineer of that neighboring
area. In this situation, pipe engineers should communicate with each other.

2.4.6 Modification Possibility

In an ideal situation, the available detailed section structure is suitable for all pipes
to be routed well. In practice, it happens quite often that the existing section
structure must be changed to enable the pipe engineer to route pipes there. In that
situation, the pipe engineer may ask to the section designer to change the existing
section design. Then, after consulting with the naval architect, if the changes do
not affect the structural integrity of the ship or other functional requirements, the
existing section can be modified.
24 PIPE ROUTING IN PRACTICE 2.5

In the other way around, there are also cases where the pipe engineer had
finished routing pipes in a certain location of a section, but due to the changes in
the ship specification, the naval architect requests the section structure designer
to change the section structure. Those changes are then communicated to the
pipe engineer, and he must make the corresponding modifications.
There is also a possibility that the pipe engineer makes a request to the P&ID
engineer to change the way some pipes should be logically arranged.
The expertise of the pipe engineer plays a big role in all of this, because the
more knowledge and experience the pipe engineer has, the more creative he can
be. He can then make a prediction that the changes in the section structure or in
the P&ID are feasible without affecting the integrity and/or the performance of
the process system in the ship.

2.5 Criteria for Pipe Routing in Ships

In Subsection 2.4.2 it is mentioned that when he routes pipes in a ship, every
pipe engineer must consider the criteria of pipe routing to ensure that the pipes
are routed according to the rules and standard. In practice, pipe engineers follow
two main criteria that can be categorized as follows:
1. Pipes must comply with the rules and standard from the Marine Classific-
ation Society involved in that ship
2. Pipes must be routed in such a way that the production, installation, and
maintenance cost of those pipes are as low as possible

2.5.1 Rules of Pipe Routing

For safety reasons, pipe routing must follow the rules from a marine classification
society. There is a special chapter regarding the rules and standards for piping
systems in ships that consists of many rules. Those rules can be categorized into
two kinds. The first one is the rules that are related with the specification of the
pipe including the piping components, such as flanges, valves, supports.

Piping system for Class I Class II Class III

p(bar) t(C) p(bar) t(C) p(bar) t(C)
Steam, thermal oil > 16 or > 300 < 16 and < 300 < 7 and < 170
Fuel, lubricating oil > 16 or > 150 < 16 and < 150 < 7 and < 60
Other media > 40 or > 300 < 40 and < 300 < 16 and < 200

Table 2.1: Classes of piping system (excerpt from DNV)

This kind of rules is applied in the P&I Diagram by the P&ID engineer, and
later on the pipe engineer must use it as a guidance. As an example, one of these
rules relates to the specification of the material, shown in Table 2.1.
2.5 CRITERIA FOR PIPE ROUTING IN SHIPS 25

In thib section, the rules ol'pipe routing as described by a marine classification

society will be described. The following list of rules only contains selective rules
that have a direct influence on the way a pipe engineer performs his task. The
complete list of rules can be seen in the classification books.
1. T h e n u m b e r of d e t a c h a b l e p i p e c o n n e c t i o n s shall b e limited t o
t h o s e which a r e n e c e s s a r y for m o u n t i n g a n d d i s m a n t l i n g
While dividing a pipe into several pipe spools, the pipe engineer must choose
to use flanges or a welded pipe. In practice, a pipe engineer must make a
clear choice if a pipe must be welded in or must be removable. For example
in technical spaces or on a connection to a ship component, every pipe must
be removable. In those cases flanges must be used. However, in electrical
spaces every pipe must be welded.
2. T h e s u p p o r t of t h e p i p i n g s y s t e m shall b e such t h a t d e t r i m e n t a l
v i b r a t i o n s will n o t arise in t h e s y s t e m
To fulfill this rule, every pipe must be routed close to the steel construction
for ease of the installation of the pipe support. One of the reasons that pipes
that run together should be routed in parallel is because the pipe support
installation is easier.
3. • I n s t a l l a t i o n of p i p e s for w a t e r , s t e a m or oil b e h i n d or a b o v e
electric s w i t c h b o a r d s shall b e avoided as far a s possible. If
t h i s is i m p r a c t i c a b l e , all d e t a c h a b l e p i p e j o i n t s & valves shall
b e a t a safe d i s t a n c e from t h e s w i t c h b o a r d or well shielded
• All d e t a c h a b l e p i p e c o n n e c t i o n s a n d valves in oil fuel p r e s s u r e
p i p i n g shall b e a t a safe d i s t a n c e from boilers, e x h a u s t p i p e s
or o t h e r h e a t e d surfaces a n d electrical a p p l i a n c e s
• D e t a c h a b l e p i p e c o n n e c t i o n s a n d valves in h y d r a u l i c p r e s s u r e
p i p i n g shall b e a t a safe d i s t a n c e from electrical a p p l i a n c e s ,
boilers, e x h a u s t p i p e s a n d o t h e r sources of ignition
Those three rules above have the same goal: to prevent the possibility of
fire. Before a pipe engineer begins to route those pipes, he must make a
good evaluation of the space management.
4. • Centrifugal bilge p u m p shall b e l o c a t e d a s low as possible
• Centrifugal s e a - w a t e r cooling p u m p s shall b e installed as low
as possible in t h e ship
Those two rules above are suggesting that the bilge and the sea-water cooling
system pipes should be routed aa low as possible too.
5. T h e overflow s y s t e m shall b e so a r r a n g e d t h a t w a t e r from t h e sea
c a n n o t e n t e r t h r o u g h t h e overflow m a i n line into o t h e r t a n k s in
case of a n y t a n k s b e i n g d a m a g e d
This means that the overflow pipes should be routed in a slope. However,
in practice it is also allowed (o have a horizontal pipe
6. P i p i n g conveying flammable liquids u n d e r p r e s s u r e in t h e engine
a n d boiler r o o m shall b e laid in well lit places, in o r d e r t h a t t h e
piping may b e kept under observation
26 PIPE ROUTING m PRACTICE 2.6

11 means that for those kind of pipes, a pipe engineer must try to route it
in a space that can be seen.
7. • Bilge s u c t i o n p i p e s a r e , as far as p r a c t i c a b l e , n o t t o b e c a r r i e d
through double b o t t o m tanks
• T a n k air p i p e s shall b e p l a c e d a t t h e h i g h e s t p a r t of t h e t a n k
a n d as far away as possible from t h e filling
• W a t e r p i p e s a n d air a n d s o u n d i n g p i p e s t h r o u g h freezing
c h a m b e r s shall b e avoided
• Fuel oil p i p e s shall n o t b e led t h r o u g h fresh w a t e r t a n k s
• T h e a r r a n g e m e n t of p i p i n g a n d valves shall b e such t h a t oil
c a n n o t e n t e r t a n k s n o t i n t e n d e d for t h i s p u r p o s e
Those five rules above are about avoiding to route a certain type of pipe in
a forbidden location,
liaijically, the list of rules above apply to all class society, however the details
may vary.

2.5.2 Minimization of Pipe Cost

In the pipe routing process, routed pipes that merely comply with the maritime
rules will not be good enough. Every pipe engineer must also consider to minimize
the pipe cost. There are three elements of the total cost of a pipe; pipe material,
production and installation cost. To calculate the pipe cost is not trivial and
most inexperience pipe engineers only consider minimizing the length of the pipe.
Meanwhile, the more experienced pipe engineer can easily estimate the total cost
beised on the last two elements above.
To minimize the production cost, it is not enough only to minimize the length
of the pipes, but the diameter and the thickness must be considered as well. A
pipe engineer should be able to identify that a pipe with a small diameter can
be much more expensive compared to a pipe with a larger diameter if the wall
thickness of the smaller pipe is larger than normal.
The knowledge of the piping system is also important to be mastered. For
example, a pipe engineer must know by head that the piping system that transfers
sea water must have a special treatment to prevent corrosion. Table 2.2 shows
systems that normally need a special pipe treatment and Table 2.3 shows some
common knowledge of pipe routing based on the pipe system type. Table 2.3 shows
that for a certain system, pipes are preferably routed below the floor rather than
through the top. Also according to the classification rules, pipes for some system
must be routed as low as possible.
The minimization of the installation cost is helped by having the pipe spool
sketch ready for production in time, so that the pipe can be installed during the
pre-outfitting stage. Beside that, a pipe must be divided into pipe spools in such
a way that the pipe spools can be handled easily. The installation cost would
also be reduced by a clever arrangement of supports and this can be achieved by
having pipes to be routed in parallel.
2.6 CRITERIA FOR PIPE ROUTING IN SHIPS 27

System Name P i p e Treatment

Seawaler Bilge & Ballast Galvanized
Fire Fighting Galvaiii/,('(i
Cooling Water RILSAN/Ai5CrrE/Galvanized
Sanitary Deck Drains Galvanized
Sanitary Galvanized
Sounding Water Tank Galvanized
Oil Tank Galvanized

Table 2.2: List of Systems Lhal need a special lreatm,eni

System Name Below Lowest Remarks

Degassing Large diameter
Jctwater Large diameter and heavy thickness
Drauglil Measuring Slope down
Lubrication Oil Bowthruster Slope down
Starting Air Expensive pipes so keep it short
Fuel Oil Transfer Y N For overflow slope down
Fuel Oil Service Y N
Lubrication Oil Transfer Y N
Lubrication Oil Service Y N
Dirty Oil and Sludge Y N Slope down
Sea Cooling Water Y Y Large diameter, expensive treatment
Fresh Cooling Water Y N
Bilge Y Y
Ballast Y Y
Firefighting and deckwash High pressure
Air, Filling, and Sounding Keep it straight, bending is < 30

Table 2.3: Fipmg system knowledge

28 PIPE ROUTING IN PRACTICE 2.6

2.6 Pipe Routing Common Knowledge

The common knowledge to route pipes is gained by every pipe engineer by ex-
perience, following the rules that are supplied by a marine classification society.
However, this knowledge is not normally properly documented. We performed
many interviews with experts in pipe routing to absorb their knowledge. During
the interview period, the pipe routing guideline was compiled by one of the very
experienced pipe engineer team leaders v.d. Berg [2009] who summarized the
most important parts of the pipe routing common knowledge.

2.6.1 General Guidance of Pipe Routing

1. For pipe routing in crowded areas it is recommended to maintain layers of
pipes with the same direction at the same elevation, (longitudinal direction,
cross direction)
2. Think ahead during routing pipelines. Large diameters first. Be aware that
the pipe you are currently working on is not the last one to put in. Especially
in machinery area the available space for piping is limited. Begin to route
pipes from Bilge, Ballast, Sea Cooling Water, or Sounding systems.
3. Pipes that are required to have a slope must be kept high as long as possible.
4. Think about the possibility of supporting the pipelines. Not too far away
from deck, bulkhead or tank top.
5. Make a clear choice whether a pipe must be welded in or must be removable.
6. For several systems we use specific construction details. These details will
be available on the corresponding pipe diagram.
7. Keep in mind that pipes should comply with the rules from classification
society. (See subsection 2.5.1).

2.6.2 Dividing a Pipe to Pipe Spools

1. In technical spaces or on a unit every pipe must be removable. Therefore
flanges must be used. Pay attention to use the right type of flange. (Welding
neck, PNIO, SAE-flange, 0-ring etc.)
2. In other parts of the ship pipes may be welded in. In electrical spaces this
is a must. Never use flanges above electrical equipment. (See subsection
2.5.1)
3. Make a clear choice whether a pipe must be welded in or must be removable.
4. Maximum length off pipe spools depends on the situation. A pipe that
will be put in during pre-outfit can often be longer than pipes put in during
outfit. In case of any doubt a shorter pipe will be the best option. Maximum
length will be 6 meter, (metric pipe 5 meter)
5. After dividing a pipeline the parts must be easy to handle. No square spools
with legs more than 2 x 2 meter.
2.7 SOME MISTAKES IN PRACTICE 29

6. Flanges in a sloped part of a pipe are difficult to assemble in the pipe shop.
This will affect the accuracy of the pipe spool. Always try to place a flange
set in a straight part of the pipe.
7. Valves, fittings, gaskets, etc. must be replaceable. Therefore it is necessary
to create an easily loosening pipe spool which is directly connected to it.

2.6.3 Bended Pipes

1. In bended pipes we will divide the pipe on rounded dimensions. Preferable,
include sufficient straight for the clamp length so flanges (or sleeves) can be
welded prior to bending the pipe in the bending machine.
2. One pipe spool with more than one bend needs to be placed on the bending
machine only once. Remember this while dividing a pipeline. So it is better
to have one pipe with two bends instead of two pipes with one bend each.
3. Route pipes in such a way that no welding is necessary between two bends.
So the minimum length must be including sufficient straight for the clamp
length of the bending machine.

2.6.4 Welding Elbow

1. Use of 1.5 diameter (LR) welding elbows must be restricted. Bending is
always preferred. Think about internal grinding after welding elbows in
pipes that must be coated afterward. The bending radius of a modern
bending machine is only a fraction larger than the radius of a 1.5 diameter
welding elbow.
2. Use of 1 diameter (SR) welding elbows must be avoided. These are far more
expensive than 1.5 diameter elbows. Often specifications require a minimum
radius of 1.5 diameter.
3. Place flanges preferably directly to a 1.5 diameter (LR) welding elbow.

2.7 Some Mistakes in Practice

In the previous section, the common knowledge of pipe routing has been described.
However, since up to now official guidance does often not exist, not every pipe
engineer is aware of it. For pipe engineers that have a lot of experience, that
common knowledge is something that they have learned by routing pipes in pre-
vious projects and by getting feedback from production and outfitting employees.
On the other hand, inexperienced pipe engineers most likely will not consider the
common knowledge sufficiently.
Fig. 2.2, 2.3, and 2.4 show basic mistakes in pipe routing; routing a pipe
without considering that the pipe must be close enough to a steel construction to
be well supported.
30 PIPE ROUTING IN PRACTICE 2.7

F i g u r e 2.2: Pipe needs to be moved near the steel construction

F i g u r e 2.3: Pipe should be moved to the right

27 SOME MISTAKES IN PRACTICE 31

F i g u r e 2.4: Pipe is not close enough to be supported

F i g u r e 2.5: Combination of pipes

32 PIPE ROUTING IN PRACTICE 2.8

Figure 2.6: Valve misplacement

In Fig. 2.5, we can see that pipes 1 to 4 are not, routed efficiently as a group
of pipes. Wfiat fiappened in this example is that the pipe engineer routes pipe 1
and 2 without considering that pipes 3 and 4 must be connected as well.
When a valve needs to be placed, a pipe engineer needs to consider that the
valve is properly accessible to be used. Fig. 2.6 shows an example where an
inexperienced pipe engineer did not think about that,

2.8 Summary
This chapter described the practical aspects of current manual pipe routing pro-
cesses in answering the second, third, and a part of the fifth research questions of
this thesis. In this section we summarize it based on the addressed questions.
The second research question relates to the needed information, the responsible
person and how to get that. To answer this question, we started this chapter by
explaining the stalceholder and their tasks and responsibilities. Then in Section
2.3, the list of tools and information that are needed are described. However, this
list relates to the manual routing process, so we will revisit this question again in
the next few chapters.
The third research question asks for the common knowledge of the pipe rout-
ing process. Section 2.4 describes in detail the process; starting with the organ-
izational part, we continued to explain the routing process itself, and how pipe
engineers collaborate with another. Then, Section 2.5 and 2.6 described in more
detail the criteria and the common knowledge of pipe routing process.
The fifth research question relates to measuring the quality of the routed pipes.
 SUMMARY 33

In Section 2.5 the pipe routing criteria were explained, and even tliough they are
not translated into quantitative measures yet, this section has answered part of
this question.
Chapter

Pipe Routing Framework

3.1 Introduction
As discussed in Chapter 1, the main goal of this thesis is to research a new
methodology to improve the current pipe routing process, and to validate the
tools based on the proposed methodology. Next to the main requirement to
find a "good" solution, there are four functional requirements of the proposed
methodology that are explained in Section 1.3.4: to minimize user input and
user decision; to be integrated easily with other systems, such as existing CAD
software; non proprietary and expandable; and able to produce results within an
acceptable time frame.
In Chapter 2, we have discussed the practice of the pipe routing process in
a ship. It highlighted the important aspects that must be followed by the pipe
engineer to route pipes in a ship according to the marine classification regulations.
Pipe engineers also need to consider to lower the cost of pipes, including material,
production and installation cost. In Section 2.6 the most important common
knowledge of pipe routing in a ship wjis described.
In this chapter the outline ol the lunctional framework is decided and the
elements that need to be lurther investigated will be highlighted. The second
part of this chapter will present a brief review of literature with a focus on the
highlighted functionality of the proposed methodology.
This chapter will describe at some length the practicalities of the functional
Iramework of our methodology. While this may seem uninteresting from the
point of view or research, we point out what our research aims at developing an
integrated pipe routing methodology, 'lb validate it, it is necessary to dwell also
on the context of the pipe routing process.
The basic outline ol the functionality framework is sliowii in Fig. 3.1. The
detail of the proposed methodology, including the architecture and the imple-

35
36 PIPE ROUTING FRAMEWORK 3.2

( Data Retrieval
)
y
f Perform Pipe Routing
)
Y
{ Send Result to CAD Software
)
Figure 3.1: Tke Outhne of the Functional i<Yamework

mentation will be explained in Chapter 5.

As one of the goals, the proposed methodology must be able to be used in
the real design process. It means that the methodology must be able to be used
together with any existing 3D CAD software package. As a proof of concept,
in our case it must be compatible with both Nupas-Cadmatic and IVibon M3
software packages.
There are three main steps in the functional framework; it begins with the
process to retrieve the data that is needed for the routing process. Then the
routing process is performed according to the common knowledge that is described
in Subsection 2.6. After this is finished, the result must be exported to the CAD
software for further processing.

3.2 Required Data

In order to define the part of data retrieval, we need to investigate three things;
what data we need, where those data reside, and what the type of those data is.
In Section 2.3 we have discussed the tools and the information that are currently
needed by a pipe engineer to route pipes in a ship. In our methodology, almost
the same data are needed by the pipe router module to route pipes in ship.

3.2.1 Piping and Instrument Drawing

In the same manner as for the manual routing process, the pipe router module
of our methodology needs to have access to the P&I diagram as the primary
guidance to route pipes. A P&I diagram (see an example in Fig. 3.2) shows to
which equipment nozzles, pipes should be connected. This can be conceptualized
by the analogy of a road trip. To travel from location A to location B, we need to
use a road map to find the direction. In a P&I diagram, both location A and B
3.2 REQUIRED DATA 37

.a. QiïLJ £L

f S!Si I lüMmMKin.

Figure 3.2: Process and Instrument Diagram

are shown as a certain nozzle of equipment A and B, and the pipe acts as a road
in a road map between A and B.
Beside that, a P&I diagram also contains some specification of pipes, such as
the pipe diameter and pipe specification. It also includes measuring instruments,
valves, and other pipe components, such as a pipe reducer.
In a P&I diagram, every pipe and other component has a unique name. In
practice, together with the P&I diagram, there is a document that contains the
detailed specification of every pipe, valve and instrument.
Very often, a P&I diagram is merely a plain drawing that does not have any
relation with a database in a CAD software package. This is problematic for
process efficiency. For example, as it does not have any connection with the 3D
model database, a pipe engineer needs to manually look for the unique name
in the P&I diagram and find the component with the same name in the CAD
software.
Starting a few years back, the so called Smart P&I diagram concept is in-
troduced. A Smart P&I diagram is a P&I diagram that has a live connection
between each object in it with the same object in the databases of the CAD soft-
ware package. For example, every pipe in the P&I diagram has a link with the
pipe specification database. With this feature, the consistency between the P&I
diagram and the routed pipes can be easily maintained.
In manual routing, a pipe engineer usually uses a hard copy of the P&I dia-
gram by making a print of it. In this case, the normal plain drawing of a P&I
diagram can be used, because the pipe engineer can mentally translate all lines
and symbol in the P&I diagram. However, the pipe router module in our proposed
methodology needs to have the Smart P&I diagram.
Looking at the development trend of the major CAD software companies, in
the future this requirement will become standard. Only a few CAD software
38 PIPE ROUTING FRAMEWORK 3.2

Figure 3.3: 3D General Arrangement

companies have already implemented the Smart P&I diagram concept. In our
laboratory case, both Nupas-Cadmatic and IVibon M3 only implement the link
between pipes and valves in the P&I diagram with the pipe specification part in
the CAD data. On the other hand, both software packages are able to generate
the old fashioned P&I diagram. Thus, to bridge the gap, a simple Smart P&I
diagram tool is included in the proposed methodology.

3.2.2 General Arrangement

Using the same road map analogy, a P&I diagram shows the road direction and the
start and end location. However, it only shows the relative topological direction
and location. The absolute start and end locations are represented in a general
arrangement.
In an early design process, a general arrangement is often merely a 2-dimensional
drawing that shows the location of every major component in a ship. Normally,
there is one general arrangement drawing for each deck. Later on, the 3D model
of those components are placed in a 3D space based on those general arrangement
drawings, resulting in what we call a 3D general arrangement.
Both Nupas-Cadmatic and TYibon M3 have a 3D modeling package, but they
use a different kind of implementation to import and export data. Thus, to
fulfill the non proprietary requirement, the proposed methodology must adopt a
common way that can serve both software packages, and indeed also more general
3.2 REQUIRED DATA 39

packages.

3.2.3 Component Details

In a 3D general arrangement, the exact location of every component is known.
However, using the 3D general arrangement alone is not enough to know the
exact location of each nozzle that needs to be connected. For this purpose, the
completeness of the 3D models in a 3D general arrangement is needed. There
are at least 3 requirements of completeness of a 3D model; it has the correct 3D
volume**, it has the exact location and type of every nozzle, and for a certain type
of component, it includes the working space area. Fortunately for our laboratory
circumstances, in both Nupas-Cadmatic and Tribon M3 this kind of detail is
available.
As we discussed before, a Smart P&I diagram has a live connection between
it and another database, thus there is a connection between every component in
a Smart P&I diagram and the 3D general arrangement. Using both the Smart
P&I diagram and the 3D general arrangement, the pipe router module in the
proposed methodology can start the routing process in a free space. However,
since our objective is to route pipes in a ship, the steel construction of the ship
needs to be known. It provides both constraints to a feasible solution and supports
for suspending the pipes.

3.2.4 3D Steel Construction

The first thing that is needed to know about a steel structural element of a ship
is the type of construction, whether it is a normal plate, watertight bulkhead,
bracket, pillar or stiffener. This is important because it has a direct effect on how
pipes should be routed. For example, if there is no other way, pipes can be routed
through a plate (if it is not aifecting the strength and stiffness of the structure),
but not through a stiffener or a pillar. However, if it is needed, a pipe engineer
can make a request to the naval architect to make a modification in the steel
structure, even though this is not recommended.
Just as other kinds of data mentioned above, the 3D steel construction is
available in both Nupas-Cadmatic and Tribon M3.

3.2.5 Tank Plan Drawing

This plan is one of many "practical plans" based on the general arrangement.
Other examples are the Watertight compartment and Watertight door plan, the
maintenance routes plan, the overview plan of the fire fighting and detection
system. They show certain aspects of the design that are related to the overall

•'SD volume is the term that widely used in the 3D CAD application to represent the physical
dimension of a 3D model. In this thesis, we adopt this term and use it in all chapters.
40 PIPE ROUTING FRAMEWORK 3.3

''H^r.

Figure 3.4: 3D Steel Construction

functioning of the ship and therefore, lilse the P&I diagram, are of prime interest
to the user.
Some of the pipe routing rules are related to a certain type of tank in a ship.
For example, it is not allowed to route an oil pipe through a freshwater tank.
To accommodate this kind of rules, the tank type and location must be known.
Unfortunately, in practice the 3D steel construction model does not contain that
information currently. The complete information is available only as a two di-
mensional drawing called a tank plan drawing.
In the current situation, in most cEises the tank plan drawing is created during
the basic design process using a two dimensional software package.

3.2.6 NoGo Area

In Subsection 2.4.4, the four types of an object constraint were explained. It was
also mentioned that an object constraint can be a real object or an area. A NoGo
area is an area that is defined as an absolute constraint or as a soft constraint or
as a combination of both.
For example, in a walking corridor inside the machinery room, a pipe should
not be routed in the middle of that corridor but it can be routed along the edge
of it. In the proposed methodology, the possibility to create that kind of NoGo
Area must be accommodated.
3.3 PIPE ROUTING 41

Figure 3.5: General Arrangement with NoGo Area

3.3 Pipe Routing

In the previous section, the data needed by the pipe router module in the proposed
methodology are discussed. The next step is to conclude the requirements of the
data retrieval part in the proposed methodology.
1. Smart P&I diagram information
2. Common interface to retrieve 3D general arrangement volume data
3. Common interface to retrieve component's detail data
4. Common interface to retrieve 3D steel construction data
5. 3D tank plan information
6. 3D NoGo area information
In case this information is not present in the current modeling tools, simple
additional tools must be provided for our methodology to work properly.
While it is a time consuming task to arrange for all the interfaces and thereby
facilitate the data retrieval, it is not of interest for our research and is not described
here.
In Chapter 2, it was mentioned that there are two main steps in the pipe
routing process. The first step is finding the actual pipe path to connect nozzles,
and after that the routed pipe is divided into one or more pipe spools. In this
thesis, we only focus on the first step and skip the creation of the pipe spool.
Since we know that the pipe spool creation process can be done quite easily in
the CAD software package, this process should be performed there. It is decided
to follow this approach because this research is intended to fill the functional gap
in existing CAD software, instead of replacing them.
42 PIPE ROUTING FRAMEWORK 3.3

The result of the routing process must be as good as possible, and it im-
plements the common knowledge that is described in Section 2.6 to ensure the
quality. However, since we decided to pursue only the pipe path finding and skip
the creation of the pipe spool, the common knowledge that is related to pipe
spools is neglected.
The basic definition of the pipe routing process is to find an optimal path
between two equipment's nozzles, more if a branch exists, while avoiding collisions.
For its implementation in ship design, that pipe routing process must also follow
the common facts for a piping system in a ship. Most of the pipes in a ship are
rigid and normally those pipes are routed orthogonally to make the installation
and maintenance easier. The pipes are divided into several different functional
systems that require a different pipe specification. In a ship, the number of pipes
is very large. Therefore, to lower the cost, it is preferred to have pipes follow
routes that are as short as possible, and also to make sure that those pipes can
be produced and installed as cheaply as possible. It means that we need to use a
shortest path algorithm. The shortest path problem has been subject of research
for years, and we will look in more detail at this in Chapter 4.
However, most of the times, the shortest path alone is not sufficient, because
that path might not comply with the rules and regulations, nor indeed might it
lead to lowest cost. For example, the shortest path algorithm that only optimize
the length of a pipe might generate the pipe path that is too far from the steel
construction, thus it is not possible to install a pipe support for that pipe. This
leads us to define the first requirement to choose the shortest path algorithm;
instead of merely trying to find the shortest distance, the shortest path algorithm
must optimize the path in a weighted manner. The shortest path algorithm must
rather be the lowest cost algorithm.
In some cases, there is also the possibility that a user wants to alter the pipe
path manually. To accommodate this requirement, the methodology must have
a functionality to allow a user to add/remove some conditions pertaining to the
environment. For example, if the best path of a certain pipe is blocked by a plate,
the user should be able to override that collision and let the pipe router module
route the pipe through it.
Most of the shortest path algorithms aim to find one optimum path at a
time, thus it optimizes only the pipe in hand without considering any other pipes
in the group. Therefore the globally optimum set of pipe routes is difRcult to
be reached. In most cases, by choosing a good sequence of pipes to be routed,
the global optimum solution can be approached to a fair measure. Whether the
optimum found is the global optimum cannot be mathematically ascertained. The
problem of choosing the right order of pipe routing is known as the combinatorial
optimization problem.
Therefore to get (or at least approach) the globally optimum solution, we need
to combine the shortest path algorithm and combinatorial optimization. So for
the proposed methodology, we need to find the optimization method that is able to
solve the combinatorial problem and can be used together with the shortest path
3.4 PIPE ROUTING 43

algorithm thai is suitable for our purpose. Ultimately this hybrid optimization
must be constructed.
In the previous section, we remarked that a user of the methodology may add
an object constraint, for example a NoGo area. Since this is a manual process,
there is always a possibility that the additional constraint might block some pipes.
If this is the case, it might add more complexity to the hybrid optimization within
the methodology. 'I'his might cause the hybrid optimization to keep trying to find
a solution that is actually blocked by the additional NoGo area. To prevent this,
the proposed methodology must be able to detect it before the hybrid optimization
process is started. For this purpose, we need to utilize a fast path finding algorithm
to be included in the methodology to act as the fast detector if a certain NoGo
area blocks one or more pipes.
Section 2.5.1 and 2.6 discussed the common knowledge and strategy to route
pipes in a ship. I^Vom the pipe router point of view, those rules must be translated
into requirements for the pipe router algorithm. One of the design requirements
stated in Section 1.3.4 mentions that a goal of the proposed methodology is to
minimize user input and user decision. To fulfill this requirement, the predefined
rules must be built as a knowledge base. It means that it should be easy to add
and modify rules by the pipe routing experts.
Section 2.4.3 discussed that the way in which pipes are routed depends on the
location of those pipes in a ship. In the manual routing process, by experience,
the pipe engineer varies his or her routing behavior based on that. Consequently,
the routing behavior of the pipe router module in the proposed methodology also
should depend on the area in which the pipes are placed. Since each type of area
has unique characteristics, the methodology has to include a different strategy of
the routing process for a different type of area.
There is one other important matter that needs to be addressed. A pipe
routing process is all about optimizing the path of pipes. The first thing that
should be known before any optimization is performed is to define the objective
of the optimization. Also, the validation criteria must be decided.
As we intend to use the 3D model built by means of the CAD software, we have
to prepared to get the 3D model with a very high level of detail (as mentioned
in Subsection 2.3.2). The number of objects in a ship can be more than one
hundred thousand. The high level of detail in the 3D model may thus reduce the
performance of the methodology with regard to the computation time to find the
routing path and also for the performance of the user interface part. Obviously, we
need to find a way to simplify the model while maintaining accuracy and routing
validity. A brief introduction of model simplification will be given in Chapter 4.
The last part of the functional framework is to send back the result to the
CAD software package, so the pipe engineer can continue to the second step
of the routing process, splitting pipes into one or more pipe spools. This part
of the methodology has been implemented, but since it is purely a computer
programming problem, it will not be discussed in this thesis.
44 PIPE ROUTING FRAMEWORK 3.4

3.4 Literature Review of Automatic Pipe Routing

In this secüon, we survey the state of the art in automatic pipe routing. Auto-
matic pipe routing has been a research topic for a long time resulting in various
approaches, not only in the shipbuilding area, but also in process plant design.
The research started with a 2-D workspace and simple obstacles, and gradually
extended to the stage of a 3D workspace with multiple constraints and multiple
objectives. In terms of the optimization technique, deterministic, heuristic, or a
combination of both methods have been used to improve the results.
We filter out the studies that have small relevance for the pipe routing im-
plementation in shipbuilding. This is important since the details of the routing
process highly depend on the area of implementation.
However, it does not mean that the research that concentrates on other fields
than shipbuilding can immediately be ignored. The similarity of the environment
and the type of the paths (in our case pipes) are the most important things. As
already explained in the previous section, in a ship the majority of the pipes are
rigid and orthogonal. Thus, we still include in our survey the pipe routing studies
which are implemented in industrial plants, even though in a ship the pipes are
not routed in dedicated pipe racks which is common practice in an industrial
plant.
Since the whole process concerns many fields of research, we focus on two main
problem domains;
1. The path finding straLegy. The in-depth review of the shortest path al-
gorithms will be done in Chapter 4. In this section, we only survey the
methods that were used in the reviewed literature, and consider if these
also take into account practical aspects, like branches.
2. The objective of the optimization. Since our main goal is to use the proposed
methodology in a real ship design process, the optimization objectives used
in previous studies is an interesting aspect to be reviewed.
Unfortunately, to date most of the algorithms are demonstrated to solve the
pipe routing problem only in academic situations, such as a system with only a
few pipes and in a simplified environment. They have paid little or no attention
to the scalability of the algorithm to be applied in the real ship design process
which consists of a much larger number of pipes. Some of the researches also
neglect the existence of pipe branches. Moreover, the algorithms are also mainly
used for the space problem - to find a pipe roxite without collision - without
considering the vitally important aspect that the solution must comply with the
marine classification rules and regulations.

3.4.1 Early Years

The automatic pipe routing research started in the 1970's when Newell [1972]
presented his work to route pipes automatically in chemical plants. He adopted
Nicholson [1966] method to find the shortest path, and in his implementation,
3.4 LITERATURE REVIEW OF AUTOMATIC PIPE ROUTING 45

Newell partly considered pipe branching. In 1974 by using the algorithm from
Dijkstra [1959], Wangdahl, Pollock, and Woodward [1974] attempted to solve
the pipe routing problem in a ship, but they only considered a two dimensional
environment. The main drawback of their research is that since the pipes are
routed one at a time, the globally optimum set of pipe routes might not be found
since they do not include a mechanism to properly order the pipes. What makes
it worse is that in some cases, by solving the shortest path problem one pipe at
a time, a situation might result where a pipe can not be routed because it was
blocked by the previously routed pipes. In his attempt to solve that problem,
Rourke [1975] reviewed several algorithms but failed to find the solution.

3.4.2 Zhu and Latombe

In the year 1991, Zhu and Latombe [1991] described a system for automatically
performing the pipe routing using robot path planning techniques. They regard
each pipe as the trace left behind by a rigid object (a robot) moving in the pipe
workspace, and a pipe routing problem as a multi-robot path planning problem.
In their work, the approximate cell decomposition approach was chosen. Then to
find the shortest path, the A* algorithm from Hart, Nilsson, and Raphael [1968]
is utilized.
They also attempt to solve the condition where some pipes are blocking each
other, using a strategy called backtracking. If a pipe was blocked by another pipe,
the backtracking strategy will be triggered. I'hat is, it must change the routes of
some of the previously routed pipes to make room for the current pipe. Zhu and
Latombe developed a sophisticated backtracking strategy that only considers the
pipes that are actually blocking the current pipe.
The most interesting part in the work of Zhu and Latombe is that they also
consider some practical rules by introducing the expert design constraints, which
consists of process constraints, structural constraints and accessibility constraints.
Process constraints relate to the process that is carried out in the pipes, e.g. a
high temperature pipe should have an expansion loop to ensure thermal flexibility,
a drainage pipe should be non-ascending and a heat sensitive pipe should be kept
sufiiciently far away from high temperature equipment. Structural constraints
relate to the mechanical properties of the pipes, more precisely their capacity to
remain in their position without falling down, e.g. pipes should be near enough
to a major support structure such as a wall or a beam. Accessibility constraints
relate to the constructibility of the pipe layout and its ease of operation and
maintenance, e.g. there must be an access path for removing all major equipments
for off-site repair and the frequently used valves should be accessible. Those three
types of constraints then were formulated into two kinds of geometric constraints;
Location constraints, and Shape constraints.
The location constraints specify the forbidden, undesirable, or preferable re-
gions for a pipe route to go through. They are conceptualized as hard virtual
obstacles, soft virtual obstacle, and virtual sinks. During the routing process, a
46 PIPE ROUTING FRAMEWORK 3.4

hard virtual obstacle acts as a real obstacle, while a soft virtual obstacle can be
traversed by pipes with some additional cost. A virtual sink acts in the same
manner as a soft virtual obstacle, but instead of getting a penalty, a bonus may
apply to the pipes that are routed through this region. The location constraints
are used during the cell generation step. The shape constraints apply to the shape
of the pipe routes. For example it is to ensure that a drainage pipe should be
non-ascending. The shape constraints are applied during the cell generation level
and at the path generation level.
Their approach, however, was focused on the study case environment with a
relatively small number of pipes and obstacles. The type of the obstacles are also
only a basic shape, which might not be sufficient for the real environment. One
other important thing that was left out is the aspect of pipe branching.
Unfortunately, they did not continue their research in the subject of pipe
routing but were more interested in the subject of robotics.

3.4.3 Kang's Expert System

In 1996, Kang, Myung, and llan proposed a method for generating the optimal
route for pipes using a knowledge-based expert system called NEXPERT, Kang
et al. [1996] and Kang et al. [1999]. The knowledge-base is constructed on the
basis of documented design knowledge and the empirical knowledge of human
experts on the piping design of a ship. The system is modeled with the following
objectives: to minimize user input and user decision; to structure the knowledge-
base for easy addition of knowledge; to make the system easy to xise; and to be
used in the real shipyard design process.
In the expert system, there are three different objects; pipe-path, pipe-element,
and space-element. Space-elements represent spaces and obstacles where pipe ele-
ments should or should not be placed. Constraints are implemented as rules, and
algorithms are implemented as sub-routines. The knowledge-base in the system
consists of three parts; the meta-control knowledge that makes main decisions,
the global designer that finds the optimal arrangement of main pipes in a two di-
mensional section plan, and the detail designer which will expand the 2D section
plan along the transverse coordinate into 3D space along the ship length.
In their research, Kang et al. constructed three different knowledge-bases that
store 167 rules and 106 supporting methods. To verify their method, the piping
design expert system had been tested to route pipes in the deck of a bulk carrier.
Using the proposed system they can reduce the working time from 4 hours to 1
hour.
However, even though the way the knowledge-base was constructed makes
it easy to be expanded with a new rule, their method is practically hard to be
used since it is difficult to define all design knowledge quantitatively Also the
knowledge base is difficult to maintain in case some predefined rules are changed.
3.4 LITERATURE REVIEW OF AUTOMATIC PIPE ROUTING 47

3.4.4 Sandurkar and Ito

In 1998, Sandurkar and Chen [1998] utilized a heuristic optimization approach
based on the genetic algorithm of Goldberg [1989] to automatically perform the
pipe routing. The interesting part of their research is that they are one of the first
to use the real tessellated model as an obstacle rather than using only the bound-
ary box of the model. Tessellation is the process of creating a two-dimensional
plane using the repetition of a geometric shape with no overlaps and no gaps.
Tessellation techniques are often used to manage data sets of polygons and di-
vide them into suitable structures for rendering. Normally, at least for real-time
rendering, the data is tessellated into triangles, which is sometimes referred to as
triangulation.
By using the real tessellated model, the working environment is a closer repre-
sentation of to the real environment when compared to only using the boundary of
the real object. However, since the number of triangles in one tessellated complex
object is large, the computation time to perform the collision detection is also
larger. To tackle this problem, Sandurkar and Chen utilized the RAPID method
from Gottschalk et al. [1996].
In their research, besides optimizing the pipe length, they also define the
desired number of bends and the angle of the bends. This approach proved to be
able to find a solution. However, it was applied only to one model with less than
10 obstacles and a single pipe and still took 18-19 hours of computations before
the best layout of the pipes was found.
Using a different approach. Ito [1999] also utilized a genetic algorithm approach
to find the best path. The workspace is defined by using the cell decomposition
approach. Then, each cell is given a potential value, according to its location and
characteristics. The potential value of the obstacle cells are high and the cells
located next to the wall have the lowest potential value because that path is more
favorable. The objective function is defined not only considering the pipe length,
but to minimize the total cell value of the selected path.
This method was only tested on a two dimensional space and using primitive
shapes for the obstacles. In practice, as described in previous chapters, the number
of pipes that need to be routed is large. Also the complexity of the problem rises
exponentially with the number of pipes to be routed. Therefore, even allowing
for the large performance gains of computers, the method still seems unsuited for
practical application.

3.4.5 Zuurmond
In 2004, Zuurmond [2004] proposed his approach to solve the automatic pipe
routing problem, by utilizing the algorithm from Dijkstra [f959] to minimize the
pipe length. Beside that, his method also restricts the drainage pipes to be non-
ascending. The workspace is defined by using the cell decomposition approach.
For the obstacles Zuurmond simplified the real model into some cuboids. This
48 PIPE ROUTING FRAMEWORK 3.4

approach has an important advantage; to have a much lower computation time

compared to the real tessellated model while still allowing a good representation
of the real model.
The most important contribution of his work is not in the routing algorithm
itself, but more in his description of the routing process in practice at a shipyard.
Also, the importance to have the Smart P&I diagram tools was mentioned.
Since the aim of his research is global pipe routing. Zuurmond used a large cell
size and always defined the pipe path in the center of axis of a cell. Therefore, even
though the density of each cell was calculated to ensure that the total number of
pipes that are routed in a certain cell always fit, the result still shows that there
are collisions between pipes.
Another drawback is that before pipes can be routed automatically, a lot of
manual setup must be done. He mentioned the importance of using the Smart P&I
diagram, but still manually defined each connection point. Also, the simplification
of the real model into cuboids is done manually.

3.4.6 Park and Storch

One of the most comprehensive research attempts in terms of considering more

practical aspects is the research of Park and Storch [2002]. They considered many
practical aspects, like branches, and practical constraints which will be translated
into total pipe cost functions. Thus, their objective value is to minimize the total
cost of the pipes.
The total cost of the pipes includes the material cost, installation cost and
operability cost. The material cost of pipes and elbows depends on pipe size and
length. T'he bending cost follows a step function because cold bending is used for
small diameter pipes and high frequency bending is used for large pipes. As for
installation cost, it is directly related to pipe-support cost. They also consider the
distance between pipes and major structural object for pipe-support installation.
The location of the valves are considered as part ot the operability cost. They
follow the guidance from ABB [1998] to measure the degree of comfort to operate
a valve, then translate it into cost.
In terms of the routing method, they proposed the cell generation method.
Using their method, the globally optimum set of pipe routes problem that was
mentioned by Rourke [1975] can be solved.
However, even though they used a part of an engine room in a real ship, the
research was intended as an early study. Until now it was not continued for more
elaborate cases. Therefore, it was tested only for a small portion of the engine
room. Also for reasons ol simplicity, the boundary boxes of the real obstacles are
used.
3.4 LITERATURE REVIEW OF AUTOMATIC PIPE ROUTING 49

Figure 3.6: Level 1 mode (excerpt from AVE [2007])

Figure 3.7: fjevel 2 and Level 3 mode (excerpt from AVE [2007])

3.4.7 Commercial Automatic Pipe Router

In the past, few years, commercial pipe rouler software liEis been developed and
deployed by ASD, Alias, and AVEVA. Judging from their commercial leaflets, they
have some basic functional similarity. However, since the software is commercial,
they did not make any technical publication on the algorithm that they use. The
most detailed documentation available is the user guide of the Pipework Design
User Guide from AVEVA AVE [2007].
The automatic router from AVEVA which is known as the PDMS Router
routes the pipes orthogonally, and uses the graph theory principles from Gibbons
[1985], and a shortest-path algorithm from Wang and Crowcroft [1992] to get the
minimum pipe length. The objective of the PDMS Router is to minimize material
cost while avoiding collision with the obstacles. To enhance its performance, the
routing algorithm used by PDMS Router has three levels of operational modes;
called Level 1, Level 2, and Level 3. Initially, Level 1 mode was used to find the
path, and if it fails, a search is conducted using Level 2. Similarly, if no free route
is found after Level 2, then Level 3 is used.
Fig. 3.6 shows how the Level 1 mode searches for an orthogonal route between
the head point PII and tail point P T of a pipe, using the minimum number of
bends and elbows. Fig. 3.7 shows the example of Level 2 and Level 3 mode
of the PDMS Router. In the left figure we can see that the bending points
can be dynamically moved along the boundary box of the path. The Level 3 is
used if both Level 1 and 2 failed. In Level 3, the boundary box of the path is
automatically extended to allow the pipe path to connect start and end points.
There are some other settings that can be defined by the user, such as pre-
defined pipe racks and routing rules. It is also possible to create a point in a space
50 PIPE ROUTING FRAMEWORK 3.5

as a routing point, and a certain pipe will be routed automatically through that
point.
It is important to keep in mind that PDMS Router does not provide any other
automation beyond the de facto routing of pipes. This means that the designer
has to perform all the manual tasks needed for PDMS Router to work. In order
to reduce the human involvement in the tasks necessary for the PDMS Router,
a research using PDMS Router has been done by Calixto, Bordoira, Calazans,
Tavares, and Rodriguez [2009].
Another attempt to use the commercial software was done by 11 Roh, Lee, and
Choi. Rather than using a generic search algorithm, 11 Roh et al. [2007] chose
to improve the function that is available in the commercial 3D CAD software
TRJBON and IntelliShip systems. There are 2 main parts in their research. The
first one is the improvement of the pipe routing function in TRIBON and Intel-
liShip systems. In principle, to route pipes they need to define the pipe tray and
then those pipes can be routed automatically. The second part of their research
is to rapidly modify the pipes that are already routed when the hull structure is
changed.
Both the original 'lYibon M3 automatic pipe routing and the two researches
above that have been done to improve its standard functionality are not focusing
on performing fully automatic pipe routing as we intended.

3.5 Summary
This chapter starts with the explanation of the basic outline of the functional
framework of our proposed methodology. The type of information that is needed
is described and the source and how to get it into the methodology is also briefly
discussed.
The second part of this chapter reviews the previous researches that have been
done in the field of automatic pipe routing. We found many interesting approaches
but none of them led to an approach that satisfies our targets, i.e. fully automatic
routing applicable for detailed design phase and for complex, real ship situations
in a practically applicable methodology.
While reviewing, we found the answer to the first research question. T'hat
question asks for the phase that we should concentrate our effort on; in the pre-
contractual or detail design phase, and the reason why we choose to that. In his
work. Zuurmond [2004] routes the pipes in a machinery room of a real ship. What
he did is the approximate routing so the result can be found almost immediately.
This approach is not suitable to be used for detail design phase. However, from
Subsection 1.3.f we knew that during the pre-contractual phase, approximate
routing is sufficient. Therefore, we can say that to solve the approximate routing
in the pre-contractual phase, we can simply adopt that method. For the sake
of widespread applicability, we focus on solving the pipe routing problem in the
detail design phase.
Chapter

Related Work

In the first part of Chapter 3 we have discussed five main points that need to be in-
vestigated; the shortest path problem, combinatorial optimization, the knowledge
base, the differences of behavior according to area type, and objective function.
In this chapter the first two points will be discussed along with the problem of
model simplification. The third until the fifth points will be discussed in Chapter
5.
We begin with surveying previous studies of each subject, then comparing
those studies to find the most suitable solution and investigate how to improve
the existing solution to fulfill our functional framework requirements.

4.1 Shortest Path Problem

In graph theory, the shortest path problem is the problem of finding a path
between two vertices (nodes) in a graph such that the sum of the weights of its
constituent edges is minimized. An example is finding the shortest way to get
from one city to another on a road map, shown in Fig. 4.1; in this case, the
vertices represent cities and the edges represent the segments of the road and are
weighted by the travel distance.
For example, if we need to answer the question "What is the shortest travel
distance to drive from Hardinxveld to Amsterdam?", we might use the graph in
Fig. 4.1 and use one of the shortest path algorithms to find the optimal solution.
Many researches have been done to solve the shortest path problem and it is
becoming one of the most prominent generic problems in various fields. One of
the reasons for this is that essentially any combinatorial optimization problem
can be formulated as a shortest path problem. Thus, this class of problems is
extremely large and includes numerous practical problems that have nothing to
do with actual shortest path problems.

51
52 RELATED WORK 4.1

80km

31 km

Figure 4.1: Example of a road map

Many approaches have been made ranging from using the deterministic (exact)
algorithm, subsequently introducing the heuristic part to improve the algorithm,
and recently researches exclusively using heuristic algorithms to solve it.
In this section, we walk through those algorithms and choose the algorithm
that suits our requirements. There are two criteria to be considered. The first
one is the performance of the algorithm itself. There are some important issues
that need to be addressed to measure the algorithm performance to solve the
shortest path problem; always find a solution if it exists, always find the optimal
solution, and use limited resources in terms of computer memory and time. In
our methodology, pipes can be routed freely in the area, therefore the selected
algorithm must be able to be implemented to route pipes in a free space.
We also consider the flexibility of those algorithms to be extended and adapted
in our methodology. In chapter 2, we have discussed some practical knowledge
that help us define the basic capabilities that must be utilized in our methodology.
We have discussed that there are some practical aspects that need to be addressed
beyond the shortest length and minimum number of bends. To have pipes routed
nicely in parallel, to route pipes close enough to the steel construction for ease of
installing the pipe support, and to follow the marine classification guidance, are
more important than only trying to route pipes such that they have minimum
length.
The nature ol a shortest path algorithm is to find the path between start and
end points that has a shortest distance. To make the shortest path algorithm
choose the path that satisfies the important practical aspects above, we need to
modify the environment to fit the algorithm, for example by using a different
weighted cost, or using potential energy techniques. Because of that, in addition
to four requirements above, the selected algorithm must also allow implementation
to find the path in that type of environment.
As a part of our methodology, we also need to have a fast function to detect
that a certain pipe can be routed at all. For this functionality the only important
requirement to choose the shortest path algorithm is that the selected algorithm
4.2 DETERMINISTIC APPROACHES 53

always finds a solution if il, exists. This will be discussed in chapter 5.

In the next few sections, some approaches to solve the shortest path problem
are discussed. We start with the deterministic approaches, followed by the heur-
istic methods to solve this problem. After that, a comparison between various
methods to solve s simple shortest path problem will carried out.

4.2 Deterministic Approaches

4.2.1 Graph Traversal Algorithms
In graph theory, one of the most fundamental tasks in an algorithm is visiting the
vertices and edges of a graph in a systematic order. There are at least three differ-
ent traversal techniques that are frequently used; Depth-first search, Breadth-first
search, and Best-first search.

Breadth-first Search
'I'he breadth-first search (BFS) algorithm is an uninformed search that systemat-
ically visits all the vertices of a graph until a goal vertex is found or all vertices are
visited. This algorithm starts at a designated start vertex and then examines the
neighbours of that vertex and puts it in the queue stack. Once a vertex has been
examined, it is marked as explored. After all neighbours have been examined,
it visits all of them one by one and examines their neighbours. This process is
repeated until all vertices of the entire graph have been visited, or until the goal
vertex has been reached. In other words, it visits the vertices of a graph uniformly
across the breadth of the frontier of its search, visiting all vertices at distance (d)
from the start vertex before looking for vertices at distance {d+1). For the order
of the search, BFS algorithm uses a First-in First-Out (FIFO) queue stack.
We use the graph of Fig. 4.2 to illustrate the Breadth-first search algorithm.
In this example, the start vertex is Hardinxveld, and we would like to find a path
to Amsterdam. Before we use BFS algorithm, we define an order of visiting the
neighbours. Fig. 4.2.a and b iUustrate BFS algorithm with visiting order from
the right to the left side and from the left to the right side respectively.
In Fig. 4.2.a, BFS starts from Hardinxveld and examines its neighbours,
Gorinchem and Rotterdam, marks it as explored and stacks it in a queue. After
that BFS visits Gorinchem and examines its neighbors Breda, Utrecht, and Hardinxveld.
Since Hardinxveld was already marked, it won't be explored further. Breda and
Utrecht are then marked and added to the queue stack. The next vertex in the
queue stack is Rotterdam. From all three neighbours of Rotterdam, only Ams-
terdam is unmarked. Therefore, Amsterdam is marked and added to the queue
stack. Eventually, our goal vertex has been reached, but in this example, we
continue to run the algorithm to build the complete tree.
Breda is now in the top of the queue stack, but all of its neighbours are marked.
From Utrecht, the search is continued and results in Amersfoort to be marked and
54 RELATED WORK 4.2

Figure 4.2: Breadih-first search on the road map tree

added to Ihe queue. Since all edges from Amsterdam lead to the marked vertices,
no vertex is added to the queue stack. Then it visits Amersfoort and then marks
Zwolle. The resulting Breadth-first search tree is shown in Fig. 4.2.C. Fig. 4.2.d
shows the tree in the left to right visiting order.
By comparing the BFS tree in Fig. 4.2.C and Fig. 4.2.d, it can be seen that
the BFS algorithm is highly sensitive to the visiting order of the vertices. If the
graph is connected, BFS will find a solution because it explores all vertices.
For unit-step cost, BFS is optimal. In general, Breadth-first search is not
optimal since it always returns the result with the fewest segments between the
start vertex and the goal vertex. As in our example above, if the graph is a
weighted graph and has costs associated with each step, a goal next to the start
does not have to be the cheapest goal available. This problem can be solved by
improving Breadth-first search to uniform-cost search, which considers the path
costs. Nevertheless, if the graph is not weighted and all step costs are equal.
Breadth-first search will find the nearest and the best solution.
The BFS algorithm begins with straightforward initialization that requires
0(1) time. In the worst case, the algorithm needs to visit aU vertices before the
goal is reached. This require time OiV) with V the number of vertices. The
algorithm also examines all edges of each vertex, and since there are two vertices
connected by one edge, the total number of examinations is 2 times the number
of edges, 0(E) with E the number of edges. We know that Emo-r < = ^max ~ li
4.2 DETERMINISTIC APPROACHES 55

Figure 4.3: Depth-first search on the road map tree

so 0{V + E) can be simplified to 0{V). Since it uses FIFO stack, it only requires
0(1) time for the dequeue process. In total, the time complexity of BFS algorithm
is 0{V).
Furthermore, BFS needs to memorize the state (marked or unmarked) of all
vertices. It also needs a queue stack memory with the maximum number of the
stack equal to the number of all vertices. So the space complexity is OiV).

Depth-first Search

Another fundamental search algorithm in graph theory is the Depth-first search al-
gorithm. Like the BFS algorithm. Depth-first search (DFS) is also an uninformed
search, so the visiting order of vertices can be arbitrary. The main difference with
BFS algorithm, which explored all the neighbors of a given vertex at a time, is
that the DFS algorithm explores only one neighbor at a time, and then explores
a path in a graph as far as possible until a goal vertex is found, or until it hits
a vertex that has no children. When it is no longer possible to go forward, the
algorithm backtracks one level and then tries again to go deeper. This process
repeats until all vertices of the entire graph have been visited.
To show how the DFS algorithm works, we use the graph of Fig. 4.1 and define
Ilardinxveld as the start vertex and Amsterdam as the goal vertex. However in
this example, we let the DFS algorithm run until all vertices are visited.
As shown in Fig. 4.3.a and Fig. 4.3.b, this algorithm highly depends on the
56 RELATED WORK 4.2

order of visiting. In Fig. 4.3.a, it visits from the most right vertex to the left.
FVom Hardinxveld, it visits Gorinchem, then Breda. Since Breda is the deepest
node in this branch, it backtracks to Gorinchem again, and visits Utrecht and then
Amersfoort and Zwolle. Again, it backtracks to Amersfoort and visits Amsterdam.
Normally, DFS should stop in Amsterdam, but we keep the algorithm running for
completeness sake. The right most side from Amsterdam is Utrecht, but Utrecht
was already visited, so it continues to visit Rotterdam. At this point, all vertices
are visited, but the algorithm does not know it yet. So it keeps backtracking
all the way until it reaches the start vertex. Like BFS, DFS also produces a
Depth-first search tree which in this case is shown in Fig. 4.3.c.
Fig. 4.3.b and Fig. 4.3.d show the results of the DFS algorithm with the order
of visiting from the most left side to the right. Also as before, even though the
goal vertex (Amsterdam) was lound, we continue running the algorithm until all
vertices are visited.
Since all vertices will be visited no matter how complex the graph is, DFS
will find a solution if it exists. DFS is typically used to traverse an entire graph.
However, there is no guarantee that the solution is the optimum one.
This algorithm visits all vertices which requires time 0{V) with V the number
of vertices. In the worst case, during forward and backtracking search each edge
is examined two times which requires 0{E) with E the number of edges. We
know that E is always less than V, such that the total time needed is 0{V).
For the sake of backtracking, the DFS algorithm needs to memorize all vertices
which requires 0{V) memory space.

Generic Best-first Search

Unlike both Depth-first search and Breadth-first search, Best-first search explores
the graph not uniformly following the depth or breadth of the graph, but it
expands the vertex with the best value. Normally, Best-first search is used on
a weighted graph and uses a "heuristic evaluation function" that estimates the
minimum cost from any vertex to the goal.
In the formal terminology, g{n) represents the cost of the path from the start-
ing point to any vertex n, and h{n) represents the heuristic estimated cost from
vertex n to the goal. Best-first search balances the two as it moves from the
starting point to the goal. In each step of the iterative procedure, it examines the
vertex n that has the lowest f(n) = g{nj + h{n).
It is crucial to choose an appropriate heuristic function because it affects the
behaviour and performance of the algorithm. At one extreme, if h{n) is extremely
high relative to g{n) (or simply neglecting g{n)), then only h{n) plays a role. In
this case, this algorithm becomes the Greedy Best-first search algorithm, which is
sometimes called the Greedy algorithm. At the other extreme, if h{n) is 0, then
only g{n) plays a role, and this turns into Dijkstra's algorithm. In between, if
both h{nj and g{n) are included in the algorithm, we get the A* algorithm (see
below).
4.2 DETERMINISTIC APPROACHES 57

Dijkstra Algorithm

One of the most famous algorithms that can be categorized as a Best-first search
technique is the Dijkstra's algorithm by Dijkstra [1959]. As previously described,
Dijkstra's algorithm only considers the cost of the path from the starting point
and neglects the estimated cost to the goal.
For a given start vertex (origin node) in the graph, the algorithm finds the
path with the lowest cost (i.e. the shortest path) between that vertex and every
other vertex. It can also be used for finding the lowest cost or the shortest path
from a single vertex to a single destination vertex. This can be achieved by
stopping the algorithm once the shortest path to the destination vertex has been
determined. For example, if the vertices of the graph represent cities and edge
path costs represent driving distances between pairs of cities connected by a direct
road. Dijkstra's algorithm can be used to find the shortest route between one city
and all other cities.
Let us use the road map in Fig. 4.1 to illustrate how Dijkstra's algorithm
works. Suppose we again want to find the shortest path between Hardinxveld
and Amsterdam, a starting point and a destination. The order is conceptually
straightforward: to start, set the distance to all cities on the map unlabeled. This
is done not to imply that there is no distance, but to note that that intersection has
not yet been examined. Some variants of this method set the distance to infinity
on all cities. In Fig. 4.4.a, Amsterdam as a destination is marked with blue color,
and Hardinxveld is marked with green colour to show that it is currently the best
vertex and will be explored.
Now, at each iteration, select a current city. For the first iteration, the cur-
rent city will be the starting point (Hardinxveld) and the distance to it (the
Hardinxveld's label) will be zero. For subsequent iterations (after the first), the
current city will be the closest unvisited city to the starting point.
Prom the current city, the algorithm updates the distance to every unvisited
city that is directly connected to it. This is done by determining the sum of the
distance between an unvisited city and the value of the current city, and relabeling
the unvisited city with this value if it is less than its current value. In effect, the
city is relabelled if the path to it through the current city is shorter than the
previously known paths.
Dijkstra's algorithm uses Breadth-first search at this stage, because it exam-
ines every neighboring city before it moves on to the next iteration and puts a
label to the examined cities, as shown in Fig. 4.4.b.
Fig. 4.4.C and d shows that Hardinxveld is marked as visited (red colour) and
Gorinchem acts as the current city because it has the lowest label (11), and Fig.
4.4.d shows the updated distance label of Gorinchem's neighbors.
In Fig. 4.4.e and f, Rotterdam acts as the current city since it now has the
lowest value, and the algorithm labels Utrecht and Amsterdam. As we can see
that the new value of Utrecht is bigger than the previous value (89 > 51), so the
link from Rotterdam to Utrecht is not a valid path. Also, Amsterdam has been
58 RELATED WORK
4.2

iif; 93
45km , .
^
^STERDAM^ >< (lÏMERSFOO^

. ' A np \67km
'ROTTERDAM^ •^i«ECin>^^^^

([TH ARD1 NXVELDjy^^ ORINC^^>-^fC!^

1-
93

(msTEHüm^

^ — \ \ -'
'^TBECHT^

TlARDlNXVELD >_j^f^GORlNCHEM "

'llARDINXVELO y^j^^'GOBINCHEM "

k. I.

F i g u r e 4.4: Dijkstra's algorithm

4.2 DETERMINISTIC APPROACHES 59

labelled at this point, but because there are still vertices having lower value than
the current value of Amsterdam, the iteration will continue.
In the next iteration, Breda is marked as visited vertex because it does not
have an unvisited neighbour. Thus, Utrecht has the next best value and it yields
labels on Amersfoort and Amsterdam. The new value of Amsterdam is lower than
its previous one (93 < H I ) . The value of Amsterdam is therefore updated to 93
and the link from Rotterdam to Amsterdam is not valid anymore as shown in Fig.
4.4.g to i.
The next best city is Amersfoort, and its neighbors, Amsterdam and Zwolle,
are labeled. The new value of Amsterdam is bigger than its previous value (109 >
93) which will cause to maintain the old link between Utrecht and Amsterdam,
and Amersfoort is marked as visited. The next best city is Amsterdam, and
since Amsterdam is our goal vertex, the algorithm stops. The result of Dijkstra'a
algorithm is shown in Fig. 4.4.1.
Of note is the fact that this algorithm makes no attempt to direct "explora-
tion" towards the destination as one might expect. Rather, the sole consideration
in determining the next "current" intersection is its distance from the starting
point. In some sense, this algorithm "expands outward" from the starting point
iteratively, considering every vertex that is closer in terms of the shortest path
distance until it reaches the destination. When understood in this way, it is clear
how the algorithm necessarily finds the shortest path. However, it may also reveal
one of the algorithm's weaknesses: its relative slowness in some topologies.
The main advantage of Dijkstra's algorithm is that it always finds the shortest
path. However, because the number of vertices in a pipe-routing application
are extremely large, it takes relatively much calculation time and occupies much
memory as it needs to memorize the state of all vertices which requires 0{V)
space of memory.
The time complexity of Dijkstra's original algorithm is 0(|1^P), because in
the worst case it needs to visit all vertices requiring time 0{V), and on each
iteration it needs to find the vertex with the best value that requires again 0{V).
In 1984, Fiedman introduced the use of Fibonacci Heap as a min-priority queue,
and it improves the performance to 0{VlogV). This is asymptotically the fastest
known single-source shortest-path algorithm for arbitrary directed graphs with
unbounded nonnegative weights.

Greedy Algorithm
The Greedy algorithm is the fastest of all varieties of the Best-first search tech-
nique, and it is well known as Best-first search algorithm itself. The Greedy
algorithm works in a similar way as Dijkstra's algorithm, except that it uses the
estimation (called a heuristic) of how far from the goal any vertex is. Instead of
selecting the vertex closest to the starting point, it only considers the estimated
distance to the goal. It runs much quicker than Dijkstra's algorithm because the
heuristic function guides the algorithm towards the goal quickly. For example, if
60 RELATED WORK 4.2

Figure 4.5: Greedy algorithm

the goal is to the north of the starting position, Greedy algorithm will tend to
focus on paths that lead northwards.
Fig. 4.5 illustrates how the Greedy algorithm works to find a shortest path
from Hardinxveld to Amsterdam. We introduce the actual distance (not a driving
distance but "as the crow flies") between Amsterdam (as the goal vertex) and
other cities, as the heuristic values of the algorithm (Fig. 4.5.a).
For the first iteration, it starts to find a city that is directly connected to
Hardinxveld and selects the city closest to Amsterdam. Fig. 4.5.b shows that
Rotterdam is closer to Amsterdam than Gorinchem.
The Greedy algorithm chooses Rotterdam as its next current city. In the
next iteration, it compares Utrecht and Amsterdam, and it obviously chooses
Amsterdam because Amsterdam is the goal as shown in Fig. 4.5.d.
In general cases. Greedy algorithms mostly (but not always) fail to find the
globally optimal solution, because they usually do not operate exhaustively on
aU the data. In terms of time complexity, the worst case of this algorithm is the
same as Dijkstra's original algorithm, 0 ( l y | ^ ) , but in practice it mostly visits a
lower number of vertices compared to Dijkstra's.
4.2 DETERMINISTIC APPROACHES 61

A* Algorithm

Historically, the A* algorithm is an improvement of Dijkstra's algorithm that

introduces the heuristic part in its cost calculation. Hart, Nilsson and Raphael
first described the A* algorithm in Hart et al. [1968]. This algorithm combines
the pieces of information that Dijkstra's algorithm uses (favoring vertices that are
close to the starting point) and the heuristic cost (favoring vertices that are close
to the goal).
Using the same example. Fig. 4.6 illustrates how the A* algorithm works to
find the shortest path from Hardinxveld to Amsterdam. The heuristic value h{n)
that is shown in Fig. 4.6.a describes the actual distance ("as the crow flies")
from Amsterdam to other cities. For a better explanation, in this figure the total
fitness value f{n) is shown as partial g{n) + h{n).
It starts by examining the direct neighbors of Hardinxveld, and compares the
total value fin). Fig. 4.6.c shows that the fitness value of Gorinchem is smaller
than Rotterdam's, so Gorinchem is visited by the algorithm. Then it examines
Breda and Utrecht and labels both with fitness value 129 and 86 respectively.
The next best city is Rotterdam. The A* algorithm examines Amsterdam
and Utrecht. For Amsterdam, it simply labels it, but for Utrecht, since that city
had been examined before, the old fitness value needs to be compared with the
new one. In Fig. 4.6.e, we can see that the new value is higher than the old one,
which means that the total cost from Hardinxveld to reach Utrecht is cheaper
through Gorinchem rather than through Rotterdam, so the link from Rotterdam
to Utrecht will not be used.
Utrecht has the best fitness value, so the A* algorithm starts to examine
Utrecht's neighbours, Amersfoort and Amsterdam, and gives a label for both
cities. Amsterdam already has a label, so the algorithm compares both fitness
values. As we can see in Fig. 4.6.g, the new fitness value is smaller that the old
one, so the valid link to Amsterdam is through Utrecht instead of Rotterdam.
Fig. 4.6.h shows that Amsterdam has the best fitness value. It means that
the solution is found as shown in Fig. 4.6.i.
The characteristic of the A* algorithm is highly sensitive to the behaviour of
the heuristic function that is used. If h{n) is always lower than (or equal to) the
cost of moving from n to the goal, then the A* algorithm is guaranteed to find a
shortest path. The lower h{n) is, the more the algorithm expands and the slower
it becomes.
If h{n) is exactly equal to the cost of moving from n to the goal, then the A*
algorithm will only follow the best path and never expands anything else, making
it extremely fast. Although we can't make this happen in all cases, we can make it
exact in some special cases. It is satisfying to know that given perfect information,
the A* algorithm will behave perfectly. If h{n) is sometimes greater than the cost
of moving from n to the goal, then the A* algorithm is not guaranteed to find a
shortest path, but it can run faster than Dijkstra or Greedy.
As in Dijkstra' algorithm, to improve the performance of the A* algorithm.
62 RELATED WORK 4.2

31km

Figure 4.6: A* algorithm

4.2 DETERMINISTIC APPROACHES 63

a b
Figure 4.7: Shortest path m a free space

Fibonacci Heap is used as min-priority queue. The time and space complexity of
the A* algorithm is the same as Dijkstra's algorithm, but since the heuristic part
directs the search toward the destination, the number of vertices that are visited
are smaller than for Dijkstra's. However, there is a possibility that the path that
is found is not optimal.

4.2.2 Maze Algorithm

In previous examples, we used a graph as the environment. In a graph, it is
assumed that the exact path (or every edge) that connects two vertices is already
pre-defined. In contrast, two vertices can be connected in many ways in a free
space. Fig. 4.7 shows that there are two solutions that connect node S with node
T with the same distance and number of turns.
Seen from another point of view, in fact a graph and a maze share the same
principle. In a maze, the space is divided into cells, and each cell has its own
associated cost. A graph is consists of vertices and edges that connect one vertex
with another. In principle, a cell is the same as a vertex, and an edge represents
the associated cost.
Starting from this subsection, we move on to survey shortest path algorithms
to find a path between nodes in a free space. In this subsection, the space is
divided into cells, and each cell has its own associated cost. Some cells are marked
as obstacles. The goal of any algorithm we study is to find the path that connects
the start cell (containing the start point) with the end cell. Fig. 4.8 shows an
example of a grid of cells in 2D space. In this example, it is shown that each cell
has eight direct neighbors, with only four of them directly accessible.

Lee's Algorithm

The most common maze algorithm was proposed by Lee [1961], and is called
Lee's algorithm. Lee presented a wavefront approach to solve the shortest path
problem. This approach is similar to Dijkstra's algorithm.
64 RELATED WORK 4.2

F i g u r e 4 . 8 : Grid of cells

"4' 3 "2 1 2 3 4 5 ff
3 2 1 S 1 2 3 4 b Ö
s 4 3 2 1 2 3 4 Ö 6
T 5 4 3 2 3 4 5 T
b 4 3 4 5
5 4 5
b

4 3 2 1 2 3 4 B 5
3 2 1 S 1 2 3 4 b B
4 3 2 1 2 3 4 b 6
'b 4 3 2 3 4 T
b 4 3 4 b
b 4 b
b

F i g u r e 4 . 9 : 2D Lee's algorithm
4.2 DETERMINISTIC APPROACHES 65

Figure 4.9 illustrates how Lee's algorithm works. The start cell is marked with
S, and the target cell is marked with T.
The first phase in Lee's algorithm is the wave propagation phase as shown in
Fig. 4.9.b. The start cell is labeled with a value 0, and it starts propagating by
walking to an unlabeled neighbor cell and marks this with its own label plus the
cost of that cell. This propagating process should not violate an obstacle cell,
and it will end when the target cell is reached, or all reachable cells in the grid
are labeled. The waving process runs as in the breadth-first search algorithm
explained above.
After the propagating process completes, it will do the backtracking phase.
This starts from the target cell, and it will move to the next neighboring cell that
has a lower value. All cells that are encountered during the backtracking are kept
as the result path. The backtracking phase can be seen in Fig. 4.9.c.
If a solution exists, this algorithm guarantees to find it. If every cell cost is
equal, the lowest number of steps is equal to the minimum cost, therefore the
propagation process can be stopped after it reaches the target. However, if the
cell cost is not uniform there is a possibility that the lowest step path is not
the minimum cost, therefore the propagation process must be continued until all
cells are labeled. In general, in the non uniform cost environment this algorithm
requires more time to find the optimal solution. The example shows that the
algorithm needs to explore many cells before it finds the solution.
If an obstacle exist, Fig. 4.10 shows that Lee's algorithm needs to explore an
even larger amount of cells.
Beside the high time complexity, Lee's algorithm also uses a lot of memory,
because it needs to save the state of all visited cells. In the worst case, Lee's
algorithm needs to memorize all cells in the space.

A* Algorithm

In the previous section, we have surveyed how the A* algorithm can be used to
find the shortest path between vertices in a graph. We also have seen that the A*
algorithm outperformed Dijkstra's algorithm in terms of the calculation time and
memory usage, and if we wisely choose the heuristic function then the solution
that is found by the A* algorithm is optimal.
Now, we investigate the performance of the A* algorithm to find the shortest
path between nodes in a free space. The A* algorithm requires the space to be
divided into a grid of cells in the same manner as when we use Lee's algorithm.
Also, as in Lee's, A* will perform two phases of searching. First it performs front
wave phase. Then after the goal is reached, it performs backtracking to the start
node.
The A* algorithm in a free space has fitness value ƒ (n) = g{n) + h{n), where
g{n) represents the cost that is needed to travel from the start node to node n, and
h{n) represents the estimated cost from node n to the goal node. The difference
 I
66 RELATED WORK 4.2

1 1 1 1 1 1 11 • 111 1^1''I^I^K I^PM I^H 1 1 1

I |s| 1 1 1 r h | 5 | 4 { 3 { 2 | i {s|i 1 1 11
•
r KlBl^hphnNH"
I T I^PI''ppr H ^ I S I Q I T I |
11 1 1 1 1 1 M^KhprhhR^r
1
11 I1 1I 11 11 I M n 1 1 1I I I|9|8
t |9|SKhprpFl''r
7{6p|6|7{8{g{ {
PI
1 I 1 11I 1 11 I I NhrhKhpl 1 1
1 11111 11111 1 1 1 kl^r p hi 1 1 1
9 8 9
9

1
"8" 7 6 5 4 3 2 1 2
7 6 5 4 3 2 1 S 1
8 7 6 5 4 3 2 1 2 10
9 8 7 6 5 4 3 2 3 9 T
9 8 7 6 5 4 3 4 8 9
9 8 7 6 5 4 5 6 7 8 9
9 8 7 6 b 6 7 8 9
9 8 7 6 7 8 d
9 8 7 8 9
9 8 9
9

F i g u r e 4 . 1 0 : BD Lee's algorithm with obstacle

4.2 DETERMINISTIC APPROACHES 67

^ s

ZZZZI IZZZ J

a b

?zzzz

Figure 4.11: 2D A* algorithm

is that normally in a graph, the value of h{n) can be pre-calculated, while in a

free space, the calculation of h{n) must be done on the fly.
Figure 4.11 and 4.12 show the performance of the A* algorithm to solve the
same problem as Lee's algorithm in the previous subsection. Comparing those
figures with Fig. 4.9 and 4.10, A* is the obvious winner.
However, if the obstacle is more complex, as shown in Fig. 4.13, the A*
algorithm still needs to observe many cells before it can find a solution, and in
the worst case, it needs time and memory as large as Lee's.

4.2.3 Line search algorithm

By using a maze algorithm, the shortest path problem in a free space can be
solved trivially. However, if the space is loaded with complex obstacle, even the
A* algorithm will suffer. In this subsection, we survey line-search algorithms that
should perform better than a maze algorithm in term of calculation time and
memory usage.

Mikami-Tabuchi's Algorithm
The first line-search algorithm was proposed by Mikami and Tabuchi [1968]. In a
two dimensional environment, this method starts with generating four lines (two
horizontal and two vertical) through the starting point S and target point T.
These lines are extended until they hit obstructions or the boundary of the space.
If a line generated from S intersects a line generated from T, then a connecting
path is found. If they do not intersect, they are identified as trial hnes of levelo.
These lines are stored in temporary storage for further processing.
68 RELATED WORK 4.2

i i
c
I d

Figure 4.12: 2D A* algorithm with obstacle

I
T

S •

c d

Figure 4.13: 2D A* algorithm with obstacle

4.2 DETERMINISTIC APPROACHES 69

1 1 1 1 0 1 2 2 2 2

«
>L X X X
f 1
'f •

's "I r "I r 1T r

» if--^-i!i--^-*--'!if
• • • • ^ 1
'» T - i
1
i - - r - * - - r
1 1 1 1 ^ 1
1 1 1 1 j , 1
T--r--i--p-*--r'
*f -;--;---;--f--j;C--;-
t 1 1 1 1 jK 1
>-K--:t-3 *•- -
1 1 1 1 1 1
1

1 1 1 1 0 1

Figure 4.14: Mikami-Tabuchi's algorithm

After that, the algorithm starts the iteration procedure. For each step i of
iteration, it creates trial hnes for each grid leveli, one at a time. Along the trial
line, all its base points are traced. Prom these base points, new trial lines (of
leveli-^-i) are perpendicularly generated to the trial lines of leveli. If a trial line
of leveli+i intersects a trial line (of any level) from the other terminal point, the
connecting path can be found by backtracking from the intersection point to S
and T. Otherwise, all trial lines of /ei)eZ,+i are added to temporary storage, and
the iteration procedure repeated.
We use Fig. 4.14 to illustrate Mikami-Tabuchi's algorithm. At the beginning,
two lines (horizontal and vertical) are generated from the start point S and the
target point T as trial lines levelo. Since those lines are not connected, it generates
new trial lines leveli from 5 and also from T. At this stage, there is still no
intersection between the trial lines from S and T. Therefore it starts to generate
trial hnes levels- In Fig. 4.14, the intersection between trial line levels from S
and trial line leveli from T is shown by a red circle. Then the algorithm performs
backtracking from the intersection point to both S and T.
Even though this algorithm guarantees to find a path if it exists, it does not
guarantee finding the optimal path. However, the calculation time and memory
requirements are significantly less than for the A* algorithm.

Escape Algorithm

Hightower [1969] proposed an escape algorithm that modifies the Mikami-Tabuchi

algorithm. Similar to Mikami-Tabuchi's algorithm, this method starts with two
70 RELATED WORK 4.3

0 2

* • O C

PI
^ - ^ ^ ^
0
0 d

--0-
0

Figure 4.15: Escape algorithm

perpendicular lines through the starting point S. However, instead of using all
line segments perpendicular to a trial line, it considers only those lines that can
be extended beyond the obstacle which blocked the preceding trial line. The
intersection point between horizontal and vertical lines is called the escape point,
and one line segment only has one escape point.

Fig. 4.15 illustrates how the escape algorithm finds a path between the start
point S and the target point T. It starts by generating two lines (horizontal and
vertical) from S (line x and y) and T (line u and v), and those lines are called
trial Hnes levelo. Then it starts establishing the new trial lines leveli. In contrast
with Mikami-Tabuchi's algorithm, Hightower considers only a line that can be
extended beyond the obstacle that blocks the previous trial line. In this example,
the trial line leveli from 5 is line r because it can be extended beyond the obstacle
that blocks line y, and the escape point is point a. In the same manner, the leveli
trial line from T is line s, and the escape point is point b. Then line t is found
as the trial line leveh from S. This line intersects with One s, so point d is the
intersection point between trial lines from S and T. Prom here, the algorithm
performs back tracking to S and T, and the path S - a - c - d - b - T i s found.

The escape algorithm is faster and uses less memory space compared to Mikami-
Tabuchi's algorithm, but it cannot guarantee to find a solution.
4.3 HEURISTIC ALGORITHM 71

4.3 Heuristic Algorithm

4.3.1 Genetic Algorithm
Genetic Algorithm (GA) is an adaptive method which can be used to solve search-
and-optimization problems. It is based on the genetic processes of evolution of
biological organisms. Over many generations, natural populations have evolved
according to the principles of natural selection. By adopting this process, genetic
algorithms are able to "evolve" solutions to real world problems, Goldberg [1989J.
In genetic algorithm, a population of chromosomes which encode candidate
solutions to an optimization problem evolves toward better solutions. lYadi-
tionally, solutions are represented in binary form as strings of Os and Is. The
parameters are converted into discrete values with a certain resolution. For ex-
ample, with 10 bits per parameter, we obtain a range with 1024 discrete values. If
the parameters are actually continuous then this discretization is not a particular
problem. This assumes, of course, that the discretization provides enough resol-
ution to make it possible to adjust the output with the desired level of precision.
If some parameters are discrete values then the coding issue becomes more
difficult. For example, suppose there are exactly 1500 discrete values which can
be assigned to variable X,. We need at least 11 bits to cover this generating 2048
discrete values. Since there are only 1500 values, the 548 unused bit patterns may
result in no evaluation. Also some parameter settings may be represented twice
so that all binary strings result in a legal set of parameter values. Solving such
coding problems is usually considered to be part of the design of the evaluation
function.
The evolution usually starts from a population of randomly generated chro-
mosomes and happens in generations. In each generation, the fitness of every
chromosome in the population is evaluated, multiple chromosomes are stochastic-
ally selected from the current population (based on their fitness) and modified
(recombined and possibly randomly mutated) to form a new population. The
new population is then used in the next iteration of the algorithm. Commonly,
the algorithm terminates when either a maximum number of generations has been
produced, or a satisfactory fitness level has been reached for the population. If the
algorithm has terminated due to a maximum number of generations, a satisfactory
solution may or may not have been reached.
In genetic algorithm, reproduction is one of the most important phases. During
this phase, first the parents must be selected and then crossover occurs. In actual
biological evolution, the genes from parents form in some way the whole new
chromosome. The newly created ofllspring can then be mutated. Mutation means,
that the elements of the DNA are changed a bit. These changes are mainly caused
by errors in copying genes from parents.
In genetic algorithms, there are several methods for selection of the chro-
mosomes from a population for reproduction; such as Roulette wheel selection,
Boltzmann selection. Tournament selection, and Rank selection. The most pop-
72 RELATED WORK 4.3

Figure 4.16: Roulette wheel selection

Figure 4.17: One point and Two points crossover techniques

ular one is the Roulette wheel selection which is based on the fitness value of
all chromosomes. For example, suppose there are six chromosomes in a genetic
algorithm population and that they are sorted based on their fitness value. Fig.
4.16 illustrates the Roulette wheel. It shows that the proportion of the wheel that
is allocated to a particular chromosome difl^ers according to the fitness value of
that chromosome. Then the selection process is similar to a Roulette wheel in a
casino. The wheel is rotated and stopped randomly, and the selected chromosome
is chosen as the new parent. The probability for the chromosome with biggest
fitness value to be chosen is bigger than for the others. However, there is still a
probability that the relatively unfit chromosome is selected.
After two or more chromosomes are selected, the crossover process is started.
Crossover is a genetic operator used to vary the chromosomes from one generation
to the next. Crossover is a process of taking more than one parent chromosome
and produce a child chromosome from them. Two of several techniques of cros-
sover are shown in Fig. 4.17. Fig. 4.17.a shows one point crossover technique,
and Fig. 4.17.b shows two points crossover technique.
4.3 HEURISTIC ALGORITHM 73

Beside crossover, mutation also might occur. Mutation is a genetic operator

used to maintain genetic diversity from one generation of a population of al-
gorithm chromosomes to the next. Mutation occurs during evolution according
to a mutation probability. This probability should be set low. If it is set too high,
the search will turn into a primitive random search. There are some different
types of mutation. The simplest one is a bit string mutation, that flips one bit at
random position from 0 to 1 or vice versa.
The notion of evaluation and fitness are sometimes used interchangeably. How-
ever, it is useful to distinguish between the evaluation function and the fitness
function used by a genetic algorithm. The evaluation function provides a measure
of performance with respect to a particular set of parameters only. The fitness
function transforms that measure of performance into an allocation of reproduct-
ive opportunities. In the canonical genetic algorithm, fitness is defined by ƒ,/ƒ
where /j is the evaluation value associated with chromosome i and ƒ is the average
evaluation value of all the chromosomes in the population. However, depending
on the selection method that is used, different formulas may be used.
The key factor of genetic algorithm is defining the evaluation function. Nor-
mally, developing an evaluation function can sometimes involve developing a sim-
ulation. The evaluation function value determination must also be relatively fast.
This is typically true for any optimization method, but it may particularly pose
an issue for genetic algorithms. Since a genetic algorithm works with a population
of potential solutions, it incurs the cost of evaluating this population.
There are many ways to utilize genetic algorithm to solve the shortest path
problem. One of them is by dividing the space into a grid, and we define a
chromosome as a turning grid location of the path. For our case of pipe routing
such location would indicate a pipe bend. Then the number of chromosomes in
a population is equal to the maximum number of turning locations. Based on
that, the evaluation function is defined as the shortest path between start and
target points. Also to avoid a collision with any obstacle, we need to add a large
penalty value in the evaluation value if there is a segment of the path that collides
with an obstacle. Genetic algorithm then evolve and eventually will give a set of
chromosomes as the result.
For example, the evaluation function could be the number of nodes in a path
plus the number of turning points plus a very large value if the path collides with
an obstacle.
Fig. 4.18 shows some snapshots of genetic algorithm evolution to find the
shortest path between node S and node T. In Fig. 4.18.b there are three turning
points found by the algorithm (marked as blue nodes). The path that was founded
contains collisions with the obstacles, so the evaluation value is very large.
Fig. 4.18.C shows an evolution snapshot of the genetic algorithm that found
six turning points. The evaluation value is 20 nodes -|- 6 turning points. Then
after several evolutions, the optimal solution is solved by the algorithm as shown
in Fig. 4.18.d (16 nodes -I- 3 turning points).
One of disadvantages of using genetic algorithm to solve the shortest path
74 RELATED WORK 4.3

Figure 4.18: Genetic algorithm search

problem is that this algorithm needs many generations of a population before it

finds a good solution. In every generation, the algorithm needs to re-calculate its
evaluation value. In the previous example, during the calculation of the evaluation
value, every segment of the path must be tested against all obstacles, and collision
detection is relatively expensive. Beside that, there is no guarantee that the
solution is the optimum one.
On the other hand, the method that we use in the previous example does not
need to memorize the state of cells. So it requires less memory space than Lee's
or the A* algorithm.

4.3.2 Particle Swarm Optimization

Particle Swarm Optimization (PSO) was first introduced and originally developed
by Kennedy and Eberhart [1995]. It refers to an algorithm that is used to find
optimal solutions of numerical and qualitative problems. It emerged from earlier
experiments with algorithms that modeled 'flocking behavior' seen in many species
of birds. PSO has been used in the last few years to solve different kinds of
optimization problems, such as training a neural network model, Salerno [1997],
tracking and optimizing dynamic systems, Eberhart and Shi [2001], power system
control, Yoshida et al. [2001], scheduling of manufacturing systems, Jerald et al.
[2004], and tuning PID parameters in an AVR system, Gaing [2004].
4.3 HEURISTIC ALGORITHM 75

Standard PSO

In the original framework, each individual in PSO, called a particle, moves in the
search space with a velocity that is dynamically adjusted according to its own
experience and its companions experience. As each particle moves through the
problem space, it evaluates the fitness function, and memorizes its best position
(the position giving its own best fitness value pbest) as a pbesix. Each particle
also memorizes the best global solution (gbesL), which is the best fitness value
among all particle in the population, and its position (gbesix), obtained so far by
any particle in the population based on the fitness function.
While moving toward its pbest and gbest location, each particle, at each time
step, changes its velocity and its current position according to:

Vi = aV, + Ciai{pt - x j + C20-2(Pg - Xt) (4.1)

a;, = a;, + Vt (4.2)

where
Ui velocity of t h e i " ' particle;
a inertia weight;
C], C2 acceleration constant;
ai,(72 random numbers in the range [0.. .1];
X, position of the i"* particle;
PI pbest of the i"' particle;
Pg gbest of population, index g represents the index of the best particle
among all particles in one population.

The first part of (4.1) represents the mertia weight which was first introduced
by Shi and Eberhart [1998]. This term serves as a memory of the previous velocity.
A large inertia weight favors exploration and a small inertia weight favors exploit-
ation. The second part is the cognition component, representing the exploiting
of its own experience. The third part is the social component, representing the
shared information and mutual cooperation among particles.
There are two types of boundaries; velocity maximum Vmax and position
boundaries Xmin and x^^ax- If the velocity according to (4.1) exceeds Vmax,
then it is limited to Vmax. In same manner, if position x, in (4.2) exceeds the
position boundaries, then the particle is placed on the boundary. Vmax needs
to be chosen wisely, because it influences the convergence of the search. If Vmax
is too high, particles might move too fast, passing a good solution. If Vmax is
too small, the particles may not sufficiently explore the search space Fan and
Shi [2001]. Early experience with particle swarm optimization indicated to set
the acceleration constant ci and C2 equal to 2.0, and Vmax equal to 20% of the
total range of the variable. The standard PSO algorithm is shown in Fig. 4.19,
Kennedy and Eberhart [1995]-Fan and Shi [2001].
76 RELATED WORK 4.3

Initialize PSO
for J = 0 to numher of iteration do
for t = 0 to number of particle do

If f: < /(p.) then

/(Pi) = fM and Pt = x,
If /(p.) < f(Pg) then
/(Pg) = /(ait) a n d Pg = x,
t), = ati, + cicr](p, - a;,) + C2(T2(P9 - a; )
Xj = Xi + n.
end
end
liPg) and Pg are the outputs

Figure 4.19: The Standard PSO AlgonLhm

Improved Version of the PSO Algorithm

Results show that PSO finds good solutions much faster than other algorithms,
Angelino [1998], Pcram et al. [2003], Tao et al [2004], Krink et al. [2002[ and Tao
et al. [2003], however for some type of problem, Angeline [1998] shows that after a
few number of iterations the quality of solutions can not be improved. Especially
in strongly multi-modal problems, PSO may suffer from premature convergence
Tao et al. [2003].
To overcome that problem, much work to improve PSO has been done. The
main work was on parameter modification, increasing diversity, and variations of
the PSO algorithm.
PSO needs predefined parameters, i.e. swarm size, maximum velocity, weight
inertia, individual and social factor. The ability to find a globally optimum solu-
tion relies greatly on the setting of these parameters. Tao et al. [2004] proposed
an adaptive PSO by choosing parameter to increase stability and avoid prema-
ture convergence. Other works on the PSO parameters are described in Shi and
Eberhart [2000] and El-Gallad et al. [2002].
The diversity of the swarm needs to be high, while low diversity would lead
to fitness stagnation of the swarm. If this happens close to a the local optimum
region, the swarm can be trapped in a local solution. L0vbjerg and Krink [2002]
introduced self-organized criticality (SOC) to maintain diversity of the swarm, by
resetting particles that are too close to each other. Another approach to reset the
swarm to increase diversity was proposed by Clerc [1999] by defining a no-hope
convergence criterion and a re-hope method so that, from time to time, the swarm
re-initializes its position, according to some gradient estimations of the objective
function and to the previous re-initialization.
Modification of the PSO algorithm itself has been done to achieve better and
faster convergence. Most of it hinges on mixing the PSO algorithm with other
4.3 HEURISTIC ALGORITHM 77

oplimizaUon methods. The combination of PSO with hill climbing is used by Lim
et al. [2003] to solve the bandwidth minimization problem. Another hybrid ap-
proach introduced by Krink and L0vbjerg [2002] apply GA, PSO and hill climbing
simultaneously.

Variations of the Standard PSO

Another approach to modily the PSO algorithm is by using different types of

probability distribution to generate the random number. Krohling et al. [2004] use
a truncated Gaussian probability distribution. IVuncation is done for the values
of the random number less than (-1) and greater than 1. To assure that the PSO
method converges, the random numbers c i , a-i have to be positive. Because of
that requirement, the generated random numbers are mapped to generate random
numbers in the interval [0,1].
To ensure the convergence of the PSO, Clerc and Kennedy [2002] introduced a
constncLion factor K that replace the stochastic terms CICTI and C21T2 into 0.729.
Another approach called Gaussian PSO was proposed by Krohling [2004]. It uses
the absolute value of the Gaussian probability distribution A' (0,1) for automatic
generation of the stochastic terms of PSO which has a mean value equal 0.798. So,
there is no more need to specify the accelerating constant ci and 02- Furthermore,
the momentum term is not used by setting the inertia weight a equal zero and
therefore the maximum velocity Vmax is no longer necessary. The Gaussian PSO
only has the swarm size and the boundary of the search space as its parameters.
However, Kennedy [2005] stated that the inertia term adds several unique
characteristics to the particle swarm, and it should be maintained. Based on this,
the Gaussian PSO with inertia term is used as one variant in this thesis. The
velocity term in the Gaussian PSO was then modified to:

v^ = avt + I/?andi KPJ - x j + \rand2\{pg - x,) (4.3)

where Bandi and Rand2 arc the random numbers generated by taking the abso-
lute value of the Gaussian probabihty distribution A' (0,1).
To solve the shortest path problem using this algorithm, we can use exactly
the same method that was used for the genetic algorithm.
In some literature, it is mentioned that PSO normally converges faster than
the genetic algorithm. However, it still needs quite a large number of generation
before it can find the solution. Smce the calculation of the evaluation value is
expensive, it still takes a lot of computation time. Also, as in genetic algorithm,
the solution from PSO is not guaranteed to be the optimal solution.

4.3.3 Ant Colony Optimization

The ant colony optimization algorithm (AGO) is a probabilistic technique for solv-
ing computational problems which can bo reduced to finding good paths through
graphs. The fundamental idea is inspired by the foraging behavior of real life ant
78 RELATED WORK 4.3

Food «4» °<» °<» °<» ° # . Nest Food Nest

^W^ ^9^ T W ° TW® TW^ ^W^ "Wf" TW^

Food °wr "wr * w Nest Food "«f Nest

•Mi»
s
^
Figure 4.20: Ant colony optimizalion

colonies in which individual ants deposit a substance called pheromone on the

path while moving from the nest to the food sources and vice versa. Thereby,
a pheromone trail is formed through which individual ants are aided to smell
and select their routes. The paths with higher pheromone doses would be more
likely to be selected by other ants bringing on further amplification of current
pheromone deposits, which leads to a positive feedback process. Due to this in-
teresting behavior of ants, after some time the shortest path from the nest to the
food source or vice versa would be formed.
Fig. 4.20 illustrates the ant colony optimization. Fig. 4.20.a shows an estab-
lished path between the nest and the food. Then if we put an obstacle as shown
in Fig. 4.20.b, the path becomes obsolete. Fig. 4.20.C shows how the colony of
ants finds two paths to go around the obstacle. Since the path below the obstacle
is shorter than the upper path then the times that is needed to travel through
the lower path is smaller. Then after a while, the number of ants that using the
lower path is larger, therefore the pheromone trail is stronger in the lower side.
Then the new shortest path is found.
The ant colony needs to register the amount of pheromone on each visited
cell. In the worst case, all cells are visited by the ants. For a graph, the space
complexity is 0{E) with E is the number of edges in graph. In a grid space, it
becomes 0{C) with C the amount of cells in the grid, the same as Lee's or the
A* algorithm.
On every repetition, a colony of ants explores a graph or a grid until the target
cell is reached. Comparing with Lee's or A*, the time that is needed by the ant
colony on one repetition is shorter. However since it needs many repetitions, it
takes much longer time before a solution is found.
4.4 COMPARISON OF ALGORITHMS 79

Table 4.1: Comparison of algorzlhm

Maze Line Heuristic

Lee A* Mikami Hightower GA PSO
Find path? 5 5 5 3 5* 5*
Optimum? 5 4 3 2 4* 4*
Time Avg 67ms 30ms 13ms llms 2min l.Smin
Space 0{C) « 0{C) 0(LM) «0{LM) O(Ppso)
Weighted? Yes Yes No No 1 Yes Yes

'J'he implementation of ant colony optimization to solve the shortest path

problem is normally applied on a graph or a map wfith predefined boundaries to
speed up the searching process and also to minimize the memory usage.
In the same way as for other heuristic methods, the solution from ant colony
optimization might not be the optimal one.

4.4 Comparison of Algorithms

In the beginning of this section, we have defined our criteria to measure the
performance of the algorithms; always find a solution if it exists, always find
the optimal solution, and use small resources in terms of computer memory and
shorter computation time. We also consider the flexibility ol those algorithms to
be extended and adapted to our methodology.
In order to compare the performance, we create a simple 3D uniform weight
grid type environment and use it as the test case to measure the performance of
the shortest path algorithm mentioned above.
Since this test is only a rough comparison, we use the standard variety of each
algorithm, for example the genetic algorithm variance that we test is its standard
form.
For the genetic algorithm, wo choose to use 50 different populations, and the
maximum number of regenerations is 100 times. Each population consists of 6
chromosomes, because the maximum number of bends that we expect is 6. The
choice to use the roulette wheel as the selection method and use the two point
crossover method.
We also use 50 different populations and a population size of 6 members for
the particle swarm optimization. The maximum number of iterations is also 100.
For both PSO and GA, the iteration process stops if at least one valid solution
has been found and the result is no longer improved after 5 further iterations.
Table 4.1 summarizes the performance of the algorithms. For the first two
criteria, the range is between 1 and 5, with 5 meaning always and 1 meaning
never. For heuristic algorithms, an asterix means that this grade can be achieved
by running the algorithm with the number of generations less than 100.
80 RELATED WORK 4.5

Mikami-Tabuchi and Hightower algorithms cannot be implemented in a weighted

environment and do not have a good performance to find the optimum solution,
which means that our methodology cannot use it as the detail routing algorithm.
However, due to its speed and low memory usage, they can potentially be used
to test whether a certain pipe can be routed or not. The table shows that even
though the Hightower algorithm has a better performance in term of memory
usage, it does not always find a solution, so it won't be considered anymore, and
we choose to utilize Mikami-Tabuchi in our methodology for that function.
The other four algorithms are suitable to be used in a weighted environment,
and they also always find a solution if it exists. First, let's take a look at the heur-
istics algorithm; genetic algorithm and particle swarm optimization. Comparing
the performance between heuristic algorithms is not trivial. 'I'here is a famous
theory introduced by Wolpert and Macready [1997] in connection with the prob-
lems of search and optimization. They say there is "no free lunch" (NFL) in
search and optimization. In short, NFL says that on average, all optimization
algorithms have the same performance, but for a certain kind of problem and
condition, one algorithm might be better than the other.
In the shortest path problem, both GA and PSO use the same implementation.
In our example case, PSO always finds a solution in fewer iteration compared to
GA. If we use the optimized variety of GA or PSO, there is a possibility that the
solution can be found faster, however it is still incomparable to the speed of the
deterministic algorithm to find the solution.
In terms of memory usage, for precise routing with a small cell size and a small
number of pipes, GA and PSO implementation is better than the deterministic
algorithm.
As we mentioned in the previous subsection, it takes many generations for
heuristic algorithms to find a good solution. So in terms of time complexity, they
are much slower than the exact algorithm; Lee's and the A* algorithm.
For completeness sake, we also tried to use GA and PSO to route a group of
pipes. As a result, we found out that both methods are able to route a small group
of pipes, but the time that was needed is much longer than for the dctermirustic
algorithm. Also, both failed to route a group of pipes if the group contained more
than 5 pipes.
If we only consider the certainty to find the optimum solution, Lee's algorithm
is the winner. However the speed and significantly lower memory usage of A*
algorithm leads us to select this as our main algorithm.

4.5 Beyond the Shortest Path Problem

4.5.1 Multiple Nodes
Most of the shortest path algorithms are basically only concerned with finding
the shortest path that connects two nodes, while in practice, more than 70% of
4.5 BEYOND THE SHORTEST PATH PROBLEM 81

'c

E
'

(a) (b) (c)

Figure 4.21: Multiple nodes

-rm ü«ife
w^-»A ^^^
• * • "
a " a

1
il b. 1 d
„
--
»
& 1
bi 1 -
J lb
H* • üü"?p^ — [ — 1r"^ ËÜÜÜÜÜ
„

_
(a) (b) (c)

Figure 4.22: Mulually miervenmg case

pipes have a branch. There is a specific algorithm to solve this optimally, called
the Steiner 'IVce Problem, Bern and Graham [J 989]. However, it is not as flexible
as the A* algorithm that we adopted in our methodology.
We use Fig. 4.21 to illustrate the multiple nodes routing. Let us say that we
have five nodes to be connected as shown in Fig. 4.21.a. Fig. 4.21.b is the result
by performing the A* algorithm lour times to connect A to B, B to C, C to D,
and D lo 1'J l'"ig. 4.21.c shows the result by using Stomer l ï e e .
In chapter 5 we will discuss a modified A* to solve the branching problem.

4.5.2 Mutually Intervening Case

In practice, it might happen that two or more pipes block each other when they
arc optimally routed. As an example, Fig 4.22 shows that if pipe a is optimally
routed, pipe b is blocked. Also in the other way around, li pipe b is optimally
routed, it blocks pipe a.
82 RELATED WORK

1
d •
•
t
• iiiliÉ

(a) (b) (c)

Figure 4.23: Routing order

In chapter 5, we will look in more detail into this problem and how our meth-
odology solves it.

4.6 Combinatorial Optimization

If we want to route more than one pipe, the order of the routing is very important.
Fig. 4.23 shows how the order of the routing might change the overall quality
of the pipes. This kind of problem is trivial if the number of the pipes is small.
However in practice, we like to route many pipes and expect optimum overall
quality.
To solve this combinatorial problem, we need to utilize the heuristic optimiz-
ation method. One of the most famous methods is the variation of the Particle
Swarm Optimization algorithm, called Discrete PSO (DPSO), Clcrc [2004],
The Discrete PSO can be formalized as follows:

•u, = aVj©Cj(pj-X,) ® C2(pg-Xj) (4.4)

Xt = Xi + Vi (4.5)

As can be seen by comparing eq.(4.1) with eq.(4.4) and eq.(4.2) with cq.(4.5),
there is no formal difference between classical PSO and DPSO. However slightly
different rules are imposed. The search space S is the finite set of all sequences
of the pipes to be routed. The position Pi is one of the possible sequences. The
interesting part is the velocity, since the meaning of movement of the particle is
not the same as in the classical PSO. In DPSO, the velocity is in a form of a list
of transpositions. For example, v = (2, 5) means that if this velocity is applied to
a position p, = (0,1, 2, 3,4, 5), it generates a new position p, = (0,1, 5, 3,4, 2).
In his work, Clerc [2004] develops DPSO and proves the performance of this
algorithm to solve the asymmetric 'IVaveling Salesman Problem, since TSP is well
known as one of the most important test grounds for the combinatorial problem.
4.7 MODEL SIMPLIFICATION 83

Figure 4.24: Example of model simphficaUon

4.7 Model Simplification

Three dimensional models play an important role in many applications such as
computation fluid dynamics and computer-aided design. I'hose applications use
3D models not merely to display the actual object as well as possible, but also
for scientific calculations like the physical and dynamic behavior of that object.
With the development of 3D scanner and computer graphic technology, a very
precise 3D model can be created. For example, the 3D models that are used in a
ship design process have a very high level of detail. Unfortunately, the complexity
of these models, measured by the number of triangles, seems to grow faster than
the ability of the hardware to calculate and render it interactively. Put in another
way, the level of detail of those models always seems to exceed the level that we
can afford.
To overcome this discrepancy the algorithm to simplify the 3D models has
been investigated for years. Most of them are dealing with reducing the number
of polygons, Garland and Ileckbert [1997], Gotsman et al. [2002], Ilua-hong et al.
[2007], F.S.Nooruddin and 'lYirk [2003], Cohen et al. [1996], Hjelmervik and Leon
[2007], Barber et al. [1996], and Kirkpatrick and Seidel [1986]. Another way of
3D model simplification is by translating the polygon soup (a group of unorgan-
ized triangles, with generally no relationship whatsoever) to volume visualization
{voxel) Kaufman [1994].
Generally there are two main objectives of model simplification, for faster
rendering and/or faster computation time of the physical aspect of the model. In
practice, one application might require a different level of detail compared with
other applications.
In some applications, the high level of detail is needed to maintain the similar-
ity of the shape of the original model and the simplified model. For this purpose,
we can use approaches of Garland and Heckbert [1997], Gotsman et al. [2002],
84 RELATED WORK 4.8

Hua-hong el, al. [2007], F.S.Nooruddin and Turk [2003], Cohen ct, al. [1996], and
Hjelmervik and Leon [2007]. The simphfied model generated by these approaches
still has a relatively high density ol triangles.
On the other hand, some applications need only the approximate shape of the
original models, for example in the application that needs to detect the collision of
two or more objects. An established method to achieve this goal is by finding the
convex hull of the model Barber et al. [1996] and Kirkpatrick and Seidcl [1986].
The convex hull of the model normally only consists of a relatively low number of
triangles. However, because the simplified model is always a convex wrapping of
the original model, it has a drawback if the original model is not close to a convex
shape.
In our methodology, we deal with thousands of 3D objects. Due to the very
large and complex calculation, and the fact that most pipes are routed in an
orthogonal way, we need a simplified model that only contains a small number
of cuboids for faster calculation. For this purpose, those approaches that are
mentioned above are not suitable to be used.
In chapter 5, we will discuss our method to simplify the 3D model as shown
in Fig. 4.24.

4.8 Summary
In this chapter, we have reviewed and compared the existing shortest path al-
gorithms to identify the algorithm that should be adopted in our methodology.
Based on our experiment described in Section 4.4, we have decided to use the
A* algorithm to find the shortest path in the single pipe router module of our
methodology. We also will use the Mikami-Tabuchi algorithm for the blockage
checker module, due to the fact that this algorithm can find a solution faster and
use less memory than the A* algorithm.
During our experiment, we found out that the heuristic algorithm is not suit-
able to route many pipes at the same time. However, it works nicely to route a
group of 2-3 pipes. This leads us to further investigate it to solve the mutually
intervening problem that will be discussed in detail in Chapter 5.
As described in the previous chapter, since our methodology routes pipes one
by one, the problem of optimizing the quality of a group of pipes becomes a
combinatorial problem. In this chapter, we also chose the heuristic algorithm
that will be used to solve the combinatorial problem.
Beside that we also described some other difficulties beyond the shortest path
problem, and our solution to these issues will be presented in Chapter 5.
Chapter 5
The Methodology Architecture and
its Implementation

5.1 Introduction
Chapter 3 discussed the core functionaUty that must be implemented in our pipe
routing methodology. These requirements are based on the current pipe routing
process in practice that has been discussed in Chapter 2.
Since a few decades ago, the pipe routing problem has been an interesting
research subject, and some selected researches that have most relevance for our
methodology have been reviewed.
In Chapter 4, we have reviewed and compared some optimization algorithms
in more detail. The comparison results led us to select the suitable optimization
algorithms to be used in the methodology. Some other important matters such
as model simplification and combinatorial optimization were also explained or
touched upon.
In Fig. 3.1, the outline of the proposed methodology architecture is shown
from the functionahty point of view. In our methodology, the function for data

Pipe Router Module

Interface Module

Figure 5.1: The Outline oj Proposed Methodology Arclaleelure

85
86 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.2

retrieval and to send result back to CAD software are combined into one module
called the interface module. Therefore, Fig. 3.1 can be represented as Fig. 5.1
that shows that the architecture of our proposed methodology consists of two
main parts; the interface module and the pipe router module. In this chapter, we
will look into both modules in more detail.
While the interface module is scientifically less interesting, it is an integral part
of our methodology since this research investigates and validates a methodology
that in principle should lend itself for practical application. For that reason we
must also dwell on it at some length.

5.2 Interface Module

As described in Chapter 3, the proposed methodology is used together with a
3D CAD software package. First it retrieves the required data from the CAD
software packages, and then sends back the result to them. All functions for data
retrieval and to send back the results are implemented in the interface module of
the methodology.
The data that are required to perform the pipe routing is already discussed
in Section 3.2 and the conclusions are summarized in Section 3.3. It is also
mentioned that some data are not available in the CAD software, therefore the
proposed methodology must consist of some other tools to generate them.
The first tool that is needed is the constraint editor tool. This tool is a simple
tool for a user to manually add an object constraint, such as a NoGo area, or a
virtual obstacle. The details of an object constraint will be discussed later in this
chapter. However, because it is basically a computer programming problem, it
will not be discussed in detail in this thesis. What is important however, is that
the methodology can handle such arbitrary constraints or NoGo area.
As mentioned above, some of data are available in CAD software. However,
most of the time the format of those data cannot immediately be used in our
methodology. Therefore it is required to include an interface between our proposed
methodology and the CAD software.
Following our discussion in Section 3.3, the smart P&I diagram tool must also
be included in our methodology.
In Section 4.7 we have discussed the importance of model simplification before
we use them for optimization calculation in the proposed methodology. Since
this matter is important to arrive at a computationally feasible solution, and
apparently no research has been done to solve it in the way we require, in this
section the algorithm to simplify the 3D models is described in more detail.

5.2.1 Interface to CAD Packages

Since most of the data that are required by our methodology are already available
in the CAD software packages, we would like to make use of it. In today situation.
5.2 INTERFACE MODULE 87

\
3DDXF

2DDXF

XML

Figure 5.2: The Interface to CAD software packages

most of the CAD software packages have an export function to dump the data
from their system to be used by external applications. However, this is not always
a trivial case, because not every data can be dumped using the standard export
functions. Moreover, every CAD software package uses its own kind of standard,
and stores its data in a different way.
Let's take a look at our case which is important for validating the methodology
and its implementation. At this moment, our laboratory implementation with
both Cadmatic and lYibon M3 software packages. Even though both can be used
to design a ship, they are fundamentally different in terms of internal data. Also,
their export functions are totally different, and will produce an output in different
formats.
To tackle this problem, it is decided that in our methodology we will use
a common type of data format, the DXF format. As widely known, DXF has
become the de facto ASCII standard file format for CAD drawing exchange.
The data are divided into three different categories; the 3D volume data are
converted to 3D DXF, the 2D drawing data are converted to 2D DXF, and data
that contains the information are restructured into an XML file. As shown in Fig.
5.2, the required data from the CAD software packages will be converted to the
common format. Further details are not necessary to appreciate the validation in
Chapter 6.

5.2.2 Smart P&l Diagram Tool

As described in Subsection 3.2.1, the smart P&I diagram acts as the primary
guidance of the pipe router module in our proposed methodology. Currently,
even though some commercial 3D CAD software vendors already implement a
concept of smart P&I diagram in their software packa,gcs, it is still relatively new
88 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.2

and only a limited number of shipyards have access 1,0 that. Therefore, to fill
in the gap temporarily and support the validation of our methodology, a simple
Smart P&I diagram tool is included in the methodology.
In a manual pipe routing process, a Smart P&I diagram tool is not necessary.
A simply plain printed drawing of P&I diagram is enough. It is because the
pipe engineer is able to read the diagram and based on that he can route a pipe
correctly. There are basically five things of importance included in a P&ID:
1. Pipe specification, such as the pipe diameter and the type of the pipe.
2. Piping system that a pipe belong to. It is important because every system
must follow a specific set of rules.
3. Geometrical location of the start and end nozzles that the pipe needs to
connect.
4. Valve specification; defining the type of valve such as a butterfly valve, and
defining the valve location type (see below).
5. Branch location type (see below).
Without implementing a Smart P&I diagram tool, the task to define this
input to the router module will take a lot of time. This would unduly hinder our
validation work and also raise question with regards to the practical applicability
of our methodology. The first three items listed above can be seen immediately
in a plain P&I diagram, but the fourth and the fifth can only be seen by the
experienced pipe engineer.

Valves

Two main things are described in a Smart P&I diagram regarding a valve. The
first one is the technical type of valve. It is important since during the pipe routing
process the valves will be placed automatically by the pipe router module. This
feature also prevents a pipe engineer to make a mistake like shown in Pig. 2.6
The second part is the valve location type. In practice, a pipe engineer places
a valve in two different ways. The first one applies to a certain kind of valves: a
pipe engineer must define the location of that valve in space and after that routes
the pipe through that location. This must be done if the valve belongs to the
group of control valves. We call this valve a fixed valve. For other types of valve,
normally the pipe will be routed first, then after that the valve that belong to
that pipe will be placed in that path. This kind of valve is called a floating valve.

Branches

Even though pipe branch cases are neglected in many studies about automatic
pipe routing, in practice most of the pipes in a ship have a branch. For that
reason, branches must be included in any relevant methodology. As with the
valve location type, the branch types are divided into two types; a fixed branch
and a floating branch.
5.2 INTERFACE MODULE 89

O © o -©

© ©

-©

Figure 5.3: Branch type

Fig. 5.3 shows the difference between a fixed branch and a floating branch.
Assume there are three nozzles that must be connected as shown in Fig. 5.3.a.
Fig. 5.3.b shows the example of a fixed branch. In a fixed branch there is a
master pipe and a child pipe. In this case, the pipe between nozzle 1 and nozzle 2
is routed, as a master pipe, then nozzle 3 is routed to the master pipe, as a child
pipe.
In a floating branch, all nozzles have the same priority. Thus, the master pipe
can be the pipe that connect nozzle 1 and nozzle 3 as shown in Fig. 5.3.c or the
master pipe is a pipe between nozzle 2 and nozzle 3 as shown in Fig. 5.3.d or
between nozzle 1 and nozzle 2 as shown in Fig. 5.3.e.

5.2.3 Simplification of the 3D IVIodei

The 3D models that are imported from the CAD software package have a high
level of detail. Not only an entity that has a large volume is complex, even a small
bolt is quite complex. As mentioned in Section 4.7, generally there are two main
objectives of model simplification, faster rendering and/or faster computation
time of the physical aspect of the model.

Simplification for Faster Rendering

In any viable methodology, this kind of simplification is needed to improve the
performance of the graphic user interface (GUI) part. As described in Chapter
3, not all data can be retrieved automatically from the CAD software, some data
such as a NoGo area is still needed to be entered manually. To make this task
easier, the GUI part of the methodology must have a good performance.
In Section 4.7, it was mentioned that many researches have been done to sim-
plify the 3D model, and some of them are useful for our piu'pose. We decided to
90 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.2

implement the quadric error metrics method for surface simplification that was
proposed by Garland and Heckbert [1997]. To make it fit our needs, some adjust-
ments in the implementation are made. For example for the steel construction,
instead of simplifying the model individually, it is better if a bunch of steel con-
structions are combined into a group and allow the simplification method to work
on the whole group. After that the simplified model needs to be split into the
original entities again so that the user can hide or show every single entity.

Simplification for Faster Calculation

Since our main goal is to derive an automatic pipe routing methodology and
framework applicable in practice and since designers need to route thousands of
pipes, the computation time that is required to route pipes needs to be short.
As we already discussed in Chapter 4, one of the most time consuming parts in
the pipe routing process is the collision avoidance. To avoid the collision between
the pipe that is being routed with other objects, such as the pieces of equipment
and the pipes that were already routed, the collision detection routine must be
performed in each step.
To test the collision detection between two real tessellated objects is compu-
tationally very expensive since basically we need to test the collision detection
for each triangle in the model. To tackle this problem many researches have been
done to speed up the collision detection routine. The common way is by simpli-
fying the 3D model into a collection of shapes that is more simple to be tested,
such as a bounding box.
In the next subsection, our proposed method to simplify the 3D model is
described in detail.

Approximate Orthogonal Simplification of a 3D IVIodel

The simplest way to simplify the 3D model is by constructing its axis-aligned
boundary box {AABB). However if the model is complex, the AABB will not
represent the model correctly. As shown in Fig 5.4, the original model in Fig 5.4a
cannot be represented adequately with its AABB in Fig 5.4b.
A better 3D model representation can be achieved by voxelizing the model,
Kaufman [1994]. A voxel representation of a model is a regular grid of cells,
in which each cell (voxel) contains a density value in the range of zero to one.
Normally a voxel-value of zero is representing a portion of unoccupied space and
a value of one is representing a voxel that is inside the model. The result of
the voxelization method can represent the original model much better than a
boundary box. However to get the best results, we need to have a large number
of voxels.
Another way to get a better 3D model representation is by representing the
model as a collection of cuboids. This approach was used by Zuurmond [2004]
to have a simple but quite representative model compared with the boundary
5.2 INTERFACE MODULE 91

Figure 5.4: Boundary Box

box approach. Basically, this approach is done by constructing the subdivision

boundary box (SDBB) from the model rather than the AABB. However, the
creat ion of cuboid models are done manually.
The method to construct the SDBB is almost similar with the voxelization
method. First, we divide the AABB boundary box into a grid of cells, but instead
of finding the density of each cell, we construct the simplified model in a diS'erent
way. The steps of this method are as follows:
Algorithm 1: Subdivision boundary box

1. Construct the AABB of the model.

2. Divide the AABB into n parts m each X, Y, and Z axis, resulting m a
uniform grid of cells (n x n x n).
3. For each cell, find the polygons of the model that lie mside or intersect with
thai cell.
4. Construct the AABB of the polygons tn step 3).
5. Clip the polygon AABB with the cell itself.
6. Loop to step 3) until all cells m the grid arc processed.

This process is illustrated in Fig. 5.5. From the original model, shown in the
upper left of Fig. 5.5, we construct the AABB. Then we apply the step 2 in the
algorithm SDBB with n equal 3. It divides the AABB to 3 parts in each axis
uniformly (cell sizes Xi = X2 = X^^Yi = Y2 = Y^; Zi = Z2 = Z^) and creates 27
uniform cells that can be seen in the lower left part of Fig. 5.5. Then steps 3-5
are applied to each cell. The final result is the simplified model that is shown in
the right most part of Fig. 5.5.
As shown in Fig. 5.5, the SDBB is able to represent the original model much
better than the AABB, while maintaining an acceptable number of cuboids.
92 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5 2

F i g u r e 5.5: Suhdivvnon Boundary Box

F i g u r e 5.6: OpLirmzed Subdivision Boundary Box

I
5.2 INTERFACE MODULE 93

Optimized Subdivision Boundary Box

Even though the subdivision boundary is more representative than the ordinary
axis-aligned boundary box, to get the best result, the size of the cells needs to be
smaller. A smaller size of the cells can be reached by dividing the AABB using a
larger value of n, at the expense of increasing numbers of cuboids in the simplified
model.
This fact is unfavorable, because our main objective is to get a simplified model
that only consists of a small number of cuboids. To overcome this problem, step
2 in SDBB algorithms needs to be modified. Instead of using a uniform division
cell size we use a non-uniform division cell size to divide the AABB. It means that
the cell sizes {Xi... X„), ( Y i . . . ¥„), {Zi ... Z„) are not necessarily equal.
Fig. 5.6 shows the comparison between the original model with the original
SDBB using the uniform division cell size, and the Optimized SDBB using the
non-uniform division cell size. As can be seen with proper division cell size, the
shape of the simplified model is closer to the original model and also contains a
smaller number of cuboids. This method will be called as the optimized subdivi-
sion boundary box method.
Moreover, because this method uses the boundary box for each cell that inter-
sects with the original model, it is guaranteed that the simplified model always
covers each whole triangle in the original model.
However, the output of this method is very sensitive to the variation of the
division cell size and to get the best result it needs to have the correct size of each
division. To find the best value manually is not trivial. Fortunately, the problem
to find the best division size can be solved using the heuristic optimization method.

Optimizing using Tribes-D

Particle Swarm Optimization (PSO) was introduced in Subsection 4.3.2 as one

of the heuristic optimization methods that can be used to solve optimization
problems in various fields. Also it was shown in Subsection 4.4 that a generic
PSO can be used to solve the shortest path problem.
The classical PSO is very dependent on parameter values. The values such as
the weight of velocity inertia and acceleration constants need to be tuned, and it
requires much time to find the optimal value. 'lYibes-D, on the other hand, is a
parameter-less particle swarm optimization algorithm. Basically we only need to
define the range of particles, and the maximum number of evaluations.
In principle, 'LYibes-D divides swarms in tribes. In the beginning, the swarm
consists of one tribe with one particle inside. At each time step, the best particle
of each tribe, called shaman, acts as an informer of other particles in that tribe.
For the shaman itself, the informer is selected at random among other shamans
in all tribes or in the archive.
Fr-om time to time, the quality of each tribe is checked. If the quality is bad,
a new particle is generated, and if it is good and has enough particles the worst
94 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.2

particle is removed. The quality of the swarm is also measured, and if it is bad
a new tribe is added, and if it is good and has enough tribes the worst tribe is
removed.
The Tnbes-D algorithm that is used was adapted from the C source code of
IHbes-D by Clerc [2008]. The difference concerns the random generator. Original
lYibes-D uses the KISS random number generator, while we use the Mersenne
Twister random generator.

Implementation of Approximates Orthogonal Simplification

We implement the optimized subdivision boundary box method and utilize 'iribes-
D to find the best division cell size. The main goal is to minimize the volume of
the simplified model while maintaining a low number of cuboids. Put in another
way, we would like to maximize the difference between the volume of the A ABB
of the original model and the sum over all volumes of each cuboid in the simplified
model. The fitness value is defined as follows;
( Cub \
JI = <^ VolBB -' Y.i'^olCuboid,) \ (5.1)

where Cub e l . , .n^, VolCuboid^ is the volume of each cuboid i, and VolBB
represents the volume of A ABB.
The lYibes-D algorithms is used to optimize the division cell size. In this case,
it optimizes {Xi... X„), (Yi . . . y„), ( Z i . . . Z„), with n the number of division, to
minimize the fitness functions fi. In our implementation, we choose to define n
equal to 3 or 5 depending on the actual size of the model and maximum fitness
evaluation is 10000 times.
The performance of the approximate orthogonal simplification method is meas-
ured by the ability to find the solution that visually represents the original model.
As a comparison, we compare the result of this method with the simplified
model using a simple voxelizalion method with two type of voxel-values, zero and
one.

Comparison with Voxelization Method

We have tried the approximate orthogonal simplification method to find the sim-
plified model of a pump model. To improve that results, we would like to capture
the quality variance in relation to the desired number of cuboids. It means that
the optimization problem becomes a multi-objective optimization, since it has to
optimize the subdivision interval and also the total of cuboids. Therefore, the
'lYibes-D was modified so it can be used to find the solution of a multi-objective
optimization problem. For this purposes, we introduce another objective function:

ƒ2 = {Cub} (5.2)
 5.2 INTERFACE MODULE 95

' ___C_^ — "1 . I ™ S. . JL "Z^ ~Z ZZ. _JZ I I "" ; iZZ

r—ijp-ijt-:;!.,.' -Zlz: np—z±—rr::;!::

.Tji: ii; zzfz—,—I zzz

F i g u r e 5.7: Pareio Pront

Table 5.1: Comparison, results proposed method and voxehzation

Cuboids Percentage
Method

I
number volume
Approximate orthogonal simphfication Method 10 42.58
Voxehzation 546 34.32
I
96 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5 2

F i g u r e 5.8: Comparisons
5.3 INTERFACE MODULE 97

Figure 5.9: Inlroduchon of The Interface Module m The Architecture

where Cub is the number of cuboid.

The result of multi-objective optimization with exactly 2 objectives can be
shown in a single graph called a pareto front graph. A pareto front is a set of
result that satisfy both objectives, where all member of that set are the optimal
solution for each combination of both objectives. Fig. 5.7 shows that lYibes-D
found the pareto front in almost all possible number of cuboids. As a comparison,
this pump model consists of 2768 triangles.
Furthermore, based on the pareto front, we can easily select the level of detail
of the simplified model. In Fig. 5.8 we can see the comparison of the original
model (5.8.a), the simplified model using the approximate orthogonal simplifica-
tion algorithm with the number of cuboids equal to 10 (5.8.b), and the simpfified
model using simple voxelization (5.8.c).
Concerning the shape of the simplified model, both methods are able to gen-
erate a representative simplified model. The summary is shown in Table 5.1. It
shows that using the simple voxelization method, we can easily get a more repres-
entative model. However, the number of cuboids that are generated by the simple
voxelization method are excessive for our purposes.

5.2.4 Summary of the Interface Module

In this section, we have defined four important parts of the interface module:
1. Constraint editor tool
2. Interface to CAD software packages
3. Smart P&T diagram tool
4. Model simplification tool
Fig. 5.9 shows the extension of the architecture with more detail Interface
Data part.
98 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.3

5.3 Pipe Router Module

The second part of the proposed methodology is the pipe router module which
uses the data from the interface module and calculates the route of all pipes. This
is the part of the system that actually performs the routing. In a simple routing
problem with a small number of pipes and simple environment, the pipe router
module is simply a path finder algorithm. However, our methodology is meant to
solve the pipe routing problem in a real design process, which means that it has
to route a large number of pipes in the complex environment, and also optimize
the solution.
As discussed in Section 3.3, pipes are routed according to the marine classi-
fication standard and the common knowledge of pipe routing. In Section 2.5.1
and 2.6, we have discussed the common knowledge and strategy to route pipes in
ship. In this section, it will be formulated in more detail in the form of routing
criteria, and will be used as the objective function of the pipe router module.
After the routing criteria are defined, the routing process can be performed by
finding the path for every pipe. In a simple situation, the path finding problem
can be solved by simply selecting one of the shortest path problem algorithms that
have been discussed in Chapter 4. However as described above, in the real ship
environment, this problem cannot be solved only by implementing one shortest
path algorithm. Therefore, we construct a pathfinder module that is capable to
handle the general pipe path finding problem.
Basically the pathfinder module contains several optimization algorithms, and
as discussed in Chapter 4, each algorithm has its own advantages and disadvant-
ages. Also, providing each optimization algorithm with the correct parameters is
very important to ensure the quality of its result. As discussed in Section 2.4.3,
the routing behavior of the pipe router module in the proposed methodology de-
pends on the area in which the pipes are placed. In the optimization problem, the
variation in the routing behavior can be achieved by using different optimization
parameters.
The main objective of the pathfinder module is to find the optimized path
for every pipe. However, since each pipe is routed and optimized one by one,
the combined solution might not be the optimized solution as a group. The
solution then might not be good enough because our main goal is to optimize the
combined solution. Therefore, an optimizer module is needed to ensure that the
group solution is optimized.
As described above, the pipe router module is no longer a simple shortest path
algorithm, but it become a complex module, as depicted in Fig. 5.10 that consists
of four main parts:
1. "Routing criteria" is the measurement of the validity and the quality of the
routed pipes.
2. "Pathfinder module" is the part ol the pipe router module that finds the
path of pipes.
3. "Optimizer module" evaluates the quality of pipe routes that are routed
5.3 PIPE ROUTER MODULE 99

Pipe Router Module

Figure 5.10: Introduction of The Pipe Router Module in The Architecture

by the single-router module to ensure that those routes are optimized as a

group.
4. "Router parameters" are the fixed parameters that are used by the single-
router module to perform its tasks. These parameter values depend on the
location where the pipes are routed.
These modules will be discussed in the next four sections.

5.3.1 Routing Criteria

According to their characteristic, the routing criteria to verify the quality of the
solution can be categorized as two part:
1. Pipe routing rules,
2. Pipe route performance.
The pipe routing rules consist of criteria that have to be fulfilled; otherwise
the solution cannot be used. If the result has passed this first criterion, then it
can be evaluated using the pipe route performance criterion.

Pipe Routing Rules

In Subsection 2.5.1, we have discussed the rules that must be followed by a pipe
engineer when he routes a pipe in a ship. Also the common strategy to route
pipes in a ship has been described in Subsection 2.6.
100 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.3

In Seclion 3.3, it is mentioned that tiie proposed metiiodology focuses on the

path finding problem and not on the division of pipes into spools. Therefore, part
of the rules and strategy that do not directly have any impact on the path finding
problem are neglected in this thesis.
The rest of the rules then can be categorized into two different groups; the
general rules that apply to all pipes, and the system rules that apply only for
pipes that belong to a certain pipe system. These two lists contains the rules that
are adopted by our methodology.

General Rules
The general rules are mentioned below:
1. Collision avoidance. All pipes that are routed must be collision free with
other pipes and with other objects in a ship.
2. Pipes must be supported. The path of all pipes that are routed must be
chosen in such a way that those pipes can be properly supported. In other
words, all pipes must be routed close enough to the major steel structural
elements and at least at every interval distance so it is possible to install
pipe supports on that pipe.
3. Parallelism. For ease of installation and maintenance, all pipe paths that
are close to one another and running in the same direction should be routed
in parallel.
4. Layering longitudinal and transverse direction. For much the same reason
as for the previous rule, if it is possible pipes are better routed in imaginary
pipe racks, and the imaginary racks that have a different direction should
be arranged into layers.

System Rules

The system rules should be implemented in such a way that they are easy to be
added or edited by a user. The system rules that are implemented:
1. Sea Cooling Water system must be placed as low as possible.
2. Bilge and Ballast system must be placed as low as possible.
3. Fresh Water system must not be routed through the oil tank.
4. Fuel Oil system must not be routed through the fresh water tank.
5. Fuel Oil system must be routed as far as possible from the electrical com-
ponent (i.e. switchgear) or combustion engine. Especially not above it.
6. Fuel Oil lYansfer system must slope down according to the flow direction.
7. Draught Measuring system must slope down according to the flow direction.
8. Lubrication Oil Bow thruster system must slope down according to the flow
direction.
9. Dirty Oil and Sludge system must slope down according to the flow direction.
10. Air, Filling, and Sounding system must be routed as straight as possible,
and the maximum bends should be limited to less than 30 degree angles.
5.3 PIPE ROUTER MODULE 101

Pipe Routing Performance

The rules that, are menlioned in the previous subsection must be followed. Failing
to comply with those rules results in a pipe that will be marked as a non-valid
solution, and it must be re-routed again. Thus, a valid solution means that all
pipes that are automatically routed comply with those pipe rules that are defined
above.
However, only trying to find a valid solution is not enough. The aim of using
the proposed methodology is to have a valid and optimal solution. To make it
easier to be measured, we adopt the condition that was used by Park [2002] by
measuring the cost of pipes. In his work, Park calculates the cost of pipes as the
sum of production cost (consists of the material and bending cost), installation
cost (related mainly with pipe support cost) and operational cost (related with
the location of the valves).
While adhering to the same principle, we use a slightly different way to cal-
culate the pipe cost. As mentioned in Chapter 3, our research is not focus on
the valve placement, therefore we neglect the operational cost. In Chapter 2, we
have mentioned that the cost of the pipe consists of three elements; pipe material,
production and installation cost. Starting from this point, the pipe material and
production costs are combined into a single category, the production cost.

Production Cost

Ftom the two elements mentioned above, the production cost is the most obvious
part. By calculating the part of each pipe completely, the cost figure can be
estimated. The raw material for a pipe consists of a straight pipe. Before it
can be installed in the ship, it must be processed according to the pipe spool
sketch. It needs to be cut, bent, flanged and if required painted or coated. In
total, the production cost of a pipe is the sum of raw material cost, the cost of
the production process, and the cost of pipe treatment.
The cost for the raw material of a pipe depends on:
1. Pipe material
2. Pipe length
3. Pipe diameter
4. Pipe thickness
5. Type of pipe bends
The cost to produce a pipe depends on:
1. Raw material of that pipe
2. Pipe surface treatment
3. Type of pipe bends
4. Type of pipe treatment or coating
102 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.3

Installation Cost

When wc discuss the inslallation cost of a pipe, it means that we discuss the pipe
spools of that pipe. Basically the installation cost depends on two aspects; the
first is whether the pipe is installed during the pre-outfitting stage or during the
outfitting stage. The latter costs are around 3 times higher than the first. Beside
that, the cost of installation of a pipe depends on the weight of that pipe, the
type of coating that the pipe has, the ease to support that pipe spool, and also
the shape of the pipe spool. For example, a simple pipe spool that only has one
bend but with a length of 2 meter for each legs, can be difficult to instaU.
Calculating the installation cost is not trivial, it is very sensitive to the labor
cost, labor skill, weather, equipment and how well the pipe spool is designed. The
version used for the analysis in this thesis includes a simplified approach, which
totally depends on the output generated by the router module.
Based on those assumptions, to calculate the installation cost can be done
by calculating the man-hours. In his book, Page (1999) describes in depth the
man-hours for piping production and installation, and it can easily be adapted to
the current situation.

Pipe Cost Value

In this research, we would like to combine the production and installation cost of
a pipe into a single pipe cost value. For the straight pipe, the total pipe cost can
be estimated to yield a single value. However, we need to differentiate the cost of
pipe bending depending on the type.
There are two different types of pipe bending; the bending that can be done by
a pipe bending machine, and the type involving the use of pipe elbows. In addition
to those two types, in practice there is another case when there are two bends in
a pipe and the distance between two bending points is too close to be bent by the
pipe bending machine. I'here are two solutions if this happens: using two elbows
or making the two bends using the pipe bending machine while allowing more
distance between the two bending points and then cut the excess pipe and weld
it back together. The second option requires additional pipe cutting and welding.
However, it is still cheaper than using two elbows.

5.3.2 Pathfinder Module

The pathfinder module is a part of the pipe router module that actually finds the
path of every pipe. I'his module routes all pipes one by one, and tries to optimize
each pipe by using the routing criteria as its objective functions.
In the routing process, this module uses pre-defined parameters from the router
parameters module, and follows the pipe order list from the optimizer module.
Basically, this module contains three separate parts; the blockage checker, the
single pipe router and the hybrid back-tracker.
5.3 PIPE ROUTER MODULE 103

Blockage Checker

As discussed before, besides the data that can be retrieved automaticaUy from the
CAD software packages, there are also some areas that must be defined manually
by the user of the methodology. Thus, there is a possibility that this areas lead
to blocking some pipes. If this situation occurs, the methodology may fail to find
a solution.
To prevent this, it is required that the system can perform the blockage test.
For this purpose, one of the shortest path algorithms that have been discussed
in Chapter 4 can be used. Since the main purpose of the blockage checker is to
test whether a pipe can be routed or not, it is not necessary to use the shortest
path algorithm that guarantees to find the shortest path, so long the algorithm
always finds a solution if it exists. Thus, the fastest computation time is the only
criterion that is considered.
In Section 4.4, we selected the shortest path algorithm from Mikami and Tabu-
chi [1968]. The Mikami-Tabuchi algorithm always finds a solution if it exists, re-
quires less memory than others, and is significantly faster than the A* or Dijkstra
algorithm.

Single Pipe Router

The single pipe router is the part that actually routes all pipes. The router get
a set of parameters from the routing parameter module according to the location
type of the pipe. During the routing process, the pipes are routed one by one
according to the list of pipes that are generated by the optimizer module. Then,
the result is evaluated using the routing criteria, and the router will send it back
to the optimizer module. Based on that, the optimizer module will optimize the
new pipe order to be used by the single pipe router. This iterative process will
continue until the globally optimized set of pipe paths is founded.
As discussed in Chapter 4, the shortest path algorithm that is suitable to be
used by the single pipe router in our proposed methodology is the A* algorithm.
There are four reasons for choosing this algorithm over the others; the A* al-
gorithm always finds a solution if it exists, the A* algorithm is faster and uses
less memory than Dijkstra, the A* algorithm most likely will find the shortest
path, and since it works in a grid environment, the A* algorithm can be used in
a weighted environment.
The basic approach to the single pipe router is as follows:
• unobstructed space is decomposed into discrete elements,
• the elements are treated as grids,
• the optimization algorithm walks through the possible path,
• an efficient and valid path through the graph is found,
• if a valid path of a pipe cannot be found, the hybrid back-tracker is invoked,
• the above steps will be repeated until all pipes have been routed or terminate
if there is no solution
104 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.3

Figure 5.11: Attraction area

Figure 5.12: Magnet area

Geometrical Constraints

Section 2.4.4 described four types of object constraints. That categorization is

based on the way a pipe engineer judges an object or an area when he routes a
pipe. However, that categorization is not sufficiently detailed to be a guidance
for a computer algorithm. Thus, it needs to be categorized into more detail, and
it will be referred as a geometrical constraint.
A geometrical constraint is a virtual area in a three dimensional space. In
this thesis, we always work in 3D environment, therefore the terms area means
a 3D volume. This virtual area represents a constraint value that will affect the
environment where it is located. In our methodology, there are eight basic types
of geometrical constraint:
• Routing point

Figure 5.13: Distraction area

5.3 PIPE ROUTER MODULE 105

Figure 5.14: Special type area

When a pipe is routed and a routing point exists, that pipe must be routed
through that point. A routing point can be added manually by the user or it
can be added automatically. For example, if the pipe that is routed contains
a fixed location valve, that pipe must be routed through the location of that
valve.
Obstruction area
Obstruction area is a virtual area where no pipes are allowed to be routed
through it. This area has an infinite constraint value. Every hard object
constraint, such as an equipment, is an obstruction area.
Sink area
This virtual area is a desirable area for a pipe to be routed in or through.
Thus, a pipe that is routed in this virtual area will get a bonus. For example,
the area near a major steel construction can be categorized as a sink area,
because it will be easier to install a pipe support.
R o u g h area
Contrary to a sink area, a rough area is a virtual area where a pipe will get
a penalty if it is routed there.
Attraction area
As shown in Fig. 5.11, an attraction area is the more complex form of a
sink area. The actual area is only the dark blue part, but there is also a
bonus effect to the surrounding area outside the attraction area itself. The
dark blue area has the highest bonus, and for the area further away from it,
the bonus is smaller.
Magnet area
Magnet area, as shown in Fig. 5.12, is almost the same as an attraction
area. The difference is that the actual area is an obstruction. Thus, area
surrounding the obstruction is preferable, but the actual area itself is a
forbidden area. As such the magnet area is a combination of obstruction
106 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.3

Figure 5.15: Without attraction area

area and attraction area.

• Distraction area
In contrast to an attraction area, a distraction area gives a penalty to the
pipe that is routed through this area and the surrounding area. Fig. 5.13
shows that the actual attraction area is red, which represent the highest
penalty area, and the penalty will be smaller the farther the location is
away from it.
• Special t y p e
Beside the basic geometrical constraint above, a combination of two or more
basic constraints is also possible. For example, the left figure in Fig. 5.14
shows a NoGo area and the right figure shows a stair area from side view.
The center of a NoGo area is an obstruction area, but the edge of it is a
rough area. The construction part of a stair is an obstruction area, and the
top part is a rough area.
Fig. 5.15 and 5.16 shows an example how the attraction area affects the search
process. As shown in Fig. 5.15, for the environment without an attraction area,
the search algorithm needs to explore a larger volume than if there is an attraction
area in the floor.

Geometrical Constraint Parameters

Each type of geometrical constraint has different influence directions. Some of
them influence all 3 directions X, Y and Z, while others only influence 2 directions
depending on its orientation.
All of them have a maximum distance of influence, called the distance para-
meter. They also have min and max parameters to set the minimum and max-
5.3 PIPE ROUTER MODULE 107

Figure 5.16: With attraction area m the floor

imum weight value depending on the distance.

Some of the geometrical constraints have a linear weight dependence on the
distance, while others are non linear.
In case two or more areas of geometrical constraint have an intersection, dif-
ferent rules are applied depending on the types. In some combination, the values
resulting from the constraints are multiplied. There is also a dominant geomet-
rical constraint which if its influenced area intersects with another's, only the
value from that constraint is used.

Grid Decomposition
The unobstructed space is decomposed into grid of cells, with the size of each cell
5 mm by 5 mm by 5 mm. This small cell size is used to improve the quality of
the routed pipes. However, there is an important disadvantage by using a small
cell size in the grid. As discussed in Chapter 4, the memory that is needed to
store a grid of cells is very large, especially in a large three dimensional space
when the cell size is small. Fortunately, in most of the cases, the A* algorithm
does not need to explore all cells in the space to find the solution. Based on this
knowledge, the cells are created during the routing process.
Every cell has a weight value. The weight value of a cell is the penalty or bonus
factor that will be used to calculate the objective cost value of the pipe that is
routed through that cell. When a cell is created, a weight value will be assigned
to that cell based on its location in the three dimensional space. As previously
described, each location in the space is affected by the constraint value of the
geometrical constraints that lie in that location. Therefore, the cell's weight value
depends on the geometrical constraints that lie on or near by that cell. If there
108 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.3

llllWIi 1 111i 111111M M 11111 1

lBHHHaHlBnia«HH.HM«Éay

Obstacle Obstacle

a.
Figure 5.17: Collision detection is needed

are more than one geometrical constraints, the weight value is the combination of
those geometrical constraints. The objective cost value is a multiplication of the
pipe cost value with the cell's weight value.
The geometrical constraints that are automatically created by the rules con-
straints are based on the pipe that is currently being routed. Thus, the constraints
might be different between pipes that belong to different groups to which different
rules apply.

Reduce Collision Detection Usage

After the grid of cells is built, the A* algorithm can be used to find the shortest
path. As described in Chapter 4, the A* algorithm finds the optimized path by
exploring the grid, cell by cell. However, since the cell size is smaller than the
diameter of the pipe that is being routed, the path that was found by the A*
algorithm might not be valid, because the real pipe might have a collision with
an obstacle. For example. Fig. 5.17.a shows the path that was found by the
A* algorithm. However, as can be seen in Fig. 5.17.b, the solution is not vaHd
because the actual pipe has a collision with the obstacle, as depicted in red.
To prevent this problem, the collision detection routine must be triggered
every time the A* algorithm explores a cell. This solution however, has a big
disadvantage because a collision detection routine is computationally expensive,
and the total computation time that is needed to find the optimized path becomes
longer.
Fortunately, this situation can be avoided by manipulating the obstacle before
the process to build the grid of cells is started. Fig. 5.18 shows how it works. In
Fig. 5.18.a, as shown in pink color, before the cells are created the obstacle was
virtually extended. Then, when the grid is built, that virtual area will not be
used as a cell. Therefore, the optimized path that was found by the A* algorithm
is always collision free. It is important to ensure that the extended area has the
proper size; enough to prevent the collision but not too large. Thus, the expansion
size must be adjusted according to the size of the pipe that is being routed.
5.3 PIPE ROUTER MODULE 109

m ^^|iWffl

Obstacle Obstacle

a. b.

Figure 5.18: Collision detection is not needed

~ " ~

1 1 i _ - u LJJ J..1..m

-
1
(3 3£>ti3C \i J
1
Figure 5.19: Backtracking

Hybrid Backtracking

In the pathfinder module, a set of pipes is routed one by one according to the
order of routing that was prepared by the optimizer module. Because of that,
there is a possibility that the path of one pipe is bloclced by the other pipes that
were previously routed. This situation is described by Zhu and Latombe [1991]
in their work, and they suggest a solution by performing a backtraclcing process.
The backtracking procedure is shown in Fig. 5.19. As can be seen in Fig.
5.19.a, there is no way to route the red pipe because it is blocked by the green
pipe. In this case, those two pipes will be re-routed to find the solution as shown
in Fig. 5.19.b.
Using the backtracking procedure will solve most of the pipe blocking prob-
lems. However, this procedure cannot solve the mutually intervening case that
was mentioned in Subsection 4.5.2. To solve this special case by utilizing the
deterministic procedure is possible. However, since it is very hard to predict the
environment of a three dimensional space that causes the mutually intervening
case, a robust deterministic procedure will be very complex and re-routing of
many pipes might be required. Thus, we propose to solve this problem by using
one of the heuristic optimization methods that have been discussed in Chapter 4,
110 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.3

the particle swarm optimization method (PSO).

Since PSO and genetic algorithm performances are almost similar in our test
case comparison, PSO was chosen over a genetic algorithm solely because we have
more experience to use it and to tweak the parameters of PSO.
To solve the mutually intervening problem, PSO is utilized to find the shortest
path for both pipes at the same time. The implementation is the same as what
we used to compare the algorithms in Chapter 4. The difference is because it tries
to optimize two pipes at the same time, the population is divided into two parts,
and each part represents the turning points of each pipe.
Using this method, the mutually intervening case of two pipes that are blocking
each other can be solved. Then, the question arises, why did we not implement
this method to find the paths for all pipes at the same time? Even though the
idea to simply use the PSO to route a group of pipes is very tempting, the answer
is short and clear; it is not possible. There are two reasons; most of the times it
cannot find any solution because of the complexity, and if the solution were found,
it is not guaranteed to be optimal. Moreover it requires a lot of computation time
to find it. However, this method is suitable to solve the shortest path problem
for up to 2 or 3 pipes at the same time.
Thus, we use the backtracking procedure from Zhu and Latombe [1991], and
if the problem is still there, it will be solved by the parallel routing using PSO.
This complete procedure is called the hybrid backtracking procedure.

5.3.3 Quality Measurement

Subsection 5.3.1 discussed the criteria of pipe routing that must be complied with
for every routed pipe. It was also mentioned that the pipe must be routed as
efficient as possible in a way that the pipe cost value is as low as possible.
However, the quality of pipes not only depends on the minimum cost, but we
must also consider the non quantitative aspects such as the possibility to install
a support and the parallelism of the pipes.
To measure a non quantitative aspect is not trivial. Therefore, we need to
convert that aspect into a quantitative value. In our methodology, we do this
by giving a different weight value for each grid cell. In Subsection 5.3.2 we have
described the various types of geometric constraints that affect the grid cell's
coefficient weight. Also we have explained that the geometric constraints can
be defined manually but also automatically created. For example, the magnet
constraints are created for every plate and stiffener to attract pipes to be routed
nearby the steel construction. The special type constraints are automatically
created for the pipe that had been routed, so the next pipes are attracted to it
and run as parallel to it as possible.
The quality of the routed pipes then can be calculated using the combination
between the pipe cost value and the weight of grid cells. During the optimization
process, this function is used as the objective function.
5.4 IMPLEMENTATION UI

5.3.4 Optimizing tlie Solution

As implied by its name, the optimizer module optimizes the solution oi the pipe
router module. After the pathfinder module finishes its task, the quality of the
result is evaluated using the routing criteria by the optimizer module. Based on
that result the next pipe order is defined by the optimizer module and sent back
to the router module. This cycle continues until a result with a defined quality is
found.
At the core of this module is the Discrete Particle Swarm Optimization. Dis-
crete PSO is chosen due to its speed of convergence as well as its performance. The
Discrete PSO version that is used here is based on the method developed by Clerc
[2004] with some adjustment on the objective function and problem formulation.
This method has been previously discussed in Section 4.6.
'J'he order of pipes as created by the optimizer module is not based on the
optimization of all pipes as one group. Instead, during the first rough routing
stage, some pipes are grouped together based on their location. Then, the op-
timizer module optimizes the order of pipes also by taking this grouping into the
consideration.
The optimizer module also optimizes the height difference parameter. This is
the parameter that arrange pipes that are routed in the same direction to have
the same height.

5.3.5 Routing Parameters Behavior

Pipe routing in a ship is hard. However there are some behavior patterns that can
be identified. An example is that in a certain area a group of pipes always runs
in the same direction, so lor this area pipes normally are arranged in parallel.
In Subsection 2.4.3 we have discussed three difl^'erent types of working area
from the pipe engineer point of view.
1. Machinery type area
2. Accommodation type area
3. Technical type area
This categorization is adopted in our proposed methodology by creating three
different parameter sets to be used by the pathfinder module.

5.4 Implementation
Fig. 5.20 shows the expanded version of the outline methodology architecture
that was shown in Fig. 3.1.
The proposed methodology has been implemented for research testing and
validation purposes as a software package. The current version of the proposed
methodology is implemented using C # programming language tools, and it can
be operated under Microsoft Windows operating system using Microsoft .NET 4,
and also under Linux operating system using Mono. However the interface part of
112 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.4

Pipe Router Module

Optimizer lUlodule

Discrete
Particle Swarm
Optimization

^T—T
Pathfinder Module Routing Criteria
Router
Parameters
Single pipe router Pipe rules

Hybrid
Blockage checker Pipe cost
Backtracker

v.-
Required Data
Interface Module

Smart P&l Constraints Ulodel

Model I
I Dii
Diagram Tools Editor Tools }|jfJcatlon
Simplification I

Interface to CAD Packages

X
Figure 5.20: The Complete Architecture of Proposed Methodology

the proposed methodology must be operated tinder Microsoft Windows because

it has to communicate with the existing CAD software which mainly works under
Windows.
One of the design requirements states that the proposed methodology has to be
expandable. Because of that the modularity concept is used. This is implemented
by dividing the proposed methodology into three main parts, the core of which
contains the pipe router module and the optimizer module. The interface contains
the interfacing between the proposed methodology and existing CAD software,
and also the P&I diagram tools. Finally, the third part contains the knowledge
rules that represent the criteria that are used as guidance for the core part.
A mySql database engine is used as the database engine of the proposed meth-
odology. The main reasons for using it are its very good performance and its
freeware nature. One other practical reason is the small size and easy installation
of mySql.
Fig. 5.21 shows the flow inside the pipe router module. It starts by reading the
data from the interface module and than performs the blockage checker to ensure
that the initial environment does not block any pipe. The blockage checker itself
has been discussed in Subsection 5.3.2. If a pipe is blocked, it gives a notification
5.4 IMPLEMENTATION 113

Pipe Router Module

Optimizer Module Pathfinder Module

Required data
1
[ Blockage Checker
~"S—
notify user

[blockage]

[no biclckage]

Calculate
parametors T<J-

Roullna Criteria
pipe rules
Slngle pipe router
pipe cost <

±
c Evaluate
solution
> [solution founded] \ / [path founded]
[path becked]

[defined qui lity achived]

[solution not founded] [blockage cleared]

Figure 5.21: Interconnection inside Pipe Router Module

to the user.
Than, using the standard parameters, the single pipe router module starts to
route pipes one at a time until all pipes are routed. If there is a pipe that cannot
be routed, the blockage checker is triggered again to find the routed pipe that
blocks the current pipe. Based on that, the hybrid backtracker tries to solve the
blockage problem. The hybrid backtracker was discussed in Subsection 5.3.2.
114 THE METHODOLOGY ARCHITECTURE AND ITS IMPLEMENTATION 5.5

If the problem is solved, the pathfinder continues to find the path for the next
pipe, and if it cannot be solved, the pathfinder module stops the iteration and
gives the signal to the optimizer module that the combination of the parameters
is not good and that a nevir set of parameters is needed to begin with.
If all pipes are routed by the pathfinder module, the solution is sent to be
evaluated by the optimizer module. Then the new set of parameters is calculated
and sent back to pathfinder module.
This cycle runs until the defined quality is achieved, or until the time fiag
limit is invoked. During the process, the pipe rules and the pipe cost which form
the routing criteria are used by both the single pipe router and the evaluation
solution.

Parameters
The predefined parameters used in the algorithm are:
• Geometrical constraints parameter
• Single pipe router algorithm parameter
• Branch solver parameter
• Grid size
• Hybrid backtracking parameter
The parameters that are defined by the optimizer module consist of
• Routing order of pipes
• Height difference

5.5 Summary
This chapter describes our proposed methodology in detail which also answers the
second, fourth, fifth and sixth research questions of this thesis. In this section we
summarize it based on the addressed questions.
The second research question asks for the needed information, the source, and
how to get it. In Chapter 2, we have described the list of information that are
needed to perform the pipe routing process. In Subsection 5.2.4 we have shown
that most of the information can be retrieved from the CAD software, and that
we also need additional tools such as the smart P&I diagram tools.
The fourth research question asks which common knowledge of the pipe rout-
ing process should be adopted in our proposed methodology, and to what extent
we should adopt it. Subsection 5.3.1 answered this question by translating the
common knowledge into the routing criteria.
Subsection 5.3.3 answers the fifth research question that concerns the way to
evaluate the quality of the routed pipes quantitatively. In this subsection, we
have explained how the translation from the non quantitative to the quantitative
value was made and that our methodology uses this as the objective function.
5.5 SUMMARY 115

In this chapter Ihe sixth research question, relating to how to implement the
routing algorithm efficiently was answered as well.
The common bottleneck for the multiple path finder is that one path is bIocl<-
ing another. If this problem can be detected as early as possible, the time wasted
can be reduced. Therefore, one of the key functions in our methodology is the
blockage checker.
During the search in a grid space with the cell's size smaller than the pipe
diameter itself, we need to perform a collision detection test for each step. By
doing this, the time that is needed to find the shortest path is very long. To
make it faster, we found a way to avoid performing the collision detection test.
Subsection 5.3.2 explained how our methodology deals with this.
Another important detail in our methodology to reduce the computation time
of the pipe routing process is the new 3D model simplification method that was
explained in Subsection 5.2.3.
With all this we have a functional method and methodology which we sub-
sequently test and validate in a complex situation taken from a real ship design
case. This is the subject of the next chapter.
Chapter 6
Pipe Routing l\/lethodology
Validation

6.1 Introduction
In the previous chapters, we have described the needs to automate the pipe routing
process and its practical aspects. Then as a means to both investigate and prove
our proposed methodology, the automatic router software package was developed
as a laboratory.
In order to validate the proposed methodology, we need to use the validation
method from Seepersad et al. [2006] called the validation square. This valida-
tion technique is chosen because the traditional approach of formal, rigorous and
quantifiable validation is problematic for the proposed methodology.
The essence of the validation square is covered by two primary tasks:
• Structural validation of the proposed methodology
• Performance validation of the proposed methodology

6.2 Structural Validation

As explained by Seepersad et al. [2006], there are three complementary facets in
structural validity:
1. Internal consistency of each parent construct
2. Internal consistency of the method
3. Appropriateness of the example problem
In the next subsections each facet is described and validated.

117
118 PIPE ROUTING METHODOLOGY VALIDATION 6.3

6.2.1 Internal Consistency of each Parent Construct

There are Iwo parts that must be shown to validate this facet. The first one is
that the requirements for the outcome and the process. The second one is to
estabUsh the internal consistency proposed methodology in its entirely.
In Chapter 2, we have described the criteria for pipe routing that contain the
requirements for both the outcome and the process. In Chapter 5, we have shown
the complete architecture that translates the pipe routing methodology into a set
of methods that subscribe to the same requirements. The interviews with pipe
routing experts, the discussions with them on the process and also the problem
analysis show that the requirements are determined.
In Chapter 5, we have also shown how the methodology architecture was built,
and we explained every part that was used to build it. The reasons of choosing
those parts were also described in that chapter, and it was shown that each part
was selected such that the part requirements were met. With this, the internal
consistency of each part of the proposed methodology is satisfied.

6.2.2 Internal Consistency of The Method

To show that the proposed methodology satisfies the internal consistency, we chose
to use the flowchart method. By using the flowchart in Fig. 5.21 in Chapter 5, it
was shown that all input required for one of the methods is always available from
a previously called method.

6.2.3 Appropriateness of The Example Problem

In order to verify the performance oi our proposed methodology, we have imple-
mented it into the automatic router software package. Then, we chose to use
this implementation to solve the pipe routing problem in a machinery room in a
complex ship.
Even though there is no formal statement claiming that the machinery room
is the most difficult one, it can be easily established from literature that the
characteristic ol the pipe routing problem in a machinery room in a ship generally
considered to be the most challenging one. Indeed, it has been the main target
of many other research to improve the pipe routing process. Based on that, we
conclude that the selected example problem is appropriate.

6.3 Performance Validation

As for the structural validation, the performance validation also consists of l,hree
facets:
1. The usefulness of the method for the chosen example problem
2. 'I'he demonstrated usefulness being linked to applying the methodology
3. The methodology being useful for domains that are broader
6.3 PERFORMANCE VALIDATION 119

6.3.1 The Usefulness of The Method for The Chosen Example Problem
To establish the usefulness of the methodology, it should be applied to a repres-
entative example problem. As explained above, the proposed methodology was
applied to solve the pipe routing problem in a machinery room in a complex ship.
The first step in establishing the usefulness is to determine how the result
for our example problem should be measured; what aspects are needed to be
examined and compared.
As discussed in the previous chapters, many researchers have made a lot of
effort to tackle the pipe routing problem. In order to prove that their algorithm
works, they use their own criteria to measure it. Since most of them merely focus
on developing the routing algorithm itself, the validation usually is performed
by showing that the algorithm has the capability to route pipes in the defined
environment, and then they claim that the algorithm is validated. Some others
include more detail in their validation criteria, by not only minimizing the length
and number of bends, but also considering the branches. Ultimately, some of
them compare the pipes that are routed by their routing algorithm with the pipes
that are routed manually. They not only compare the total cost (using total
meter-inch and number of bends), but also compare the path of those pipes in
the space. The main argument is that since the automatic pipe router algorithm
is created by mimicking how a person routes the pipes, both results are expected
to have the same path, or at least almost the same.
In our research, beside minimizing the total cost, we have tried to include
most of the practical aspects, for example, all pipes must be routed according to
the marine standard. There was a point where we were also tempted to compare
the path of the pipes that are automatically routed with the one that arc routed
by a pipe engineer. However, we decided only to make a comparison of the cost
to show that the quality of the automatic pipe routing is on the same level witli
the pipes that are routed manually in terms of cost.
The main reason why we choose not to compare the path of the pipes is because
we believe that the argument above was not completely correct. That argument
is true if we only route one pipe from one nozzle to another. Then the path of
the pipe most likely is the same for both automatic and manual routing. This is
because both methods will route the pipe as efficiently as possible. However, there
are many pipes that need to be routed in a ship, and it is impossible to route all
pipes if we only consider minimizing the cost of the pipe that is currently being
routed. Every pipe engineer must consider that all pipes are routed efliciently as
a group of pipes. In practice, there are no exact rules to tackle this condition.
Thus, every pipe engineer must think to solve this problem individually. Then
as a result, the combination of pipes that are routed by a different pipe engineer
might not be the same.
In short, it is a fact that there is no standard of how a combination of pipes
must be routed in a ship. 'Thus, it is no longer interesting to compare every single
path of the pipes that are routed automatically and manually.
120 PIPE ROUTING METHODOLOGY VALIDATION 6.3

However, in other aspects, we still need to make a comparison between the

result of automatic and manual routing. For example, we still need to compare
the total cost of the pipes, but merely as a measurement that the total cost of
the pipes that are automatically routed are acceptable. Beside that, in a certain
difficult area, for example in a very small area that contains many pipes in it, we
also would like to show that the automatic algorithm is capable to route pipes
in that area, and compare it with the way a pipe engineer handles that kind of
situation.
Moreover, rather than focusing on comparing the automatic result with the
manual result, we are more interested to show that the proposed methodology has
the capability to solve the pipe routing problem in a ship. There are four aspects
that are used as measurement criteria:
1. The capability of the methodology to route pipes in a difficult area of a ship,
i.e. in a machinery room.
2. The routed pipes must comply with the marine standard.
3. The routing algorithm must produce high quality pipe paths, with regard
to the practical aspects. Especially the parallelism and the possibility to
install pipe supports for every pipe.
4. Show that the total cost of the routed pipes are reasonably optimized.

Test Case: The Machinery Room

In order to establish the usefulness of the proposed methodology, we have imple-
mented it into the automatic pipe router tools package. This package includes the
simple smart P&I Diagram tools, library tools, and the automatic router tools.
The automatic pipe router tools package is used to route the pipes in the most
difficult area of a ship. We choose to use it in an area that is considered to be a
difficult area by a pipe engineer, rather than to route all pipes in a ship. 'I'his is
simply because in other areas, the manual routing can be done conveniently using
3D CAD software.
To satisfy the usefulness of our proposed methodology in the machinery room,
we need to prove that the measurement criteria are satisfied. However, it is
not trivial to measure every criterion from the test case. For example, to show
criterion number four that the cost of the pipes that are routed automatically
is reasonable, we have to compare many pipes, but to prove the measurement
criterion number three, we have to investigate a specific group of pipes.
As the test case, we test our proposed methodology to route pipes in the
machinery room in a ship. In this case, all pipes in this part of the compartment
will be routed automatically. By proving that the proposed methodology capable
to route all pipes in this compartment, we prove that (part of) the first criterion
is satisfied.
Then the pipe cost of the results of both manual and automatic routing are
compared, by only considering the production and installation cost as described
in Subsection 5.3.1. By showing that the total cost similar to the value when the
6.3 PERFORMANCE VALIDATION 121

Figure 6.1: Machinery room

pipes are routed manually, we validate criterion number four.

The automatic routing process was performed in a HP Z400 computer, with
Intel Xeon CPU @ 3.2GHz, and 24GB RAM. The operating system is Windows
7 x64. This computer is also used as the database server using mySQL 5.5 which
consumed 8GB of RAM. Therefore, the computer memory that can be used by
the automatic pipe routing was limited to 12GB.
The machinery room is the most common test case in automatic pipe routing
research area. The main reason is because the machinery room is considered to
be the most difficult area for a pipe engineer.
As our test case, we use the machinery room of a real vessel. Fig. 6.1 shows
the original situation that was imported from CAD software.

Model Preparation
As mentioned in the previous chapter, our methodology uses the simplified 3D
model from the CAD software. The simplification of the 3D model is performed
during the 3D model import process.
There are two parts in the simplification process; for the plates in the steel con-
struction and for other parts. Steel plates are simplified by settings the thickness
122 PIPE ROUTING METHODOLOGY VALIDATION 6.3

Table 6.1: Calculation time

Run Time (Minutes) Remarks

1 3233
2 3015
3 5760 stopped
4 2683
5 2658 memory overload
6 4373
7 3554
8 2745
9 3111
10 2890

to zero. For the other parts, they are simphfied into a small number of boundary
boxes using the simplification method that was discussed in Subsection 5.2.3.
During the simplification process, the basic geometry constraints are created
automatically, i.e. the attraction constraint areas are also created in the steel
construction part. The intention of this is to attract a pipe to choose a path
nearby the steel construction for the ease of installing the pipe support.
There are some manual setups that need to be done. One of them is to place
some valves that are required to be arranged as a group in a certain location.

Discussion of the result

We have run the implementation of the methodology to route the pipes in the
machinery room 10 times, and the best result is shown in Fig. 6.2. This is
achieved with 5 mm cell size, which means that the center pipe is routed in a 5
mm grid space. For example, the automatically routed pipe path can only lies in
coordinates that are a multiple of 5 (8255, -555, 645), and cannot be in coordinate
(8257, -557, 643).
The smaller the grid space size is, the better the chances of the methodology
to find the solution. However, it needs more memory and longer computation
time.
As shown in Fig. 5.21 in Chapter 5, the pipe router module contains some
heuristic parts, in the optimizer module and hybrid backtracking part. Therefore,
on each run we expect to get a different solution.
Table 6.1 shows the calculation times that were needed by the methodology.
As we can see, in run # 3 and run # 5 the methodology failed to find the solution.
In run # 3 , the methodology was stopped because it still did not find the complete
solution after 4 days. The limitation of 4 days was chosen based on the previous
two runs that found the solution in about 2 days time. However, without the
limitation, the methodology might also able to find the complete solution in run
6.3 PERFORMANCE VALIDATION 123

Figure 6.2: Automatically routed

#3.
In run # 5 , the methodology crashed due to memory overload. There are some
aspects that might generate this problem. There is a possibility that another
application running on the computer was causing this. However, since it is only
a research implementation, at this moment the automatic pipe routing software
package is not yet optimized for real production in terms of robustness, such
as memory management and prevention of crashes caused by the mistake of the
operator. Here we only focus on the algorithm to solve the pipe routing problem.
If we conclude that the methodology failed on run # 3 and # 5 , the success
ratio is 80% which can be considered as acceptable at this stage. Therefore, the 5
mm grid space is good enough for our present purpose. Considering only the runs
that were able to find the complete solution, the average time needed to route
144 pipes in the engine room is 3200 minutes.
At this moment our focus in not to minimize the calculation time. But it is
important to show that the proposed methodology is able to automatically route
many pipes within an acceptable time, and 3200 minutes to route 144 pipes in an
engine room can be claimed to be an acceptable result.
To shorten the calculation time and make the success ratio higher, we need
to focus on the improvement of the software package, and keep the proposed
methodology intact. However, this is not our main focus because it is a software
engineering problem.
Also, with the rate of hardware improvement in the past years, we can safely
make an assumption that in the near future the increasing computational power
of the computer will make the calculation time shorter. The fast growth in com-
124 PIPE ROUTING METHODOLOGY VALIDATION 6.3

Figure 6.3: Manually routed

putation power is one of the main motivation to route pipes automatically; the
time that is needed to do the manual routing is basically almost the same from
year to year, but the computation time to perform automatic pipe routing will
grow shorter and shorter.
Our main focus was to translate the way a pipe engineer routes the pipes in
a very difficult area of the ship into the procedural methods that can be done by
the computer. This is very interesting and challenging and until today, there is a
general assumption that only experienced pipe engineers are able to route pipes
properly in the engine room.
Fig. 6.3 shows the pipes in the machinery room that are routed manually. If
we compare it with Fig. 6.2, even though there are important similarities, the
differences between those two results can be noticed easily. As we mentioned
earlier, those differences are expected, since there are many ways to route the
group of pipes.
One of the reasons for the differences is that the routing algorithm in the
proposed methodology is more strict with respect to parallelism, while in practice,
pipes might not have to be routed in parallel that strictly. This can be seen in
Fig. 6.4 and Fig. 6.5.
Beside that as shown in Fig. 6.6, to ensure the minimum number of bends,
the automatic routing algorithm is strict with respect to maintaining different
heights for X and Y direction, while in manual routing, shown in Fig. 6.7, a pipe
engineer is more flexible in this respect.
It is interesting to compare the total cost to make sure that the result of
our proposed methodology is on par with the manual routing result. Table 6.2
shows the comparison of the pipe cost per system. The cost value in Table 6.2 is
6.3 PERFORMANCE VALIDATION 125

F i g u r e 6.4: Parallehzation m automatic routing

F i g u r e 6.5: Parallehzation m manual routing

126 PIPE ROUTING METHODOLOGY VALIDATION 6.3

Figure 6.6: Mamtaining difference heights for different direction

Figure 6.7: Difference m heights more flexible for manual routing

 6.3 PERFORMANCE VALIDATION 127

> CU <5 CO o o CD 00 1--

t- CO
C3>
CO
O
T-H
lO CO T-H CO 00 o lO
ir- CM CM lO
00
CM
O
CD CM lO
o
CO CO '^ T-H
T-H
T-H
o o
^1 00 CD
CO
O
i-O
1---
T-H
CO T-H
CM CM
1>-
co t-H
CO CM T-H CO r-H T-H rH

00
CM CJ2 lO CD 00 lO CM
.a 3 CM
CO
CD
T-H
T-H CM 00 O T-H
CJ5 t--
00 o
O T—t CO T-H O CO co O LO
CM
CO
00
o CO
CO O 00 lO CO
CO CO
CO no
PL, CO CO r—1 T-H CM T-H CO T-H
T-H CO
oo T-H o
CO

t- <M
(TO
00 CM T-H
T-H CO
CM lO 00 o CD Oi
00 CD T-H T-H T-H ^H CJ3 tr- o o 00 CM
CO O CM
CM
T-H
CO in 00 o io lO T-H
CM
^H
CM O 00
I—t T-H T-H 00 QO
in
r-H CO CO in

u
CM T-H CO 00 O o CO 00 05 o
0) 1=1
CM
CO CM
CD
00
CO
lO CO
en
00
CO
00
ira
05
CD o
O
T-H
T-H
CO
'^
o o 05
^H T-H o 00 lO
CO T-H
CO CO
CO
S3,
Number of
pipes

o CO
00
-* l O tr- CO T-H
CM CO T-H T-H CM T-H CO l O 00
t-H T-H T-H T-H T-H i-H
rH IN
CO

PI 1 Ö
PH

<; fa
Ö
1
B
'i a;
PH CD
a.
a. o >
^ T

,<D
^4H
CP
PH
CD
CD
C^

'>
t4H
a;
!H

<
"(3
cn

1 I cn s fa£
CO
o pi bX)

i 1
Ö SH
(D CO CD
03
CD Ö Ö
'S +H d CO M bO
03 *+^
m O
CJ
p Oo .3 O >>
CD p-1

o
03

X fa fa o
•Q Ü O
o +j
-l-J
Cfi 1
< CD
w Ü
Ü '3
03
Ü p ,JPI CO

1
fa
128 PIPE ROUTING METHODOLOGY VALIDATION 6.3

calculated based on the pipe cost value that was used as part of the objective cost
function as described in Chapter 5. It must be noted that the pipe cost value is
not the same as the objective cost value that was used during the optimization
process. The pipe cost value is a constant value per pipe in every location, while
the objective cost value varies depending on the cell location, and each cell has a
weight value that depends on the geometric constraints.
Also, since our methodology does not split pipes into spools, we exclude the
cost of the pipe flanges, with the assumption that both manually and automatic-
ally routed pipes have almost the same number of flanges.
As shown in Table 6.2, the total meter inch of the automatically routed pipes
is around 5% higher than the pipes that are routed manually. This result is
reasonable because the current routing algorithm uses 90 degree of bending angle
as its first choice. It only will choose other bending angles for 3 reasons; the
distance between two bending points is shorter than the minimum requirement,
the pipe is a member of a pipe system that requires the pipe path to have a slope
and depends on the predefined bending parameter, to select a cheaper bend type.
The total estimated cost in automatic routing is around 1.7% lower than the
total cost of manual routing. This is also a reasonable result since the routing
algorithm intends to route pipes as parallel as possible. Also the consistency of
maintaining constant height for X and Y direction reduces the number of bends
that are needed.
Since we only compare the estimated cost of the pipes, we are not claiming
that the automatically routed pipes are definitively cheaper than the manually
routed pipes. Nevertheless, from this comparison, we can claim that the solution
obtamed by the proposed methodology is quite good and that the validity of the
cost criterion is established.
In a previous subsection we mentioned that for operational reasons, some of
the valves must be located nearby in an accessible area. Fig. 6.8 and Fig. 6.9 show
that the proposed algorithm is also capable to route pipes with this condition.
Since all criteria are met, it proves that the proposed methodology is able to
route pipes in difficult areas of a ship. Thereby we demonstrated its usefulness.

6.3.2 The Demonstrated Usefulness is Linked to Applying The Methodo-

logy
During our attempt to demonstrate the usefulness of the methodology in the
previous subsection, we applied it and as a corollary did also prove that the
usefulness is linked to the utilization of the proposed methodology.

6.3.3 The Methodology is Useful for Domains that are Broader

In order to identiiy what can be categorized as broader domains relevance to our
research we need to look back to its main objectives. These were to identify the
expertise of pipe engineers and translate it into a set of procedures to route pipes
6.3 PERFORMANCE VALIDATION 129

1 :iJ^
Wi ^ ^ ^ ^ ^ ^ ^ ^ ^ .
ii\'^^Bi ^_a

1 i Figure 6.8: Predefined valves group

F i g u r e 6.9: Predefined valves group

130 PIPE ROUTING METHODOLOGY VALIDATION 6.4

in a ship. From that, it was clear that our broader domain is the pipe routing
problem in an arbitrary ship.
As mentioned above, we took the machinery room as the example problem.
It was also stated that the machinery room is the most difficult part of the ship.
Based on this, it can easily be deduced that if the methodology is able to solve
the pipe routing problem in that room, it must be able to be used in other areas
in a ship also.
Indeed, our research addresses pipe routing. There are also other distribution
systems that need to be routed: ducts, cables and walkways for example. While
these distribution channels obey different rules, our methodology can most likely
be tailored to these situations by implementing these rules. Then the relevant
domain for our research is indeed considerably broader.

6.4 Sensitivity Analysis

Sensitivity analysis is the study of how the uncertainty in the output of a math-
ematical model or system (numerical or otherwise) can be apportioned to different
sources of uncertainty in its inputs, Saltelli et al. [2008].
Pannell [1997] says that sensitivity analysis can be useful for a range of pur-
poses, including:
1. Testing the robustness of the results of a model or system in the presence
of uncertainty.
2. Increased understanding of the relationships between input and output vari-
ables in a system or model.
3. Uncertainty reduction: identifying model inputs that cause significant un-
certainty in the output and should therefore be the focus of attention if the
robustness is to be increased (perhaps by further research).
4. Searching for errors in the model (by encountering unexpected relationships
between inputs and outputs).
5. Model simplification - fixing model inputs that have no effect on the output,
or identifying and removing redundant parts of the model structure.
6. Enhancing communication from modelers to decision makers (e.g. by mak-
ing recommendations more credible, understandable, compelling or persuas-
ive).
7. Finding regions in the space of input factors for which the model output is
either maximum or minimum or meets some optimum criterion.
Our methodology consists of many parameters, that are categorized into three
categories:
1. R o u t i n g a l g o r i t h m p a r a m e t e r s ; As described in chapter 5, our method-
ology utilizes the hybrid optimization method. Therefore many parameters
need to be tuned properly. Both the general and system rules that are
explained in subsection 5.3.1.1 are included in this category.
6.5 SENSITIVITY ANALYSIS 131

2. G e o m e t r y c o n s t r a i n t s ; The parameters of the influenced area of the geo-

metry constraints must be chosen properly to ensure that the pipes are
routed as intended.
3. O b j e c t i v e criteria; This is related to the cost of the pipes. Differences in
how pipe cost is calculated might affect the way pipes wiU be routed.
During the validation of the methodology, we used the optimized parameters
that are tuned properly.
To measure the parameter sensitivity of our methodology, we show how a
certain parameter affects the behavior of the methodology. Since the number of
parameters is large, only some of them are chosen to be varied.
Since the core of the routing algorithm has been discussed in detail in chapter
4 and 5, it is not interesting anymore to try to change these parameters. Also,
we will maintain the pipe routing rules that are used as parameters.
In the routing algorithm, basically to create pipes that are easily installed,
we would like to route pipes as orthogonally as possible. However, there are
conditions that require that the pipe should be routed non orthogonally, e.g.
when there are two consecutive bends, and the distance between two bends is
smaller than a minimum value. For this kind of situation, the algorithm allows
that particular part of the pipe to be routed non orthogonally. This parameter
needs to be tuned and as part of the sensitivity analysis, we vary this parameter
and compare the results.
In order to show the importance of the geometry constraints parameters, we
try to route the pipes with modified parameters for the influenced area of the
attraction, magnet, and distraction area.

6.5 Sensitivity Analysis

As described in the previous section, the sensitivity analysis is performed by
varying certain parameters of second category, the geometry constraints. In this
case, to isolate the effect of each parameter, the routing process is performed
individually for each parameter change.
The first parameter that is tried is the orthogonal restriction parameter. This
parameter decides whether the pipe is allowed to be routed non-orthogonally or
not. The second parameter is the attraction coefficient of the geometric con-
straints. As discussed in Chapter 5, this coefficient is used to influence whether
the pipe is routed as close as possible to a steel structure or as far as possible
from a particular object.
During the parameter tuning, we have found that the effect of both parameters
is non linear for the quality of the routed pipes. Therefore, we only show the
extreme lowest and highest parameter values. The results are compared with the
optimal solution.
Table 6.3 shows the summary of the comparison with the optimal solution.
Strictly orthogonal means that every pipe must be routed orthogonally, except
132 PIPE ROUTING METHODOLOGY VALIDATION 6.5

Length Cost
Optimal 100,00% 100,00%
Strictly orthogonal 100,03% 104,35%
No restriction orthogonal failed failed
Very high attraction coef. 102,33% 110,97%
Very low attraction coef. failed% failed%

Table 6.3: Comparison mill the results for oplimal parameter settings

if the distance between two bending points is shorter than allowed. In the other
way around, the non restriction orthogonal means that the pipes are allowed to be
routed non-orthogonally without any restriction at all. In the case of a very high
attraction coefficient, the objective cost value is low if the pipe is routed close
to a geometric constraint. Vice versa, with a very low attraction coefficient, the
objective cost value is not influenced by the distance between pipe and geometric
constraint.
Fig. 6.10 shows the example of pipes that are routed with the strictly ortho-
gonal parameter. As can be seen in that figure, the distance between two bending
points are too close to each other. In that case, it is required to use two elbows
rather than using a single pipe with two bends, which in terms of cost, can be
between 4-6 times more expensive than the normal pipe bending as shown in Fig.
6.11.
In the other way around, allowing pipes to be routed non-orthogonally without
any restriction might prevent the proposed methodology to find a complete solu-
tion. In our case, the methodology failed to find a complete solution if there is no
restriction at all for the orthogonality. Apparently the optimization procedure or
the parameters in this case need to be adjusted. Alternatively we could consider
starting the optimization from the standard case and slowly relax the orthogonal-
ity parameter and find a solution in that way. We did not pursue that possibility
in this thesis.
To achieve the optimal result as shown in Fig. 6.11, the restriction parameter
must be chosen properly. The parameter tuning must be done properly and it is
not a trivial task. As we can see in Table 6.3, the cost result using the strictly
orthogonal parameter is around 4% higher than the optimal one. Meanwhile, the
non restriction orthogonal parameter case failed to find the solution. From this
fact, it can be deduced that it is safer to choose the restriction parameter value
that is too high.
The variation of the geometry constraints parameters is also interesting. By
lowering the distance of influence in the attraction and magnet area, some of the
pipes are routed too far away from the steel construction. Therefore it is very
hard to put the pipe supports on it, and in most of the cases during operations
pipe vibration may occur. In our test, the low attraction coefficient value failed
to find the complete solution.
6.5 SENSITIVITY ANALYSIS 133

F i g u r e 6.10: Only allowed orthogonal pipe

F i g u r e 6.11: Pipes is allowed to be routed non orthogonally to allow a cheaper bending

134 PIPE ROUTING METHODOLOGY VALIDATION 65

F i g u r e 6.12: Optimal distance of influence and coefficient m geometry constraints area

F i g u r e 6.13: Large distance of influence and coefficient of the geometry constraints

area
6.6 SUMMARY 135

However, if this coefficient is too large, tfie solution won't be good anymore.
As can be seen in Fig. 6.13 and Fig. 6.12, by using the very large attraction
coefficient, it produces many unnecessary bends. Table 6.3 shows that the cost is
more than 10% higher.
Therefore, the importance to have an optimal parameter for the attraction
coefficient is higher than the restriction orthogonal parameter.

6.6 Summary
In this chapter we have validated our proposed methodology and its implement-
ation using the validation square from Seepersad et al. [2006]. This evaluation
aims to answer our main research objectives of this thesis:
1. to what extent can the expertise of a pipe engineer be identified and trans-
lated into procedures that lend themselves to be computerized?
2. if we build a pipe routing methodology based on the results of the first ques-
tion and combine it with advanced optimization techniques in a practically
applicable method, how good will it perform?
In this experiment, we measure the performance of our proposed methodology
to route pipes in the most difficult area in a ship. We have compared the result to
pipes that were routed manually by a pipe engineer. The comparison proves that
the methodology is able to perform the pipe routing process and is on par with
the quality of a pipe engineer result. Our experiment shows that the methodology
is able to implement the common knowledge.
We also learn from this experiment that the success ratio of the tools that were
built from the methodology is only 80%. Since we found that the cause of the
error probably was on the software engineering part, this number is good enough
to prove that the methodology is valid, but not yet perfect.
At the end of this chapter, we also performed some experiments by varying
some of the key parameters to identify the methodology's sensitivity to these
parameters. The results shows that great care must be taken to use the cored
parameter settings. Further work in this area is recommended.
Chapter

Conclusions and Recommendations

7.1 Research Questions

'J'he work reported in this thesis is motivated by the main research objectives
below:
1. to what extent can the expertise of a pipe engineer be identified and trans-
lated into procedures that lend themselves to be computerized?
2. if we build a pipe routing methodology based on the results of the first ques-
tion and combine it with advanced optimization techniques in a practically
apphcable method, how good will it perform?
In the course of experiments conducted to answer the question, we raised the
following six sub-questions:

1. On which design phases should we focus and what is the reason for that?

2. What information is needed to perform the pipe routing process, who is

responsible to provide it and how can one get it?

3. What is the common knowledge to route pipes that is used as guidance by

a pipe engineer?

4. Which knowledge from Research Question # 3 should be adopted in our pro-

posed methodology and to what extent can the current practical knowledge
be absorbed into a programmable methodology?

5. How to evaluate the quality of a set of routed pipes quantitatively?

137
138 CONCLUSIONS AND RECOMMENDATIONS 7.2

6. How to efficiently implement the routing algorithm in a given 3D environ-

ment?

We provide conclusions to those research questions in Section 7.2, and add

some recommendations in Section 7.3.

7.2 Conclusion
In order to stay competitive in the shipbuilding industry, European shipyards
need to be more efficient in design and production processes. One of the useful
directions is to be more innovative in the pipe routing process. A lot of effort has
been done in the past to be able to route pipes automatically in a ship. However
to date automatic routing is not applied nor possible in ship design.
As much tacit basic knowledge on the pipe routing process exists, we decided
to absorb the expertise of the pipe engineer and translate it into procedures that
lend themselves to be programmed and executed in a computer. We started by
establishing the information that is needed to route pipes, how to get it and who
is responsible for that. This step provided answers to our research question # 2
and the result is described in Chapter 2. By having that knowledge, in Chapter
5 we could focus on the selected information that is needed for our methodology.
To translate the expertise of a pipe engineer, we made an inventory of the
common knowledge of the pipe routing process. This step was not easy to be
done, especially since currently there is no official guideline for a pipe engineer to
route pipes in a ship. We performed many interviews experts in pipe routing to
absorb their knowledge. During the interview period, the pipe routing guideline
is compiled by v.d. Berg [2009]; the experienced pipe engineer team leader. The
interview results are formulated as the common knowledge of pipe routing. Thus,
with regards to research question # 3 , we should refer to the whole of Chapter 2.
Then we survey the state of the art in automatic pipe routing and in Section
3.6 we described some of the previous researches that have most relevance for
our problem domains. While reviewing those researches, we found the answer
to the research question # 1 . FYom Subsection 1.3.1 we knew that during the
contractual phase, approximate routing is sufficient. Combining suitably selected
existing methods with the completeness of our methodology, e.g. the ability to
extract the required data easily from CAD software and automatically perform the
cell generation, we can build the pipe routing application for the pre-contractual
phase easily. For the sake of widespread applicability however, we placed our focus
on the scientifically much more interesting and demanding detail design phase.
With regards to research question # 4 , Subsection 5.3.1 provided the answer.
The essential common knowledge with direct impact on the path finding problem
is selected and reformulated as general and system rules.
To perform the optimization process we defined the objective function that
needs to be minimized. As described in Chapters 2 and 5, the objective in the
7.2 CONCLUSION 139

pipe routing process can be divided into two categories. First is the quantitative
value that consists of pipe production and installation cost. This is relatively easy
to be calculated, and since our research attempt to use the practical environment,
the actual total pipe cost from the pipe contractor is used. The second part is of
non quantitative nature. It concerns issues that involve parallelism and aesthetics
considerations.
With regards to research question # 5 , we conclude that the best way to meas-
ure the quality of the pipes is by assigning quantitative measures to them. The
quantification process was done by giving a weight to every cell based on the geo-
metric constraints reflecting the space and objects surrounding it, then combine
the pipe cost value with the cell's weight. The results are used as the objective
function.
Routing pipes is hard. Implementing the expertise of pipe engineers into
computer procedures is even harder. In this thesis, we proposed a methodology
and methodology to answer that challenge. It does not merely focus on the
algorithm to find the pipe path, but also includes the data preparation step and,
more generally, its embedding in the practical design process.
In Chapter 5, we described our proposed methodology in detail. We started
with the interface module that contains three important parts; the functionality
to extract data from CAD software, the smart P&ID tools and the model simpli-
fication. The last part, the model simplification, is very important to ensure that
the whole automatic routing process can be done within an acceptable time. We
established a novel method to automatically simplify the 3D model heuristically
to ensure that the model is simple yet sufficiently detailed.
In our methodology, we reformed the 3D space into a grid of cells. Balancing
between the quality of the routed pipes and the computation time, the cell size
was fixed at 5 mm by 5 mm by 5 mm. The fact that the cell size is smaller than the
diameter of pipes raises the need of collision detection. We know that collision
detection is computationally expensive, therefore we introduced the expanding
obstacle method as described in Chapter 5. Even though this method looks
simple, the performance of the shortest path algorithm significantly improved.
For the path finding itself, we investigated both deterministic and heuristic
shortest path algorithms. We made a comparison by implementing the standard
form of 5 well-known shortest path algorithms, and selected the A* algorithm as
the main single pipe routing algorithm and the Mikami-Tabuchi algorithm as the
blockage tester.
Since our goal is to route many pipes, there is a possibility that two or more
pipes axe blocking each other. We call this the mutually intervening problem. We
tackled this by implementing the particle swarm optimization to route multiple
pipes at the same time.
Since our methodology basically routes pipe one at a time, the order of pipes
is very important. To optimally order the pipes, we implemented the discrete
particle swarm optimization which minimized the objective values of the routed
pipes.
140 CONCLUSIONS AND RECOMMENDATIONS 7.3

In Chapter 6, we have appHed the methodology to route pipes in the machinery

room of a ship. The results prove that our proposed methodology is able to
perform automatic routing processes in a real ship. The results are excellent and
support detailed engineering research questions. We also showed that we can
build the methodology in such a way that it uses the selected common knowledge
of the expert pipe engineers. In this way, it is knowledge based.
In the course of this application we have also validated the methodology using
the validation square method.

7.3 Recommendations
As we mentioned in the previous section, we have proved that our proposed meth-
odology can be used to solve automatic pipe routing problems in difficult areas
of a ship. However, our work is not complete, since there are some important
aspects that have not been addressed in this thesis.
To improve the proposed methodology, there are at least three main points that
should be focused on in future research. The first point concerns the number of
common pipe routing knowledge rules that should be included in the methodology.
As mentioned in Chapter 5, not all of the common knowledge is integrated in the
methodology described in this thesis. By including more items, the quality of
the routed pipes may be further increased. However, the selection must be done
carefully, by only selecting the most important ones that have most influence.
The result from our proposed methodology is only in the form of pipe routes.
It still requires manual work to split it into pipe spools. Therefore, it is also very
interesting to complement the methodology with the method to split pipes into
spools and use the result as an input for the research of Wei [2012] to generate
the assembly sequence automatically.
As explained in Chapter 6, our methodology uses 90 degree angles as default,
and only chooses other bending angles if required. This situation is good enough
for steel pipes, which form the majority of the pipes in a ship. However, a PVC
pipe might require to have different bending angles, and the current behavior of
the methodology is not yet entirely suitable for that situation.
Another interesting subject for further research is to extend our methodology
to route HVAC and cable trays in a ship.
Bibliography

ABS. Guidance notes on the application of ergonomics to marine systems. The

American Bureau of Shipping, 1998.
P. J. Angehne. Using selection lo improve particle swarm optimization. In Proc.
IEEE Congress on Evolutionary Computation, pages 84-89, May 1998.

Andi Asmara and Ubald Nienhuis. Optimum routing of piping in real ships under
real constraints. In Proc. Int. Conference on Computer and IT Application in
Maritime Industry, pages 271-281, April 2008.

Pipework Design User Cuide. AVEVA, 2007.

C B Barber, D P Dobkin, and H T Iluhdanpaa. The quicklmll algorithm for
convex hulls. ACM Transactions on Mathematical Software, (22), 1996.
Marshall W. Bern and Ronald L. Graham. The shortest-network problem. Sci-
entific American, 260:84-89, 1989.
Ewerton E.S. Calixto, Paula G. Bordeira, Hugo T. Calazans, Cesar A. C. Tav-
arcs, and Marco T. D. Rodriguez. Plant design project automation using an
automatic pipe routing routine, 2009.

M. Clerc. The swarm and the queen: towards a deterministic and adaptive particle
swarm optimization. In Proc. IEEE Congress on Evolutionary Computation,
volume 3, pages 1951-1957, Washington, DC, July 1999.
M. Clerc and J. Kennedy. I'ho particle swarmexplosion, stability, and convergence
in a multidimensional complex space. 6:58 73, February 2002.
Maurice Clerc. Discrete particle swarm optimization. In Godfrey C Onwubolu
and B. V. Babu, editors. New optimization techniques in engineering, pages
219 240. Springer, 2004.
Maurice Clerc. lYibes-d, 2008. URL h t t p : / / c l e r c . m a u r i c e . f r e e , f r / p s o / \ #
TRIBES-D.

141
142 BIBLIOGRAPHY

J Cohen, A Varshney, D Manocha, G 'l\irk, H Weber, P Agarwal, F P Brooks,

and W Wright. SimphficaUon envelopes. In In: Rushmeier, H. (Ed.), SIG-
GRAPH 96 Conference Proceedings. Annual Conference Series. ACM SIG-
GRAPH, pages 119-128. Addison Wesley, 1996.

Edgar W. Dijkstra. A note on two problems in connexion with graphs. Numerische

Mathematik, 1:269-271, 1959. ISSN 0029-599X.

R. C. Eberhart and J. Kennedy. A new optimizer using particle swarm theory. In

Proc. IEEE International Symposium on Micro Machine and Human Science
6th, pages 39-43, Nagoya, Japan, October 1995.

R. C. Eberhart and Y. Shi. lYacking and optimizing dynamic systems with

particle swarms. In Proc. IEEE Congress on Evolutionary Computation, pages
94-97, Seoul, Korea, May 2001.

A. I. El-Gallad, M. E. El-Hawary, A. A. Sallam, and A. Kalas. Enhancing the

particle swarm optimizer via proper parameters selection. In Proc. Canadian
Conference on Electrical and Computer Engineering, pages 792-797, 2002.

H. Y. Fan and Y. Shi. Study of Vmax of the particle swarm optimization al-
gorithm. In Proc. of the Workshop on Particle Swarm Optimization, Indiana-
polis, IN: Purdue School of Engineering and Technology, April 2001.

X. Fan, Y. Lin, and Z. Ji. The ant colony optimization for ship pipe route design
in 3d space. In World Congress on Intelligent Control and Automation, October
2006.

R. Edward Freeman and David L. R,eed. Stockholders and stakeholders: A new

perspective on corporate governance. California Management Review, Vol 25
Issue 3:88-106, 1983.

F.S.Nooruddin and G lYirk. Simplification and repair of polygonal models using

volumetric techniques. IEEE Trans. Visual Computer Graphics, 9:191-205,
April 2003.

Z. L. Gaing. A particle swarm optimization approach for optimum design of pid

controller in avr system. 19:384-391, June 2004.

Michael Garland and Paul S. Heckbert. Surface simplification using quadric error
metrics. In Proceedings of the 24 ih annual conference on Computer graphics and
interactive techniques, SIGGRAPH '97, pages 209-216. ACM Press/Addison-
Wesley Publishing Co., 1997. ISBN 0-89791-896-7.

A.M. Gibbons. Algorithmic Graph I'fieory. Cambridge University Press, 1985.

ISBN 9780521288811.
 BIBLIOGRAPHY 143

D. E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learn-

ing. Kluwer Academic Publishers, Boston, MA, 1989.

C Gotsman, S Gumhold, and L Kobbelt. Simplification and compression of 3d

meshes. In Tutorials on Multiresolution in Geometric Modelling, pages 319-361.
Springer, 2002.

S. Gottschalk, M. C. Lin, and D. Manocha. Obbtree: a hierarchical structure for

rapid interference detection. In Proceedings of the 23rd annual conference on
Computer graphics and interactive techniques, SIGGRAPH '96, pages 171-180,
1996. ISBN 0-89791-746-4.

R. Guirardello and R. E. Swaney. Optimization of process plant layout with pipe

routing. Computers and Chemical Engineering, 30:99-114, 2005.

R.L. Harrington. Marine Engineering. The Society of Naval Architects and Marine
Engineers, 1992.

Peter E. Hart, Nils J. Nilsson, and Bertram Raphael. A formal basis for the
heuristic determination of minimum cost paths. IEEE 'Prans. Systems Science
and Cybernetics, 4(2): 100 107, 1968.

David W. Hightower. A solution to line routing problems on the continuous plane.

In Proc. 6th Design Automation Workshop (IEEE), pages 1-24, 1969.

,1 Iljelmcrvik and J.-C. Leon. Gpu-accelerated shape simplification for mechanical-

based applications. In Proc. IEEE Int. Conference on Shape Modeling and
Applications, pages 91-102, June 2007.

Chen Hua-hong, Luo Xiao-nan, and Ling Ruotian. Mesh simplification algorithm
based on quadrangle collapse. In Proc. IEEE Int. Conference on Image and
Graphics, pages 960-965, August 2007.

Teruaki Ito. A genetic algorithm approach to piping route path planning. Journal
of Intelligent Manufacturing, 10:103-114, 1999.

J. Jerald, P. Asokan, G. Prabaharan, and R,. Saravanan. Scheduling optimisation

of flexible manufacturing systems using particle swarm optimisation algorithm.
The Int. Journal of Advanced Manufacturing Technology, 2004.

Sang-Scob Kang, Se-Hyun Myung, and Soon-Hung Han. Design expert system
for auto-ro>iting of ship pipes. In Proc. Pacific Conference on Manufacturing,
October 1996.

Sang-Seob Kang, Se-Hyun Myung, and Soon-Hung Han. A design expert system
for auto-routing of ship pipes. Journal of Ship Production, 15:1-9, 1999.
144 BIBLIOGRAPHY

A Kaufman. Voxels as a computational representation of geometry. In in The

Computahonal Representation of Geometry. SJGGRAPII '94 Course Notes,
page 45, 1994.

J. Kennedy. Why does it need velocity? In Proc. IEEE Swarm Intelligence

Symposium, pages 38-44, June 2005.

J. Kennedy and R. C. Eberhart. Particle swarm optimization. In Proc. IEEE

International Conference on Neural Networks IV, pages 1942-1948, Piscataway,
NJ, December 1995.

David G Kirkpatrick and Raimund Seidel. The ultimate planar convex hull al-
gorithm. SIAM J. Comput, (15), 1986.

J. Klein Woud and D. Stapersma. Design of propulsion and electric power gen-
eration systems. IMAREST, 2003.

T. Krink and M. L0vbjerg. The lifecycle model: Combining particle swarm op-
timisation, genetic algorithms and hillclimbers. In Proc. International Confer-
ence on Parallel Problem Solving from Nature, pages 621-630, Granada, Spain,
September 2002.

T. Krink, J. S. Vcsterstrom, and J. Riget. Particle swarm optimisation with spa-

tial particle extension. In Proc. IEEE Congress on Evolutionary Computation,
volume 2, pages 1474-1479, May 2002.

R. A. Krohling. Gaussian swarm: A novel particle swarm optimization algorithm.

In Proceedings IEEE Conf. on Cybernetics and Intelligent Systems, pages 372-
376, Singapore, December 2004.

R. A. Krohhng, F. Hoffmann, and Ld. S. Coelho. Co-evolutionary particle swarm

optimization for min-max problems using gaussian distribution. In Proc. IEEE
Congress on Evolutionarij Computation, pages 959-964, June 2004. '

C. Y. Lee. An algorithm for path connections and its applications. IEEE 'Prans.
on Computers, EC-10:346 365, 1961.

E.V. Lewis. Principles of Naval Architecture. Society of Naval Architects, 1988.

ISBN 978-9991181417.

A. Lim, Jing Lin, and Fei Xiao. Particle swarm optimization and hill climbing to
solve the bandwidth minimization problem. In Proc. The Fifth Melafieuristics
International Conference, Kyoto, Japan, August 2003.

Myung 11 Roh, Kyu-Yeul Lee, and Woo-Young Choi. Rapid generation of the
piping model having the relationship with a hull structure in shipbuilding. Ad-
vances in Engineering Software, 38:215-228, 2007.
 BIBLIOGRAPHY 145

M. lj0vbjcrg and T. Krink. Extending particle swarm optimisers with sell organ-
ised criUcaiity. In Proc. IEEE Congress on Evolutionary CompuiaLion, pages
1588 1593, Honolulu, Hawaii, 2002.

Lawrence Markosian, Philip Newcomb, R.ussell Brand, Scott, Burson, and 'I'ed
Kitzmiller. Using an enabling technology to reengineer legacy systems. Com-
municaiions of the ACM, 37(5):58-70, 1994.

Y. Matsui, 11. Takagi, S. Emori, N. Masuda, S. Sasabe, C. Yoshimura, T. Shirai,

S. Nioh, and B. Kinno. Automatic pipe routing and material take-oil system
for chemical plan. In ConJ. on Design Automation, pages 121 427, 1979.

K. Mikami and K. Tabuchi. A computer program lor optimal routing of printed

circuit connectors. In Proceedings of IFIPS, volume H47, pages 1475-1478,
1968.

R. Newell. An interactive approach to pipe routing in process plants. J. Inform-

ation Processing, pages 121-127, 1972.
T. A. J. Nicholson. Finding the shortest route between two points in a network.
The Computer Journal, 9:275-280, 1966.

David J. Pannell. Sensitivity analysis: strategies, methods, concepts, examples.

Agricultural Economics, 16:139-152, 1997.

Jin-IIyung Park. Pipe-Routing Algorithm Development for A Ship Engine Room

Design. PhD thesis, University of Washington, 2002.
Jin-Hyung Park and Richard L. Storch. Pipe-routing algorithm development:
case study of a ship engine room design. Expert Systems with Applications, 23:
299309, 2002.

T. Peram, K. Veeramachaneni, and C. K. Mohan. Fitness-distance-ratio based

particle swarm optimization. In Proc. IEEE Swarm Intelligence Symposium,
pages 174-181, April 2003.

Patrick William Rourke. Development of a Three-Dimensional Pipe Routing Al-

gorithm. PhD thesis, Brunei University, 1975.

J. Salerno. Using the particle swarm optimization technique to train a recurrent

neural model. In Proc. IEEE International Conference on Tools with Artificial
Intelligence, pages 45-49, Newport Beach, CA, November 1997.
A. Saltelii, Marco Ratto, Terry Andres, Francesca Campolongo, Jessica Cariboni,
Dcbora Gatelli, Michaela Saisana, and Stcfano Tarantola. Global Sensitivity
Analysis: The Primer. 2008.

Sunand Sandurkar and Wei Chen. Gaprus - genetic algorithms based pipe routing
using tessclatcd objects. The Journal of Computers in Industry, 1998.
146 BIBLIOGRAPHY

Carolyn C. Seepersad, Kjartan Pedersen, Jan Emblemsvag, Reid Bailey, Janet K.

Allen, and Parrokh Mistree. The validalion square; How does one verify and
validate a design method? In Kemper E. Lewis, Wei Chen, and Linda C.
Schmidt, editors. Decision Making In Engineering Design. ASMl'], Now York,
2006.
Y. Shi and R. C. Eberhart. A modified particle swarm optimization. In Proc.
IEEE International Conference on Evolutionary Computation, pages 69-73,
Piscataway, NJ, 1998.

Y. Shi and R. C. Eberhart. Comparing inertia weight and constriction factors in

particle swarm optimization. In Proc. IEEE Congress on Evolutionary Com-
putation, pages 84-88, San Diego, CA, May 2000.
Cai Tao, Pan Feng, and Chen Jie. Adaptive particle swarm optimization al-
gorithm. In Proc. IEEE Congress on Intelligent Control and Automation,
volume 3, pages 2245-2247, June 2004.
Li Tao, Wei Chengjian, and Pei Wenjang. Pso with sharing for multimodal func-
tion optimization. In Proe. IEEE Int. Conference on Neural Networks and
Signal Processing, volume 1, pages 450-453, December 2003.
D. A. Taylor. Introduction to Marine Engineering. Elsevier Butterworth-
Heinemann, 1996.
Mario v.d. Berg. Guidelines piping design. Internal Document IHC Merwede
O&M, 2009.
Zheng Wang and Jon Crowcroft. Analysis of shortest-path routing algorithms in
a dynamic network environment. SIGCOMM Comput. Commun. Rev., 22(2):
63-71, April 1992.
Glenn E. Wangdahl, Stephen M. Pollock, and John B. Woodward. Minimum
trajectory pipe routing. Journal of Ship Research, 18;44 49, 1974.
Yan Wei. Automatic generation of assembly sequence for the planning of outfitting
processes in shipbuilding, 2012.
David Wolpert and William G. Macready. No free lunch theorems for optimiza-
tion. IEEE Transactions on Evolutionary Computation, 1997.
H. Yoshida, K. Kawata, Y. Fukuyama, S. Takayama, and Y. Nakanishi. A particle
swarm optimization for reactive power and voltage control in electric power
systems. In Proc. IEEE Congress on Evolutionary Computation, pages 87-93,
Seoul, Korea, May 2001.
David Zhu and Jean-Claude Latombe. New heuristic algorithms for efficient hier-
archical path planning. IEEE Transactions on Robotics and Automation, 7:
9-20, 1991.
 BIBLIOGRAPHY 147

]l..J. Zuurmond. Automaüsch routeren van pijpleidingen in schepen. Master's

thesis, 'I'echnischc Universiteit Delft, 2004.
I
Summary

Pipe routing consumes a large part of the total required effort in the ship design
process. In current practice, it takes around 30-40 thousands man hours to route
pipes of a middle size complex ship. Reducing the time of this process will have
a large impact in total engineering cost.
Judged by the results, the current process produces an excellent solution al-
though an objective, quantitative assessment of this is difficult given the fact that
there is no mathematically determined optimum known. So, in order to speed
up the process, the current process was investigated, adopted and subsequently
translated into the computer procedures.
The main objective of this thesis is to create a pipe routing methodology that
can be used in ship detail design process in practice. The methodology consists
of the functional framework, the architecture and its implementation.
In order to do it properly, the pipe routing process was investigated. Since no
formal documentation that explains how pipe routing process should be done in
practice is available, this knowledge was gained by carrying out many interviews
and discussions with experienced pipe engineers and other stakeholders in the
pipe routing process.
Beside the know-how on how a pipe engineer route pipes in ship, those inter-
views and discussions provide us with the criteria of pipe routing in ship
Based on that knowledge, the functional framework of our methodology was
derived. The elements needed by the functional framework were investigated and
reviewed, starting with reviewing the previous research attempts on improving
pipe routing process. Next the well known algorithms needed by each part of the
functional framework were compared and selected.
After all elements had been completed, the architecture of pipe routing meth-
odology was built. Finally it was implemented into a prototype software package
that can be used to route pipes in a real ship.
The validation ol the proposed methodology was carried out using the valid-
ation square technique by performing the structural and performance vaUdation.
The implementation of the proposed methodology was used to solve the pipe rout-

149
150 SUMMARY

ing problem in a machinery room of a real ship. I'he parameter variances was
also performed and compared.
As conclusion the pipe routing methodology was developed and validated on
the difficult area of a complex ship. As shown by the quality of the test case
result, the main objective that was previously mentioned is fulfilled indeed.
Samenvatting

Routing van pijpleidingen beslaat een groot deel van de totale benodigde in-
spanning in het scheepsontwerpproces. In de hedendaagse praktijk zijn 30-40
duizend manuren vereist om de pijpleidingen van een middelgroot, complex schip
te routeren. Vermindering van de tijd die benodigd is voor dit proces, zou een
grote impact hebben op de totale engineering kosten.
Wanneer men naar de resultaten van het handmatige routeerproces kijkt, valt
op dat dit zeer goed is, hoewel een objectieve, kwantitatieve maatstaf moeilijk
aan te leggen is omdat er geen wiskundig bepaald optimum bekend is. Daarom
werd, om het proces wezenlijk te versnellen, het huidige proces als uitgangspunt
genomen en onderzocht. Vervolgens werd dit vertaald in algoritmen die door een
computer opgelost kunnen worden.
Het hoofddoel van dit proefschrift is om een pijp-routing methodiek te ontwikkelen,
die in de praktijk gebruikt kan worden, namelijk gedurende het detail-ontwerpproces.
De methodiek bestaat uit het zg. functionele raamwerk, de architectuur en de im-
plementatie.
Ten behoeve van deze ontwikkeling werd het pijp-routing proces onderzocht.
Aangezien er geen formele documentatie van het proces bestond waarin praktijk-
voorschriften worden beschreven, is deze kennis verkregen door middel van vele
interviews en discussies met ervaren pijpschetsers en andere belanghebbenden in
het pijp-routing proces.
Het resultaat van deze interviews en discussies is een vastgelegde rationale
betreffende de manier waarop een schetser de pijpen routeert en een set formele
criteria.
Van deze kennis kon het functionele raamwerk van de beschreven methodiek
worden afgeleid. De elementen benodigd voor het raamwerk werden onderzocht,
te beginnen met een onderzoek naar eerdere pogingen om het pijp-routing proces
te verbeteren. Vervolgens werden alle bekende algoritmen die van toepassing
zouden kunnen zijn op de verschillende onderdelen van het functionele raamwerk
vergeleken, en de meest geschikte oplossingen geselecteerd.
Nadat alle afzonderlijke functionele elementen ontwikkeld waren, werd de ar-

151
152 SAMENVATTING

chitectuur van do methodiek gemaakt. Tenslotte werd alles geïntegreerd in een

testversie van een softwarepakket dat gebruikt kan worden om daadwerkelijk auto-
matisch pijpen te routeren in een echt schip.
De validatie van de voorgestelde methodiek werd uitgevoerd met behulp van de
zg. validation square techniek (validatickwadranten), die kijkt naar de structurele
aspecten van validatie, maar ook kijkt naar de geleverde (ontwerp)prestaties van
de methodiek.
Tenslotte werd de pijp-routing methodiek ontwikkeld en gevaÜdeerd op een
moeilijke ruimte van een complex schip. De kwaliteit van het resultaat van de
testcase Iaat zien dat het onderzoeksdoel, zoals dat eerder werd vermeld, inder-
daad wordt behaald.
Acknowledgments

This research was conducted within the CE3P project (Concurrent Engineering,
Production, Planning and Pricing) and the Integraal Samenwerken project. The
pipe routing research subject was carried out in IHC Offshore and Marine who
employed me to do this research. I gratefully acknowledge management of IIIC
Offshore and Marine for this opportunity and support.
I would like to thank the doctoral committee for your time and effort review-
ing this thesis. My gratitude for your comments, suggestions and critiques has
improve the quality of this thesis.
This research project would not have been possible without the support of
many wonderful people. I wish to express my gratitude to my promotor, prof.dr.Ir.
Ubald Nienhuis who was abundantly helpful and offered invaluable assistance,
support and guidance, not only work related but also helped me to get through
life in The Netherlands. I would like to thank him for the trust he gave me, the
independence he allow me in conducting this research, his experience he shared
with me and his advices and support he always offered me when I needed them.
Other most valuable support has come from Teus van Nordennen who has
given me the opportunity, trust and support to conduct this research. He was
also the one who kept me in balance between practical and scientific worlds.
I would to thank Jenny Coenen who helped me a lot, especially during the
beginning of my research and also her willingness to proofread this thesis.
Since the beginning of my journey, I received many supports and assistances
from a large group of people. There are too many of you to mention but Pd
especially like to thank the following people.
Frank Dekker and Gerco Brand who gave me the first insight about pipe
routing in practice, introduced me to the CAD software package and provided
me with many kinds of data through all the research. Mario van den Berg who
shared his expertise in the pipe routing process and especially for compiling the
document on pipe routing in practice.
The engineering department from IHC O&M in general and Gert Rook, Sjaak
van Dorsten, Bert Rissema, Wim van Mannen, Cor Molenaar, Gerard Smouter,

153
154 ACKNOWLEDGMENTS

Andre Dubbeldam, Krijn Ooms, Ilja van Deurzen, Henk Scheep, Peter de Rek,
Nuur Nuur, Ken Schie and Edwin Keizer in particular. The discussion on pipe
routing process in practice provided invaluable knowledge for this research.
Peter Wagenaar, your insight on pipe cost calculation during the pre-contract
phase is highly appreciated.
Jeroen Kaarsemaker and other members of process management department,
thank you for the discussion that we had during these past years. Especially
Reinier Zuurmond who started this research project and gave me a solid base to
start with.
Prom IHC Piping, Peter Gelderbloom and Mark Whitney for the knowledge
on how pipes are produced, Ad Hazelaar and Edwin van Leeuwen who showed
me how to calculate pipe cost in practice, and especially Edwin, thanks a lot for
the data.
Also my colleagues in TU Delft, in particular Erik Ulijn for sharing his of-
fice with me, Ria Nieuwland-Jobse for her helping me with many administrative
things, Jan Jaap Nieuwenhuis for letting me use some of his typesettings, Guus
van der Bles, late Hugo Grimmelius, Bart van Oers, Robert Hekkenberg, Jeroen
Pruyn, Wei Yan, Elena Moredo for the company and the discussions we have had
over the past years.
Many thanks to Dina Shona Ijaila for reviewing this manuscript.
My parents and family back home in Indonesia, who understand my choice to
live far away from all of you.
Last, but most certainly not least, I would especially like to thank my awesome
family for the love, patience, support, and constant encouragement. Without
them I would not come this far. Erin, Paza, Gafa, Kai, this is for you.
Curriculum Vitae

Andi Asmara was born in Jakarta, Indonesia on February 18, 1974. After finishing
his highschool in SMA 70 Jakarta in 1992, he studied Automation and Control
Engineering from Electrical Department at Bandung Institute of Technology. He
graduated in 1996, and his BSc. final project title is 'Application of Programmable
Logic Controller for Analog and Discrete System'.
He worked as an Automation Product Engineer at Schneider Electric Indonesia
from 1996 to 1999. In 1999, he became a co-founder of Ekakarsa Cita Dinamika, an
automation system integrator company which was responsible for programmable
logic controller system during the commissioning and warranty period at Musi
Pulp Mill.
In 2002, he continued his education in automation and robotics program at
Universitat Dortmund, Germany. He graduated in 2005 as master of science in
automation and robotics, and his MSc. thesis title is 'Accelerated Co-evolutionary
Particle Swarm Optimization for Computed-Torque Controller parameter tuning
for Robot Manipulator'.
In 2005, he started with his PhD research as a researcher at Merwede Shipyard
(now IHC Offshore and Marine) and carried out one of the sub-project of the CE3P
project that was done in co-operation with the TU Delft. Within CE3P, he focused
at the automatic pipe routing project. After the CE3P project finished in 2009,
the automatic pipe routing project was continued in the Integraal Samenwerken
project.
After the pipe routing research finished in 2010, but while still writing his
dissertation, he continued to work in IHC Offshore and Marine and also worked
on other sub-project in Integraal Samenwerken to develop the 3D Information
Viewer.
He is currently employed as one of the project managers in the One IHC
Program inside IHC Merwede.

155
 789065"623263
13W34870/ TV 9789065623263
i