ELSEVIER SCIENCE Ltd

The Boulevard, Langford Lane

Kidlington, Oxford OXS 1GB,UK

Elsevier Science Internet Homepage

http://wWW ,elsevier.com
Consult the Elsevier Homepage for full catalogue information on all books, journals and electronic products and services.
IFAC Publications Internet Homepage
http://www .elsevier.cowl ocatelifac
Consult the IFAC Publications Homepage for full details on the preparation of IFAC meeting papers, published/forthcoming IFAC books,
and information about the IFAC Journals and affiliated journals.

Copyright @) 2002 IFAC

All Righls Reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted in any form or by any means:
electronic. electrostatic, magnetic tape, mechanical, photocopying, recording or otherwise, without permission in writing from the copyrighl
holders.

First edition 2002

Library of Congress Cataloging in Publication Data

A catalogue record for this book is available from the Library of Congress

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

ISBN 0-08-043236 0
ISSN 1474-6670

These proceedings were reproduced from manuscripts supplied by the authors, therefore the reproduction is not completely uniform but neither
the format nor the language hove been changed in the interests of rapid publication. Whilst every effort is made by the publishers to see that
no inaccurate or misleading data, opinion or statement appears in this publication, they wish to moke it clear that the data and opinions
appearing in the articles herein are the sole responsibility of the contributor concerned. Accordingly, the publisher, editors andtheir respective
employers, officers and agents accept no responsibility or liability whatsoever for the onsequences of any such inaccurate or misleading data.
opinion or statement.

Printed in Great Britain

To Contact the Publisher

Elsevier Science welcomes enquiries concerning publishing proposals: books, journal special issues, conference proceedings, etc.
All formats and media can be considered. Should you have a publishing proposal you wish to discuss, please contact; without
obligation, the publisher responsible for Elsevier's industrial and control engineering publishing programme:

Dr Martin Ruck
Publishing Editor
Elsevier Science Ltd
The Boulevard, Langford Lane Phone: +44 1865843230
Kidlington, Oxford Fax: +44 1865843920
OX5 1GB, UK E.mail: m.ruck@e1sevier.co.uk

General enquiries, including placing orders, should be directed to Elsevier's Regional Sales Offices - please access the
Elsevier homepage for full contact details (homepage details at the top of this page).
IFAC CONFERENCE ON CONTROL APPLICATIONS IN
MARINE SYSTEMS 2001

Sponsored by
International Federation of Automatic Control (IF AC)

Co-sponsored by
Technical Committees on:
- Transportation Systems
- Intelligent Autonomous Vehic1es
Advanced Control Technology Club

Organized by
Industrial Control Centre, Department of Electronic and Electrical Engineering, University of
Strathclyde, Scotland

International Programme Committee (IPC)

Job van Amerongen (NL) Sauro Longhi (IT)
Joan Batlle (SP) Adam Lozowicki (PL)
Daniele Bertin (IT) Sadko Mandzuka (CR)
Mogens Blanke (DN) Kohei Ohtsu (JP)
Roland Burns (UK) Antonio Pascoal (PR)
Manuel Haro Casado (SP) Malek Pourzanjani (UK)
Roberto Cristi (USA) Guang Ren (CH)
Decio Donha (BR) Juhani Sukselianen (FL)
Ferial El-Hawary (CA) Asgeir S~rensen (NR)
Philippe Fabiani (FR) Robert Sutton (UK)
Thor Fossen (NR) Antonio Tiano (IT)
Mike Gerigk (PL) Zoran Vukic (CR)
Kazuhiko Hasegawa (JP) Dana R Yoerger (USA)
Thomas Holzhueter (GR) Junku Yuh (USA)
Claes Kallstrom (SW) Yuuzi Terada (JP)
Katsuro Kijima (JP) Markos Papageorgiou (GR)
Naomi Leonard (USA)

National Organizing Committee (NOC)

Grimble, M.J. Chairman
Katebi, Reza Deputy Chairman
Roberts, G.N. Programme Chairman
Yamamoto, I. Programme Co-chairman
Johnson, M. Finance Chair
Dinwoodie, S. Secretary
 PREFACE

The IFAC Conference on Control Applications in Marine Systems (CAMS 2001) was successfully
held on I8-20th of July 2001 at the University of Strathclyde in Glasgow, Scotland. More than 120
delegates from over 20 countries attended the Conference and discussed and exchanged their research
results in marine control systems design and applications. CAMS 2001 attracted the highest number of
papers and delegates compared to the previous ones. This shows the importance and the increasing
trend in the control design and application in marine systems. The papers were presented in the areas
of Autopilot Systems, Dynamic Positioning Systems, Tracking Systems, Roll Stabilization, High
Speed Crafts, Underwater Vehicles, Robotic Off-shore Systems, Ship Navigation Systems, ship
propulsion, Traffic Guidance and Control Systems, Engine and Machinery Control Systems, Safety
and Fault Tolerant Control, and Training and Vehicle Simulation.

The three plenary sessions were well attended and presented by Mogens Blanke, Danish Technical
University, who reviewed recent advances on maritime safety using fault tolerant control systems. Ken
Cartledge, Rolls Royce, UK, presented the second plenary on 'In pursuit of lean manning: ship
automation and the value of simulation' and discussed the importance of simulation in reducing
manning in ship automation from an industrial viewpoint. The third plenary was presented by John
Moscopoulos, US Navy on Ship automation from the US Navy perspective and summarized the
requirement for new areas of research in marine control. The three semi-plenary papers were presented
by Ikuo Yamamoto, Mitsubishi Heavy Industry, Japan, who discussed the result of their research on
flexible fin propulsion and robotic fish. Gregory Provan, Rockwell Automation, USA, presented a
motivating talk on Dynamically reconfigurable and scalable distributed shipboard automation system
for improving sustainability and survivability and A.J. Haywood, Maritime Dynamics Inc., presented a
talk on various aspects of ride control systems that have been developed for high-speed ferries over the
last decade.

There were also three invited sessions on underwater vehicle control and high speed ship control
which gave an excellent review of the current work in these areas.

The papers presented in CAMS200I Conference gave an excellent view both of the state-of-the-art as
well as insight into the future of control in marine systems.

Dr Reza Katebi
Industrial Control Center
Dept of Electronic and Electrical Engineering
University of Strathclyde
Glasgow, Scotland
September 2001
Email: r.Katebi@eee.strath.ac.uk
 CONTENTS

PLENARY PAPER I

Enhanced Maritime Safety Through Diagnosis and Fault Tolerant Control 1

M. BLANKE

SHIP MOTION MODELLING AND CONTROL

A New Approach Towards Designing Ship Motion Control Systems 21

P.G.M. van der KLUGT, C. de KEIZER

An Improved Numerical Model for the Study of Controlled Ship Motions in Extreme Following and Quartering Seas 27
Z. AYAZ, K.J. SPYROU, D. VASSALOS

Ship Stability Monitoring by Motion Frequency Analysis 39

R. FERREIRO GARCIA, C.A.F. AMEAL

SHIP CONTROL IN RESTRICTED AREAS

The Analysis of a Ship Fuzzy Domain in a Restricted Area 45

Z. PIETRZYKOWSKl

Automatic Ship Berthing Using Parallel Neural Controller 51

N. 1M, K. HASEGAWA

An Electronical Aid for Manoeuvring Ships in Restricted Waters 59

R.~LER,M.DEMUTH

MARINE APPLICATION OF FUZZY LOGIC AND NEURAL NETWORK

Design of Automatic Collision Avoidance System Using Fuzzy Inference 65

K. KlJIMA, Y. FURUKAWA

Autopilot Designed with Fuzzy Set Theory 71

A. ZIRILLI, A. TIANO, G.N. ROBERTS, R. SUTION

Fault-Tolerant Control Based on Adaptive Neural Network 77

T.K. CHANG, D.L. YU

Neural Networks for System Identification of Coupled Ship Dynamics 83

P. MARTIN, M.R. KATEBl, 1. YAMAMOTO, K. DAIGO, E. KOBAYASHI, M. MATSUURA, M. HASHIMOTO,
H. HIRAYAMA, N. OKAMOTO

POSITION, ROUTING AND PATH CONTROL

Modelling and Control of Towed Marine Seismic Streamer Cables 89

E. PEDERSEN, AJ. S0RENSEN

Non-Linear Semi-Submersible Positioning System Design Using an 1i_ Controller 95

D.C. DONHA, E.A. TANNURI

Methods of Optimal Ship Routing for Weather Perturbations 101

K. STAWICKI, R. SMIERZCHALSKl

v
 Assisted Dynamic Positioning System for a FPSO Based on Minimization of a Cost Function 107
E.A. TANNURI, C.P. PESCE, D.C. DONHA

Positional Game Control of Ship in Collision Situations 113

1. LISOWSKI

SEMI-PLENARY PAPER I

Research on Flexible Oscillating Fin Propulsion System and Robotic Fish 119
!. YAMAMOTO, Y. TERADA

SHIP PROPULSION AND ENGINE CONTROL

Observer-Based FDI for Gain Fault Detection in Ship Propulsion Benchmark - A Geometric Approach 125
T.F. LOOTSMA, R. IZADI-ZAMANABADI, H. NIJMEUER

A Nonlinear State-Space Model of Diesel Propulsion Plant Operation Using Neural Nets 131
N.!. XIROS, N.P. KYRTATOS

Marine Engine Room Monitoring and Control System for Simulating Real Processes on a Ship 137
R. SMIERZCHALSKI

UNDERWATER VEHICLES MODELLING AND CONTROL (INVITED PAPERS)

A 2-D Visual Servoing Technique for Underwater Vehicle Station Keeping 143
J.-F. LOTS, D.M. LANE, E. TRUCCO, F. CHAUMETTE

Self-Tuning Position and Force Control of an Underwater Hydraulic Manipulator 149

AC. CLEGG, M.W. DUNNIGAN, D.M. LANE

Docking Techniques and Evaluations Trials of the Swimmer AUV 155

K.M. KELLER, J.C. EVANS, J.S. SMITH

Predictive Control and Dynamic Planning of an Autonomous Underwater Vehic1e 161

R. SMIERZCHALSKI, M. KWIESIELEWICZ, M. SZYMANSKI, R. SUTTON

AUVS' Dynamics Modeling, Position Control, and Path Planning Using Neural Networks 167
H. SAYYAADI, T. URA

SHIP MANOEUVRING AND CONTROL

The Manoeuvring Characteristics on Tug-Towed Ship Systems 173

T. KISHIMOTO, K. KIJIMA

Study on a Decoupling Controller for Ship Harbour Manoeuvres 179

MINH-DUC LE

Concept of a Modem Manoeuvre Prediction System for Ships 185

H. KORTE, H.-D. KACHANT, 1. MAJOHR, T. BUCH, C. KORTE, M. WULFF

On a Ship Track-Keeping Controller with Roll Damping Capability 191

D. LOZOWICKA, A TIANO, A LOZOWICKI

MONITORING, FAULT DIAGNOSIS, AND FAULT

DETECTION IN MARINE VEHICLES

Classification and Detection of Changes in the Hydroacoustic Signals 197

E. SHPILEWSKI, M. SZPILEWSKI

vi
 On the Design of Alerts for Future Platfonn Management Systems 203
M.A. TAINSH

Neural Manoeuvre Detection of the Tracked Target in ARPA Systems 209

A. STATECZNY

Enhanced Neural-Network Modelling for Process Fault Diagnosis 215

T.K. CHANG, D.L. YU, D. WILLIAMS

PLENARY PAPER II

In Pursuit of Lean Manning: Ship Automation and the Value of Simulation 221
KJ. CARTLEDGE

APPLICATION OF GENETIC ALGORITHMS TO MARINE SYSTEMS

Genetic Algorithm Optimisation of Oil Tanker Control Systems 227

E. ALF ARO-CID, EW. McGOOKIN

Genetic Method of Optimization of Evacuation Ways in Cases the Fire Growth at Ferryboat. 233
D. LOZOWICKA, P. NIKONCZUK

A Genetic Algorithm for Autonomous Underwater Vehicle Route Planning in Ocean Environments with Complex
Space-Time Variability 237
A. ALVAREZ, A. CAITI

CONTROL OF SPECIAL SHIPS

Modelling and Control of a Trawl System in the Transversal Direction 243

V. JOHANSEN, O. EGELAND, A.J. S0RENSEN

On the Design of an IGV-Based Servo System for Air Cushion Vessels 249
S.M. SAVARESI, D. BERTIN, S. BITT ANTI

Simulation Model of the Shiphandling Training Boat "Blue Lady" 255

W. GlERUSZ

MODELLING AND CONTROL OF FAST SHIPS (INVITED PAPERS)

Experimental Study of Controlled Flaps and T-Foil for Comfort Improvement of a Fast Ferry 261
J.M. GIRON-SIERRA, S. ESTEBAN, B. De ANDRES, J.M. DIAZ, J.M. RIOLA

A Simulation Tool for a Fast Ferry Control Design 267

S. ESTEBAN, B. De ANDRES, J.M. GIRON-SIERRA, O.R. POLO, E. MOYANO

Decreasing of the Motion Sickness Incidence by a Multivariable Classic Control for a High Speed Ferry 273
J. ARANDA, J.M. DiAz, P. RUIPEREZ, T.M. RUEDA, E. LOPEZ

Control Code Generator Used for Control Experiments in Ship Scale Model 279
OR POLO, S. ESTEBAN, A. MARON, L. GRAU, J.M. De la CRUZ

A Research on Motion Smoothing of Fast Ferries 285

J.M. De la CRUZ, P. PEREZ de LUCAS, J. ARANDA, J.M. GIRON-SIERRA, F. VELASCO, A. MARON

MODELLING AND CONTROL OF UNDERWATER VEHICLES I (INVITED PAPERS)

A Sensor Fusion Scheme for Autonomous Underwater Vehicle Localization in Cable Following 291
A. BALASURIY A, T. URA

VII
Control of Limit Cycles for Underwater Vehicle Systems Using Describing Functions 297
Y.J. HUANG, Y.-J. WANG

Combined Trajectory Tracking and Path Following for Underwater Vehicles 303
P. ENCARNA<;::AO, A. PASCOAL

02CA2: A New Hybrid Control Architecture for a Low Cost AUV 311
P. RIDAO, M. CARRERAS, J. BATLLE, J. AMAT

AUV Sliding Mode Autopilot Optimisation Using Genetic Algorithms 317

E.W. McGOOKIN

SEMI-PLENARY PAPER II

Shipboard System Diagnostics and Reconfiguration Using Model-Based Autonomous Cooperative Agents 323
S. CHill, G. PROVAN, Y.-L. CHEN, F. MATURANA, S. BALASUBRAMANIAN, R. STARON, D. VASKO

DIFFERENT ASPECTS OF MARINE CONTROL

Wave Spectrum Correction with the Ship's Speed and the Incidence Angle 331
V. NICOLAU, E. CEANGA
"
On Board Artificial Stable Platform System Free from Roll and Pitch Motions 337
J. KAY ANO, K. OHTSU

Inertance Control of Underwater Installations 345

S.I. SAGATUN, T.I. FOSSEN, K-P. LINDEGAARD

SHIP AUTOPILOT CONTROL DESIGN

Ship Autopilot Tuning Using the Vishnegradskii Aizerman Diagram 351

D.CLARKE

Adaptive Self-Tuning PID Control Algorithm for a Yacht 357

C.M. XIAO, P.C. AUSTIN

High Performance PI(D) Controller on Ship Steering with a Conditional Reset Windup 363
R. FERREIRO GARCIA, C.A.F. AMEAL

Design ofPID Controllers for Ship Course Keeping Using Approximations to H_-Solutions 369
S. RUMY ANTZEV, A. MIROSHNIKOV, E. POPOV A

Predictive PID Control for Ship Autopilot Design 375

M.H. MORAD!, M.R. KATEBI

PLENARY PAPER III

Ship Control Technology; A US Navy Perspective 381

J. MOSCHOPOULOS

SENSORS, OBSERVERS, IDENTIFICATION AND CONTROL

On Global Model Based Observer Designs for Surface Vessels 389

K.-P. LINDEGAARD, T.!. FOSSEN

Model Identification of a Low-Speed UUV 395

P. RlDAO, J. BATLLE, M. CARRERAS

The Study of H_ Control theory on Ship Lift Feedback Fin Stabilizer 401
ZHANG XIAOYU, JIN HONGZHANG, LI GUOBIN, II MING

viii
Relay-Based Process Identification in the Frequency Domain 505
Y.G. WANG, W.J. CAI

SHIP TRACK KEEPING

Optimization of Program Tracks for Research Ship 511

V.M. AMBROSOVSKY, S.P. HABAROV

Implementation of a Real-Time Target Tracking Behaviour Using Video Sensors 517

V. lLA, P. RIDAO, J. De La CRUZ, J. BATLLE

Algorithms for the Ship Trajectory Planning and Its Tracking in the Collision Avoidance Process 523
Z. ZWIERZEWICZ

Author Index 531

x
Copyright © IF AC Control Applications in Marine Systems,
Glasgow, Scotland, UK, 2001

ENHANCED MARITIME SAFETY THROUGH

DIAGNOSIS AND FAULT TOLERANT CONTROL

Mogens Blanke *

* Automation at Oersted.DTU, Build. 326, Technical University

of Denmark, DK 2800 Lyngby, Denmark
e-mail: blanke@iau.dtu. dk

Abstract: Faults in steering, navigation instruments or propulsion machinery are

serious on a marine vessel since the consequence could be loss of maneuvering ability,
and imply risk of damage to vessel, personnel or environment. Early diagnosis and
accommodation of faults could enhance safety. Fault-tolerant control is a methodology
to help prevent that faults develop into failure. The means include on-line fault
diagnosis, automatic condition assessment and calculation of remedial action to avoid
hazards. This paper gives an overview of methods to obtain fault-tolerance: fault
diagnosis; analysis of properties of a faulty system; means to determine remedial
actions. The paper illustrates the techniques by two marine examples, sensor fusion
for automatic steering and control of the main engine. Copyright © 2001lFAC

Keywords: fault diagnosis, fault-tolerant control, marine systems.

1. INTRODUCTION Fault-tolerant control ((Blanke, 1996),(Patton,

Faults in steering or propulsion machinery on
a ship are particularly serious, since the conse-
quence could be loss of maneuvering ability, which
a risk to cause major damage to vessel, personnel
or environment. Faults related to steering include
faults in sensors and local rudder control, fault
related to prime propulsion include faults that
cause shut-down of ancillary or auxiliary sys-
tems or of the main engine. Alarms and simple
faults are fairly common on board ships and they
cause a stress impact on officers on the watch.
Human reaction under stress is sometimes erro-
neous and if manual intervention to silence an
alarm is erroneous, a local fault can escalate to
a hazard. Automatic detection and warning of
faults is hence desirable before they need urgent
attendance. Even better, simple faults could be
accommodated by autonomous action within the
automation systems, the aim being to maintain
availability of critical functions whilst issuing a
diagnostic message about the occurrence.
1997), (Blanke et al., 2000)) is a set oftechniques
developed to handle faults autonomously, increase
plant availability and reduce the risk of safety
hazards. The aim is to prevent that simple faults
develop into serious failures. To automate the
handling of faults, it is essential to have tools to
analyze the complexity of a case and determine
which remedial actions could and should be auto-
mated.
Fault diagnosis techniques have been the subject
of research over the last couple of decades and the
field has gained widespread international interest
and acceptance (Gertler, 1998), (Chen and Pat-
ton, 1999), (Patton et al., 2000). With an accel-
erated development, recent results include fairly
sophisticated design procedures that can assure
sufficient robustness to false detection. This is
crucial for automatic fault handling since false
detection could deteriorate overall reliability (Wu
and Klir, 2000).

1
The effects of faults critical to ships' safety are
required assessed by legislation of all countries
within the International Maritime Organization
(IMO) since the start of this millennium. This
means there is available information to support
analysis of faults and their consequences. This
paper contributes with an overview of methods
that can help design automation systems to diag-
nose and handle faults and contribute to enhanced
safety.
The paper focus on diagnosis and on-line handling
of faults whereas diagnosis for maintenance is not
within the scope. An overview is first presented of
tools to analyze structure and other fundamental
properties of an automated system in normal
and in faulty conditions. The aim is to find
inherent redundancy and utilize this to maintain
availability if faults occur. Methods for diagnosis
of faults are subsequently discussed, emphasizing
essential properties for hard real-time use. Two
examples illustrate the methodology, one is sensor
fusion for navigation, the other is main engine
propulsion control. Bibliographic references are
included in a final section of the paper.
2. REQUIREMENTS AND SAFETY
An essential requirement in manoeuvring of a ship
or operation of a machinery is that the transient
state that might follow due to a fault should not
exceed a specified limit. Such requirements are
vessel and operation dependent, extremes being
a fast ferry at 50 kt operating in narrow waters,
an off-shore supply vessel dynamically positioned
close to a rig, a container vessel at ocean passage.
Requirements specify the allowed erroneous ship
motion: a fault in the navigation or automatic
steering control should never cause a heading de-
viation larger than 5 degrees; a fault in equipment
servicing or controlling the main engine should
never cause an over speed shut-dow of the engine;
ships' speed should not exceed the desired by
more than 4 kt.; ships' position should remain
within a virtual box of dimension 0.7 (ship length,
beam); manoeuvrability could be required intact
or recovered within 15 seconds, as an example.
3. FAULT-TOLERANT CONTROL
Automation for safety-critical applications, where
neither function failure nor performance degrada-
tion could be tolerated, normally requires redun-
dant hardware to facilitate fault recovery. Fail-
operational systems are insensitive to any single
point component failure. Fail-safe systems make
controlled shut-down to a safe state when a sensor
measurement indicates a critical fault. In contrast,
fault-tolerant control systems, employ redundancy
2
in the plant and in the automation system to make
"intelligent" software that monitors behavior of
components and function blocks. Faults are iso-
lated, and appropriate remedial action is taken to
prevent that faults develop into critical failures.
The overall FTC strategy is to maintain plant
availability high and accept reduced performance
when critical faults occur.
Fault-tolerance can be obtained as follows. A fault
diagnosis schemes operates on-line. A discrete
event signal to a supervisor-agent is generated
when a fault is detected. This in turn activates
fault handling, which can be predetermined for
each type of critical fault, or can be obtained
from real-time analysis and optimization. Figure
1 illustrates a plant as a set of constraints, e.g.
differential equations, the structure of which is de-
termined by the state of an automaton. The logic
states can change as a consequence of changes in
continuous state or of the occurrence of faults.
A fault-tolerant control system employs diagno-
sis and autonomous supervision for handling of
faults. Reconfiguration (Blanke et at., 2000) will
change the use of interface between plant and
control system as a consequence of a new (faulty)
condition that has been diagnosed. Accommoda-
tion would change controller parameters but not
the interface with the plant.
Fig. 1. A plant is a hybrid entity. Changes in
state or faults can cause changes in structure
or in parameters of the constraints that de-
scribe the plant. Diagnosis and autonomous
supervisor functions implement appropriate
handling when a fault occurs.
3.1 Detailed Requirements
Requirements to limit ship erroneous motion are
closely related to the effect of the fault and the
structure of the control loop. The impact factor
of the fault is a function of control loop structure
and properties of the faulty system, magnitude of
the fault, time to detect and time to reconfigure.
An important parameter is the time to detect that
a fault has occurred, and the impact the fault has
during this time.
• If a component fault is abrupt and the di-
agnosis method is used for operator assisted
change of operational mode, the detection
time can be seconds to minutes.
• If abrupt faults in setpoint values to a closed
loop control are used by a controller for au-
tomatic accommodation, the time to detect
the fault can be within five to ten samples.
• If abrupt faults in a feedback element in a
closed loop control occur, time to detect and
reconfigure is one to two samples.
The categorization above indicated there is a need
to distinguish between the time-wise development
of a fault (abrupt or incipient) and the way in
which the information is used (from maintenance
to autonomous accommodation). A total fault
management system which may use robust design
methods provides fault tolerant control. Process
faults are only considered here as their effects
propagate through the control system and the
ways in which the control system can alter their
propagation.
4. FAULT MONITORING AND DIAGNOSIS
Fault diagnosis involves different steps to provide
increasing knowledge about the fault.
• Fault detection: a change from normal is
detected.
• Fault isolation: the location is determined
and perhaps also the time of occurrence.
• Fault estimation: the magnitude of the fault
is determined. In some cases, the time of
occurrence is estimated.
The term diagnosis is commonly used when at
least the first two tasks are accomplished.
Fault diagnosis is based on available input/output
signals of the process. When dynamic characteris-
tics of the system are known with a reasonable
degree of precision, said model is used to gen-
erate an estimate of expected output. The dif-
ference between estimated and measured output
are primary residuals. The primary residual is
asymptotically zero and within the envelope of
measurement noise in the fault-free case,.
5. DESIGN FOR FAULT-TOLERANCE
The first step in a fault-tolerant design is to de-
termine which failure modes could severely affect
the safety or availability of a plant. Analysis of
3
failure of parts of a system is a classical disci-
pline and the Failure Mode and Effects Analysis
(FMEA) is widely used and appreciated in the
marine industry. The traditional FMEA does not
support the analysis of fault handling, only their
propagation. In automated systems, 'the goal of
fault-tolerance is to continue operation, if this is at
all possible. An extended method for Fault Prop-
agation Analysis (FPA) can be employed using
an algebraic approach for propagation analysis.
The aim of the FP A is to show the end effects
of faults. It can also assist fault tolerant design
so that severe consequences are avoided (if the
system structure makes this possible). If the FPA
analysis finds that serious failure can occur, due
to certain faults, these can be included in a list
of fault effects to be detected. Whether this is
possible is disclosed in a later analysis of struc-
ture, that shows which redundant information is
available in the system. The final step is to de-
termine actions, preferably within the software of
the controller, that can accommodate the fault
and prevent the serious failures from occurring. To
determine whether this will be possible is a part of
the analysis procedures for the quality measures
used to judge the possible recovery of the system.
5.1 Analysis of Fault Propagation
For the reasons given above, fault analysis needs
to involve detailed analysis throughout the sys-
tem. In order to do this a component-based
method can be used, in which possible component
faults are identified at an early stage of the design.
The method uses FMEA tables for the description
of faults in components. In this context compo-
nents are sensors, valves, motors, programmable
functions etc. Programmable parts are considered
as consisting of separate function blocks that can
be treated similarly to physical components in the
analysis; bearing in mind that their properties
may be changed by software modification if so
desired. Considering terminology, a component
failure is considered a fault at sub-system level.
Failure at the subsystem level is what the system
strives to avoid.
An FMEA scheme shows how fault effects in
components relate to faults in inputs, outputs, or
in parts within the components.
 Abstract
This paper discusses a controller architecture that allows continuous
improvement, upgrading and expansion of complex control systems. Its success
has been proven in practice with the development of a DpDt system for
dredgers. The paper discusses the merits of the approach both for Industry and
for University. Copyright © 2001 IFAC

Keywords: DpDt, autopilot, controller architecture

1. INTRODUCTION As one would expect, these papers follow one other

Since the coming of the computer era, and even years
before that time, many papers have been written
about the problem of controlling the motions of a
ship. The emphasis continually shifted, following the
possibilities offered by new technology or just the
interest of those involved. When regarding the
seventies, one will recognise the shift trom analog to
digital systems. In the eighties, new control schemes
appeared that took full advantage of the opportunities
offered by computers. Also less common control
problems could be addressed such as controlling roll
and heading simultaneously with one rudder.
Together, these developments have resulted in the
currently available mix of control systems; autopilot,
stabilisers, Dynamic Positioning, Ride Control, etc.
Upon.. studying papers about ship. motion control,. one
easIly recogmses
. that they fall mto three categones
. d ·
• The.oretical papers about new control algonthms
(Gn~ble & others, 1994).
• Practical pap~rs about the performance of new
control algonthms (Van Amerongen & others,
1986) ...
• CommercIally mfluenced papers about the
performance of new control systems (Van der
Klugt & de Meijer, 1999)
up i~ time; a sub-set of newly developed control
algonthms shows enough merits to be tested in
practice and a sub-set of those tested will appear
eventually in commercial control systems.
However, it is doubtful whether that will be true in
the next decade. Pressure on companies to reduce
development time and to be the first on the market
with a new system or a new function is mounting. At
the same time, system requirements become more
diverse as technology allows customers to tailor ship
designs to their specific demands. Control systems
either become flexible enough to follow these
changing demands or have to focus on the relatively
dwindling market of 'standard' ships.
Th e R otter d am b ase d company Imtec h Manne . &
.
In d ustry h as a Iways concentrate on bemg a leader m
new technology, in particular in ship control systems,
rather than a follower. Therefore, this company has
felt more keenly than others the pressure to reduce
development time of new systems; preferably to the
time span of building a new ship. At first glance, this
seems to be utterly unrealistic. For example, the
development of an RRS autopilot (together with
,

21
Delft University of Technology and the Royal
Netherlands Navy) from scratch to sea acceptance
test took 9 years (Van der Klugt, 1991). Nevertheless,
as shown by the development of a DpDt system for
hopper dredgers, it can be done, conditionally.
One of these conditions is the focus of this paper: the
selection of an appropriate controller architecture.
This paper presents a controller architecture that has
been proven to be flexible enough
• to realise diverse ship control systems such as an
RRS autopilot and a DpDt (Dynamic positioning /
Dynamic tracking) system for different types of
ships such as frigates, LPDs and hopper dredgers,
• to allow embedding of parts that are supplied by
third parties,
• to reduce the development time of complex ship
motion control systems substantially and
• to realis~ a performance that will be difficult to
match wIth other approaches.
2. A UNIVERSAL CONTROLLER
ARCHITECTURE
2.1 Requirements
From the introduction, one might deduce that the aim
is to have a control algorithm for ship motion control
that can be easily reconfigured or modified to act as
any (new) type of control algorithm. To think that
such an algorithm is possible is of course an illusion.
Rather, the aim is to have an appropriate controller
architecture; a network of replaceable components
that together form the control algorithm. As this has
to work within a commercial control system, a
number of restrictions are in order that, rather than
making it more difficult, make it more easy to define
an appropriate architecture. A secondary aim,
formulated by Imtech Marine & Industry as a long
term goal, is to be able to control all relevant ship
motions. This goal further restricts the design
freedom:
• The architecture comprises 'recognisable'
replaceable parts.
The definition of 'recognisable' parts encompasses
more than such things as 'a filter' or 'a controller',
Many different types of ship have to be regarded with
many potential configurations of sensors and
actuators that, from the point of view of the
controller architecture, have to be treated as
replaceable parts.
..•
Complex parts are allowed, but It IS paramount that
..
they can be bUllt from smaller recognlsable parts.
Otherwise, a manufacture becomes too reliant on the
expertise of individual employees. A typical example
of that is the use of an 'optimal' controller. It is
important to have a direct relation between problem
(i.e. heading control of a specific ship), solution (i.e.
controller structure and gains) and optimisation (of
22
those gains). If such a relation exists, it will be
possible for a designer to quickly add modifications
or even replace this part by some alternative.
Otherwise, too much time will be lost to verify the
impact of the change on the overall design. Also, it
allows him to start simple, say with a PID control
algorithm for controlling a specific motion, to be
replaced with a better alternative as soon as that is
convenient. Finally, it enables co-designing; i.e. parts
are supplied by a third party and merely integrated.
An often-overlooked problem with complex control
systems is, that if there is a problem encountered in
practice, it will be difficult to find. Having
recognisable parts with familiar signals that can be
monitored (such as rudder, heading and filtered
heading to verify the impact of the rudder on the
heading or the heading on the filtered heading) is a
great help for finding the root of the problem.
• The architecture is built from parts that have
simple alternatives in today's technology.
When regarding a 'universal' control architecture for
ship motion control, one has to realise the complexity
of the underlying control problem. Potential motions
to be controlled are many. Heading, rate of turn, roll,
pitch, position, speed, depth are only some of them.
With al those different types of rudders and thrusters,
stabilisers, tanks, planes, etc. there are many potential
actuators. And, when zooming in on special purpose
ships such as dredgers or tugboats, there are many
other ways to exert a force or a moment to the ship.
It is no wonder that today's control systems focus on
a small part of the ship motion control problem.
Common are the autopilot (focussing on heading and
rate of turn), the roll stabilisation system, the ride
control system (focussing on the roll, pitch and
heave) and the dynamic positioning system,
If such parts are still recognisable within one
common controller architecture, it is possible to
design relatively quickly a control system that covers
all these control options. In a later stage, the
interaction between such individual parts (such as the
influence of a rudder on the roll or the influence of a
stabiliser on the heading) can be used to design better
alternatives.
Combining components of existing technology into
one system is well suited to satisfy the commercial
requirement: within a short time span a product on
the market, one that can be easily upgraded to meet
future demands.
The arch'Itecture can be rea 1Ise · d .
WIth the
'
standards 0 f to day s techn0 1ogy
From a commercial point of view, it is important to
adhere to the main standards for selecting
programming language, operating system and
communication protocol. System components have to
be world-wide available (service requirement) and
should meet regulation requirements.
In this diagram, the following holds:
• It comprises a number of sensors (heading, roll,
rudder angle, etc.) providing the data required to
calculate (filtering / estimation) the signals upon
which a controller has to act.
• The block 'control' calculates actions that are to
be distributed over the available actuators.
• An allocation algorithm distributes the control
outputs over the available actuators
• Filters, controllers and allocation algorithms are
adjuste? according to operator requirements given
the avai!able sensors en actuators.
• Data i~ ~xchanged by means of some
commumcatiOn layer . practice, these add 'states' to the system as well as
• Alarm management provides data about the state
of the system and its components to the operator
and for co~figuration ~anagement p~oses.
• The blo~k configuratiOn management selects the
appr~pnate sens~rs, actuato~s an~ control
algonthms dependmg on what is possible (for a
part indicated by Alarm Management) and what is
requested (by the operator).
• The 'adjust' blocks can be very simple (such as
changing the course control accuracy upon
operator request) or may comprise complex
adaptation schemes (such as to fmd optimal gains
with changing weather conditions. The same
applies for the other blocks; they can be very
simple or so complex that they have to be built
from a network of building blocks themselves.
The actual implementation is such that the actuators
and sensors can be either the actual systems on board
a ship or simulated alternatives. That way, most
testing can be executed in a (design friendly) office
environment.
The state-space domain provides filter and controller
solutions that match naturally the indicated
architecture. This is illustrated with the filter design
as used within the baseline DpDt system of Imtech
Marine & Industry.
23
The filter / estimator structure resembles a Kalman
filter structure. It shows model-based estimation of
the ship motions using 'known' actuator and sensor
inputs. From these ship motion estimates, sensor
estimates are derived. The difference between
estimate and actual value is used to update the filter
states.
When regarding the above diagram, one has to be
aware of some simplifications that would otherwise
clutter the diagram:
• The consequence of having coloured noise on
sensors is not shown neither is the colouring of
system noise other than a constant disturbance. In
coefficients that describe the colouring.
• Some sensors (dashed text) need coordinate
transformation to bring them to the ship
coordinates. In theory, it is irrelevant whether a
comparison between transformed sensor and
estimate or between sensor and transformed
estimate is used. In practice, one solution may be
more convenient.
The word :optimal' cannot really be used with
respect to this filter. The force model may have non-
linear coefficients that prohibit the calculation of
'optimal' update gains. In practice this is not a
serious drawback as the variances of the system noise
and measurement noise are generally not known.
Satisfactory settings can be found by following some
rules of thumb from control theory with respect to the
relation between time constants and the underlying
process. However, as indicated in Fig. I, the structure
does allow provisions for optimising such gains
(block 'adjust filter / estimator').
The above structure has been used in several
realisations ofImtech's DpDt control system:
• As part of the RRS autopilot for the M-class
frigate, the structure provides estimated roll (-
rate), heading (-rate) and waterspeed. The force
Fig. 3 shows a simplified model of the impact of
rudder and stabiliser setpoints on the roll moment and
the yaw moment. It is valid for a ship with one rudder
and one stabiliser set (i.e. two fin stabilisers). The
gains kop ... koo-2 are considered to be known
coefficients (part of 'force model' shown in Fig. 2)
that may be non-linear or may change in time.
The above model describes a part of the overall
problem of controlling roll and heading of a ship. It is
also the part that causes the greatest difficulties in
calculating an 'optimal' controller (i.e. the
interference between roll and heading control).
Furthermore, it is unconventional in the sense that it
regards the two fin stabilisers independently (usually,
both stabilisers operate in tandem). But this part is
well suited to illustrate the practical benefits of
regarding 'allocation' as an independent part of a
controller design (the appropriate allocation
algorithm for this example goes beyond the scope of
this paper):
1. Making changes in the actuator configuration
does not affect the controller design
Adding a bow-thruster or adding stabilisation
tanks merely extends the freedom of an allocation
algorithm to find the optimal set of actuator
setpoints. If an actuator fails, the allocation
algorithm will automatically introduce a different
set (if possible). In principle, redesign of the
controller is not necessary.
2. It IS relatively easy to coop with actuator
limitations such as rate limit and angle limit.
Upon reaching a limitation, an appropriate
allocation algorithm will automatically try to
compensate by adjusting the setpoints of the other
actuators.
3. It is relatively easy to introduce priority schemes
An allocation algorithm for the above example
could be realised such that rudder motion for roll
reduction will be introduced only when the
stabilisers reach one of their limitations. It is even
possible to change the priority scheme on line, for
instance to let the stabilisers compensate for
rudder induced roll during a turning manoeuvre
only. This does not affect the design of the
controller or the filter/estimator.
4. A conflict of interest between more than one
process output to be controlled can be clearly
recognised.
The allocation algorithm will recognise when it
reaches the boundaries of the system's capability.
For example, the difference between desired and
possible roll moment can be used as input to some
adaptation scheme (' adjust control' in Fig. 1) to
find a better compromise between controlling the
heading and the roll. This information cannot be
obtained as easily from comparing the difference
between actuator setpoint and actuator position.
5. It is possible to allow combinations of automatic
and manual control
25
The allocation algorithm will automatically adjust
the actuator setpoints if heading control is done
by the operator. If access to the rudder is denied,
it will use the rudder position to calculate the
stabiliser setpoints. Or, if that is allowed by the
operator, it may add a stabilisation component to
the rudder setpoint.
6. Other types of manual control are possible
Joystick control, in which the operator gives a
force setpoint in the ship's X-Y plane, is a typical
example; an allocation algorithm is responsible
for positioning the available actuators.
2.3 Practice
Just before the turn of the century, the first ships
sailed with motion control systems that are based on
the control architecture described above. Besides the
short time span to develop appropriate algorithms, a
number of other constraints had to be met
• In all cases, actuator capability is not guaranteed
to be sufficient under all operational conditions.
This implies that the control system has to be able
to operate continuously with limited actuators.
• The performance requirements were close to what
is possible in theory. In case of the hopper
dredgers, that would result In a substantial
improvement for track keeping In dredging
conditions.
• The performance had to be demonstrated within
the time span of the sea acceptance trials. Hardly
any time would be available for testing or
modifying algorithms.
• Actual systems had to be supplied with user-
friendly human interfaces that encourage the use
of such complex systems.
• Extensive functionality, especially in case of the
hopper dredgers. This is illustrated in Fig. 4
giving an overview of the baseline architecture of
the actual system. All different operational modes
of a hopper dredger had to be considered resulting
in a mix of manual and automatic control with
different requirements to be executed with
different sets of actuators and different control
schemes.
Applying the new controller architecture may be
considered as one of the prime reasons why these
projects have been concluded successfully. It was
possible to work with a team where each team
member was responsible for a priori defined parts of
the overall system. A simulation environment was
created including models of ship, disturbances,
actuators and sensors for testing purposes.
Subsequently, while executing tests with their
individual parts, the team members automatically
validated the contributions of each other. Thus, the
different design problems could be tackled in
parallel.
3. CONCLUSIONS
Universities have to be wary that a too strong
commitment to long time research will increase the
gap with companies. This is particular true in those
areas where companies face tough competition and
where a short development time of new products has
become essential. In case of control systems, focusing
on algorithms that have not a clearly recognisable
relation with the systems to be controlled is merely
increasing this problem.
The paper introduces a way to look at control systems
that may be used as a way out of this growing
problem. It is a way that has its roots in Object
Oriented Engineering and it demands adhering to a
certain controller architecture. It has been proven to
be beneficial for industry, but it seems to have merits
for Universities as well:
• A company that uses such a architecture will be
more lenient towards letting Universities research
potential improvements. Modifications that
adhere to this architecture can be easily embedded
within the existing system and therefore brought
quickly to the market. Upgrading complex control
systems can be done gradually over time.
• A University that adheres to this architecture
sacrifices some of its freedom to a more organised
form of research. But it will gain better facilities
to work in changing teams at complex control
problems. It will be more easy to compare
different solutions (and thus finding their merits)
without having to resort to simplifying the
processes to be controlled. And, most important,
it will enhance their success in finding companies
interested in their work.
A general-architecture for controlling all ship
motions is feasible.
The available structure includes the means to control
all relevant ship motions of a growing number of ship
designs. At this moment in time, Dynamic
Positioning, Track Control, Speed control and RRS
are examples of modem control options that are
available. Compensation is possible for wind, current
and dredging forces. The allocation algorithm
26
addresses rudders, bow thruster(s) and main thrusters.
It is expected that in future this will expand with
more types of actuators and control of the remaining
ship motions (pitch, vertical acceleration, etc.).
A substantial improvement of the development time
of complex control systems is possible!
Although not the only driving factor, the introduction
of the controller architecture above (and the
consequential parallellisation of the tasks that had to
be carried out) has been a key element in reducing the
development time ofImtech's baseline DpDt system.
The other key element has been the embedding of a
simulation architecture within the same control
system. The combination was crucial for developing
the actual systems for a number of hopper dredgers
and other ship~ ~imultaneously and in time.
The replaceabIhty property enabled the developers to
add quickly basic filter, control and allocation
algorithms (often re-use of existing technology) as a
starting point in a simulation environment prior to
defining the fin~l algorithms. Thus, in a v~ry early
~hase of t~e desI~, the customer could be Involve.d
In the ~roJect. D:sIgners cou!d take adv~tage of hIS
extensIve operatIOnal expenence and Increase the
chance of acceptance of the final design.
..
Contmuous Improvement of complex control systems
is feasible! ..
The .controller archIte~ture allows Improvement of
relatively small parts In a way that does not affect
other parts. Within Imtech's baseline, examples are
already many: othe~ control alg~rithms, ~dding a
stem thruster, replacing the allocation algonthm by a
more adv~nced s~h~me, etc: .
CommercIally, It IS a sIgnIficant advantage that
changes are confined and that resource consequences
(co~ts & m~npower) c.an be overseen. ~xtras are
eaSIer to estimate causing development nsks to be
reduced.
4. REFERENCES
Amerongen, J. van, H.R. van Nauta Lemke and
P.G.M. van der Klugt, Application of Adaptive
Control to Rudder Roll Stabilization of Ships,
Proceedings ACC, 1986.
Grimble, MJ., S.A. Carr and M.R. Katebi, Integrated
Ship Control using Hoc Robust Design Techniques,
Proceedings IEE, 1994.
Klugt, P.G.M. van der, Rudder Rolls Stabilization in
Hr. Ms. Karel Doorman, J. Nav. Eng. 33(2), 1991.
Klugt, P.G.M. van der and S.A. de Meijer,
Autopilots: the next step, Proceedings 12th SCSS,
1999.
allowing for a meaningful combination of manoeuvring
and seakeeping motions in the prediction of extreme
ship motions. Furthermore, convolution terms (so
called "memory effects") are being incorporated that
would improve the prediction of the behaviour of the
vessel in non-zero encounter frequencies.

Within the above framework the following sections of

the paper present the formulation of the proposed new
set of motion equations and the associated terms in the
equations. This is followed by the presentation of the
solution procedure for these equations. The
implementation of the proposed memory effects is also
described. The motion equations are then applied to a
purse seiner model, for which seakeeping test data for
the following and quartering seas are available, and
results are discussed.

2. EQUATIONS OF MOTIONS

Traditionally in both manoeuvring and seakeeping,

hydrodynamic and inertial forces are formulated in
terms of general body axes that rotate and translate
with the ship. The Horizontal Body Axes, which are
closely related to, but not a special case of General
Body Axes, is quite a common system and it has been
used in many other studies of ship manoeuvring that
include roll (Chislett, 1990, Eda, 1978, and Son et ai,
1981). Chislett (1990) also explained the use of this
system for both seakeeping and manoeuvring in 4
degrees-of-freedom. Hamamoto et af. (1993, 1992a, b,
1994) presented the application of the system in studies
of manoeuvring motion of ships in waves especially for
the study of capsizing motion and the dynamic stability
of ships in following and quartering seas.

Notice that in contrast with conventional methods, in

order to solve motions equations more accurately in
extreme seas, no assumptions are made for small
angles.

In deriving the basic equations of motion, normally

three different coordinate systems are used as shown in
Figure 1. The first is an Earth fixed system, defined by
O-;T\~.

The second is a general body axes which is fixed in the

ship with the origin G being located at the centre of
gravity of the ship defined by G-xyz. The third is the
Horizontal body axes fixed in the ship with the origin
at G and defined by G-x'y'z'.
5.2 Discussion
As mentioned earlier, the numerical model was
designed to estimate dangerous motions in following
and quartering seas. Therefore the results from the
numerical method can be assumed as reasonably
satisfactory.
In the first run (See Figure 3), the method
approximated the roll motion rather well apart from the
initial disturbance. This might be caused by the initial
values or by not having enough damping force during
this short period. For the yaw and rudder, the autopilot
model seems to be too stiff However, it catches up the
overall characteristics of the motion.
With the increasing speed and when the heading angle
is getting closer to astern, the model is suffering from
surf-riding. After about 8 s, the yaw angle violently
increased to port while the rudder controlled by the
autopilot responded to prevent this yaw motion.
However, despite hard starboard, the mcrease
continued for a while and here the centrifugal force due
to large wave-induced force and rudder force the model
rolled uncontrollably to starboard. As a result, the
model capsized due to broaching (See Figure 4.).
On the other hand, in the numerical model the
behaviour of the motion does not display the same
characteristics. When the model experiences wave crest
at amidships for a while, the roll increased when a
wave crest passed amidships and the model capsized at
the wave crest. This is due to the reduced restoring
moment at the wave crest amidships, which cannot
counteract the heeling moment. It is known as
capsizing due to loss of stability on a wave crest. The
stiff autopilot model might cause these differences.
Also, in the numerical model capsizing occured before
the actual capsizing in model tests.
changed. However, the propeller revolution is assumed
constant due to simplicity in the numerical model. Note
also that, wave forces are modelled in mean water level
instead of using the instantaneous wave surface. This is
important in calculating the restoring moment as
explained earlier, however if the GZ curve is obtained,
it is not required in such motions. It should be taken
into account in simulating 6 DOF motions and in
identifYingdangerous situations in lower speed regions.
6. SIMPLE NUMERICAL TEST INCLUDING
MEMORY EFFECTS
As mentioned above, first-order convolution terms (so
called "memory effects") are being incorporated that
would improve the prediction of the behaviour of the
vessel in non-zero frequencies of encounter. For this
purpose, a simple independent module is developed to
compute the retardation (Kernel) function (23) for the
ship. These computations will eventually be
incorporated in the equations of motions to investigate
the behaviour of the ship in non-zero frequencies.
Some of the results of the module are presented in the
figures below (See Figure 7-10). Damping values of the
ship were obtained by 2-D strip theory in 52 different
frequencies. In these figures, all the results are non
dimensional, for zero heading angle and zero speed.
Damping values have been made non dimensional as
follows,

With increased wave steepness and reduced wave

length to ship length ratio, the model undergoes
sinusoidal motions in the experiment while the
numerical model approximate well this behaviour apart
from some disturbances in roll motion (See Figure 5.).

When the speed increased, similar characteristics with

the second run are displayed in the model tests (See
Figure 6.). However, the model reaches the capsizing
limit earlier than in the previous case due to increased
wave steepness. The numerical model, this time,
estimates the motions rather well. The difference
caused by the stiff autopilot model continues. The
period of the numerical model is also slightly shorter
than the actual model runs.

When surf-riding occurs, the relationship between

propeller revolution and the ship forward speed can be
An improved numerical method is presented to identify
dangerous situations in following and quartering seas
with ships advancing in waves. Some comparisons with
model tests were undertaken demonstrating satisfactory
agreement.

In the light of the above figures, it is clearly understood

that the increase of wave steepness and the speed are
the governing factors in the high-speed region.

The results are still preliminary and the program is still

undergoing development. The experimental
investigations are required to study these motions and
also to provide a basis for developing and verifying
numerical methods capable of predicting broaching and
capsizing of ships in extreme astern seas. For this
purpose and for obtaining manoeuvring coefficients in
heave and pitch, which are important in non-zero
encounter frequencies, model experiments are planned
in the near future.

The autopilot model has an important effect on the

motions. Use of more sophisticated autopilot models
will be attempted.
 SHIP STABILITY MONITORING BY MOTION FREQUENCY ANALYSIS

R. Ferreiro Garcia (1), C. Antonio F. Ameal (2)

(1) Dept. of Ind. Engineering, (2) Dept. of Maritime Sciences

E.S Marina Civil, Paseo de Ronda 51- 15011. E-mail:ferreiro@udc.es
University of A Corona. Spain

Abstract: This paper describes a methodology to implement and operate an onboard expert
system to be applied on real time ship stability computation. The system comprises a data
acquisition system associated to a software based tool on a PC computer. Such instruments
measure roll motions and process the acquired data to give information on actual ship stability,
such as an alarm when the actual stability deviation exceed a predetermined limit value: Ship
stability is computed by means of an FFT algorithm associated to further filtering based in the
frequency domain analysis of ship motions associated to sea state parameters. Copyright ©
20011FAC
.
Keywords: Ship stability, Roll motion, Fast Fourier Transform, Low pass filter,

parameters and therefore cannot always guarantee

1. INTRODUCTION safety of navigation.

The basic methodology for most of the
methods to compute ship stability in any operating
condition, is the designer's information supplied as
"stability and model loading manual". The
information serves to define the stability
characteristics in typical and unspecified cases of
ship loading based on the data of masses and static
moments for the deadweight components with regard
to the influence of liquid cargo free surfaces.
As is well known, the actual weights of many
categories of standard cargo units, like containers or
trailers may differ significantly from those specified
in the terminal documentation.
Also important is the error of the basic data
on the cargo centre of gravity co-ordinates, that
together with errors in the necessary amount of
ballast contribute to show that the calculation
method of stability does not satisfy with permissible
Some methods to compute seagoing stability
have been developed. In (Brown and Witz, 1996)
roll motion records have been used to implement the
stability computation procedure by means of roll
decrement method. In (Haddara, et aI., 1994),
validation of the random decrement technique with
experimental data was also performed to extract roll
natural frequencies for a static stability assessment in
waves. Some other methods like solution, the
autocorrelation equation, the modulation function
technique, (Haddara and Zhang, 1994), (Nechaev,
1995) or other using signal processing, (Zeng, et al.
1999).
1.1. The Basic Motion Model
Modelling the motion of a ship in a seaway is
extremely complex and the full solution to the
problem has yet to be established, see (Okhusu,
1996).The difficulties lie m the non-linear

39
interaction between the rigid body structure of the
ship and the surrounding random fluid field.This
Issue IS further complicated if the structural
deformation of the vessel is taken into account when
considering its response, changing from a six degree
of freedom model to an infinite modes of motion
elastic body.
The motion considered in the ship is mainly
due to the wave disturbance, which in effect, is the
input to the system considered.
For a rigid ship travelling with constant speed
v at a arbitrary angle to sea waves, the resultant
motions in the six degrees of freedom are governed
by a set of second order differential equation, that for
all linear motions of the ship, has the following
standard form, (Okhusu, 1996).
(m + a)q(t)+ bq(t)+ cq(t) = Q(t)
where in equation (1),
q(t)={x(t),y(t),z(t), lft(t),O(t), !fI(t)] T, x(t), y(t), z(t), are
the surge, sway and heave linear motions and lft(t),
O(t) and !fI(t) are the roll, pitch and yaw angular
motions, and Q(t)={X(t), Y(t),Z(t),K(t),M(t),N(t)f
X(t), Y(t), Z(t) are the amplitudes of surge, sway and
heave forces and K(t), M(t), N(t) are the amplitudes
of roll, pitch and yaw moments; the components q(t),
are the corresponding responses.The matrix m
includes as its elements, the mass of the ship and the
principal moments of inertia.The elements of the 6x6
matrix a are a combination of hydrodynamics forces;
b is a damping matrix and c is a matrix which
account for hydrostatic contributions.
The system of equation (1), has been written
in the co-ordinate system associated with a ship.
System (1) could be simplified if only selected
motions were considered, like symmetric or vertical
motions, heave and pitch, or antysimetrical or
horizontal motions like roll, yaw or sway that usually
are uncoupled in such equation.
The system described by equation (1) can be
analysed by the classical methods of random
differential equations, associated to difficulties of
computed nature due the interactions of
hydrodynamic forces.
Because of such factors as free surface
conditions, viscous effects, geometric properties of
the hull, etc., in spite of the non-linear description of
ship motions is often necessary.To make the problem
tractable only one degree of freedom is considered
from the point of view of ship stability.For normal
ships, rolling is probably the most obviously non-
linear mode and it is also considered to be the
motion which can be most realistic treated in
isolation, related with the other modes of motion,
(Okhusu, 1996).
To obtain de initial stability characteristics of
the ship, like GM, or the KG, the height of the
gravity centre over the keel of the ship it is necessary
to estimate, first the normal or natural frequency or
roll period of the ship, through a relation between
roll motion, wave excitation, and other quantities
like ship roll inertia and roll damping moment, all of
them strongly dependent of the ship draft.
2. THE SIMULAnON MODEL
The problem for calculate the initial stability
based on instrumentation data , i.e. true KG of the
ship demands the realisation of three different
alternative tasks:
(1)
• ship stability calculation at the berth on the
basis of an operational heeling experiment.
• real time stability estimation by simulation
of ship under actual sea state from
estimated sea state and ship characteristics.
, • real time ship stability computation and/or
monitoring under actual sea state
conditions from roll motion response
analysis.
From ship characteristics, basically from draft
information, it is possible to obtain information
about metacentric height, KM, approximate roll
radius of inertia, as fraction of ship beam, and the
value obtained from ship cargo documents about the
vertical position of the centre of gravity, KG. It
could then give an approximately idea of ship roll
period value, based in the knowledge of ship mass
displacement, but this single calculation could have
accumulated an excessive amount of error to be safe
and useful
In the actual state of marine technology,
vessels like containerships and ro-ro, could easily
implement onboard devices destined to measure roll
motions in waves as well as data processing
equipment to calculate real time initial ship stability
values, by means of frequency domain analysis of
the information gathered from ships motions related
or not with sea state parameters.
The illustration In figure 1 shows how
stability GM is estimated by simulation of ship
motion and frequency domain analysis of roll motion
under given sea state and ship characteristics,
(Ferreiro and Ameal, 2000).
40
 THE ANALYSIS OF A SHIP FUZZY DOMAIN IN A RESTRICTED AREA

Zbigniew Pietrzykowski

Maritime University ofSzczecin

Waly Chrobrego 1/2 str., 70-500 Szczecin
zbip@wsm.szczecin.pl

Abstract: The paper presents an analysis of the fuzzy domain of a ship in a restricted area. The
notion of the ship fuzzy domain has been defined. Also, a method of its determination has been
presented. Boundaries and areas of the ship fuzzy domain have been analysed for various levels
of the navigational safety. Copyright ©2001 IFAC

Keywords: Navigation systems, Safety analysis, Artificial intelligence, Fuzzy logic,

Neural networks

1. INTRODUCTION intimate, private, social and public. This approach

Evaluation of a navigational situation plays an
important role in decisions made by the navigator.
This kind of evaluation is affected by a number of
factors, such as ship parameters, area parameters,
traffic situation (other ships), hydro-meteorological
conditions and others. In most shipboard systems the
situation, based on previously defined criteria, is
evaluated as safe or dangerous one. This means that
certain intermediate states, e.g. less / more dangerous
situation are omitted. Identification of how the
evaluation tends to change is thus impossible, so that
an imminent danger cannot be signalled in advance.
For example, criteria such as the closest point of
approach (CPA) and time to closest point of
approach (TCPA) are used. A situation is identified
as a collision situation when the closest point of
approach (CPA) is not maintained. Attempts are also
made to use the ship domain as a safety criterion. The
area is most often determined by means of statistical
(observations) or analytical methods (systems of
equations). Due to a large number of factors -
including the human factor - that must be accounted
for, one may have doubts as to the correctness of ship
domains thus determined. These doubts are even
larger in the case of restricted areas where ship
manoeuvring possibilities are significantly decreased.
The approach in which intermediate states are taken
into account enabling graded evaluation reflects the
findings of the psychological analysis of human
behaviour in which diverse areas are distinguished:
may be imposed on the ship movement situation
(Zhao, et al., 1993). Consequently, the evaluation of
a navigational situation has to have a more graded
range. It should be emphasised that the boundaries of
particular areas (evaluations) are imprecise and
equivocal which results from the nature of the
phenomenon and the human behaviour. Therefore,
attempts have been made to apply that more flexible
approach. The major problem one encounters while
building a system of navigational situation
identification and safety level evaluation is the
creation of an appropriate knowledge base. For its
acquisition and representation tools of artificial
intelligence are applied. These include fuzzy
systems, neural networks, genetic algorithms, expert
systems and hybrid systems utilising positive
properties of the mentioned tools.
2. THE IDENTIFYING AND ASSESSMENT OF
ANA VIGA TIONAL SITUATION
2.1 The ship domain
The notion of ship domain was defined in the 1970s.
It is used as a safety criterion in an analysis of
a navigational situation. The significance of ship
domain grew when the equipment and systems of
automatic ship steering were introduced.

45
 AUTOMATIC SHIP BERTHING USING PARALLEL NEURAL CONTROLLER

Namkyun Im*, Kazuhiko Hasegawa**

*Graduate school of Osaka Univ., Osaka, Japan

Email: im@naoe.eng.osaka-u.ac.jp

**Graduate school of Osaka Univ., Osaka, Japan

Email: hase@naoe.eng.osaka-u.ac.jp

Abstract: In this paper a parallel ANN(artificial neural networks) for the automatic berthing will
be discussed. This controller has a separated hidden layer each control an engine and a rudder
respectively. Using this controller simulations were carried out where the initial conditions such
as ship's positions and heading angle are different from teaching data provided. Finally
comparison of separated hidden layer and united hidden layer will be described. Copyright ©
20011FAC

Key words: parallel neural networks, automatic controllers, artificial intelligence,

expert systems, fuzzy control.

1.INTRODUCTION pattern and then produces similar output with the

Automatic berthing control is one of unresolved
problems in ship control. Many methods and theories
were adopted to achieve the goal. A typical example
is feedback control, which has been used as a
controller in some research. (K. Kose et al.,1986).
Even if conventional feedback controller is great tool,
ship's berthing is so complicated that a lot of
limitations still are found. Therefore many studies
(Koyama.1987, Yamato.1990&1992, Hasegawa.1993,
Itoh 1998) suggest other controllers such as fuzzy
theory, neural network, and expert system. A neural
network controller is one of them. As it is known, a
neural network has good ability to recognize the
pattern. The feature was used to control a ship in
berthing problem. Yamato (H. Yamato, I 990)
suggested automatic berthing by a neural network
and produced excellent results, but he soon changed
his research field to others such as expert system (H.
Yamato,1992). So deep study for neural networks
controller was not conducted. Hasegawa (K.
Hasegawa, 1993) took over the study. Excellent
results were produced even if many things should be
overcome: general simulations have same initial
value with teaching data and oscillation phenomenon
occurred in controller. In other hands, when you look
to the existing research that is done for the automatic
berthing, it is found that main concept is based on

51
d2 is the remaining distance to the goal point. These
factors can explain the fact that navigators usually
make imaginary line to goal point under berthing
work. Even if ; ,1] can explain the location
information, they are not enough to explain the ship's
location information.
3. PROCEDURE FOR LEARNING AND
MAKING TEACHING DATA
3.1 Preparing for the teaching data
This paper focuses mainly on how a parallel neural
network will work effectively. So authors didn't try
the automatic berthing problem using a tug or side
thruster. This problem will be a next challenge.
Automatic berthing mentioned here means that a ship
stops near the goal point within O.2m1sof speed and
between 250-270deg of heading angle. Basically 6
sets of teaching data were adopted as like the Fig. 4-1
where disturbance is not considered. 4 sets of
teaching data are also adopted to compare their
effectiveness with 6 sets of teaching data. Fig. 4-2
shows 4 sets of teaching data.
3.2 Procedure of training
Popular method of training, back propagation, is
adopted. The Neural Network Toolbox from the
MATLAB package has been used to train the model.
Just variable learning rate method is used to minimize
time taken. Generally learning rate is held constant
throughout training. If the learning rate is set too high,
the algorithm may oscillate and become unstable. If
the learning rate is set too small, the algorithm will
take too large time to converge. So variable learning
rate is adopted here. Since a neural network in this
paper have separated hidden layers, training is needed
twice to make one set of weight and bias which will
produce one set of output units. The group for rudder
and the group for engine are trained separately. One
example of these training is presented in Fig. 5. The
figure explain very well the before and after training,
where the circle shapes indicate the original data,
teaching data, triangle means before training data and
square means after training data respectively. It is
easy to understand how much training is conducted
well by this figure.

4. SIMULATION RESULTS
4.1 Effect of parallel hidden layer control
Fig. 6 shows comparisons with the parallel hidden
layer and conventional layer. The red line (heavy
line) is the results of separated control. The thin line
is the results of united hidden layer. Cross marks
indicate initial positions of the teaching data used for
controller. This figure shows that even if the ship's
states of two models are same in the initial stage, the
output of the rudder and the engine are different
because of the different hidden layer design. It is easy
to understand that the parallel hidden lay~r has .more
stable and corrective control than the umted hIdden
layer by these figures. Take a look at the last case of
Fig. 6,even if both cases failed successful berthin~,
the parallel NN is showing its improve~ent .Ill
stopping ability and keeping track companng ':Ith
normal NN. Especially in the case oflast two ofFlg.6,
even if weights and bias of Fig. 4-2 are used, in other
words, starting positions and states are different with
training data provided, but the parallel hidden layer's
results is showing good ability comparing with that of
united hidden layer.
4.2 Simulations having the same initial conditions as
training data provided
Even if 6 sets were simulated, only 3 sets are
presented as examples here. As it is shown in Fig. 7,
the ability of stopping near the wharf and seeking a
goal point is very good. Especially the stopping
ability is good. It was possible to end within O.2m1s at
the wharf in all the cases. Also the final heading
angles were within 250-270deg. The details are
55
shown in Fig. 7.
4.3 Results having different conditions with
training data provided and Funnel Effect
In this section, the funnel effect will be discussed.
Many simulation which have different initial
conditions with teaching data provided, are presented
in this section. Fig. 8 shows the details. While the No.
1 represents simulation case where the same initial
data as the teaching data are used, the No. 2 has
different initial data with teaching data, but they
carried good automatic berth. Figs. 10 shows the
results where cross marks indicate initial positions of
the teaching data. It is easy to understand with these
figures that successful automatic berthing has been
accomplished even if they have different initial
conditions and different starting point with the
teaching data provided. This is due to the
interpolation ability of neural networks. ANN has
great power of interpolation to solve a faced probl~m
even if the situations are different with the teachmg
data. In Fig. 8 the square is marked. It is an area,
which guarantees safe automatic berthing to the
wharf from that area. The authors would like to call it
the 'funnel area'. For example, when an object comes
into the entrance of a funnel, the object should reach
the opposite side of the funnel without escaping from
the funnel passage like in Fig. 8. In this paper all
simulations, which are done at the square area, are
finished with good automatic berthing like Fig. 7 and
10. This funnel effect suggests the possibility that if
more of these funnel areas are established like in Fig.
9 automatic berth can be realized from every
direction and every distance.
 5. CONCLUSIONS AND DISCUSSIONS
K. Kose et al., "On a Computer Aided Maneuvering
In this paper, a parallel ANN for ship berthing was System in Harbors", Journal of the Society of Naval
discussed. In this paper, two groups of input units are Architects of Japan, Vol.160, Dec. 1986, p.l 03-11 0
considered to compose a parallel NN where hidden (in Japanese)
layers are split into two. The first group consider all
of input units and anther group include only two K. Kose et al., "On a Mathematical Model of
input units such as remaining distance, d2, and a ship Meuvering Motions of Ships in Low Speeds", JSNA
speed, u. But it can be said that additional research is of Japan, Vol. 155, June 1984, p 132-138 (in
needed to determine how input units should be Japanese)
separated. Conclusions of this paper can be drawn as
follows T. Koyama and Y. Jin, "A Systematic Study On
1) Newly designed ANN was used in automatic Automatic Berthing Control(151 Report)", Journal of
berthing problem as a controller. the Society of Naval Architects of Japan, Vol. 162,
2) A parallel ANN has good control ability December 1987, p.201-21O (in Japanese)
comparing with normal ANN of united hidden
layer.
3) The funnel area suggests the possibility that
automatic berth from every direction can be
accomplished.
4) Successful berthing has been accomplished even if
under different initial condition and different
starting point with teaching data provided.

6. REFERENCES

Hiroko Hoh, "Berthing Control with Multi-Agent

System", Journal of the Society of Naval Architects
of Japan, Vol. 184, Dec. 1998, p.639-648 (in
Japanese)

H. Yamato, T. Koyama and T. Nakagawa, "Automatic

Berthing using Expert System", Proc. Of Workshop
on Artificial Intelligence Control and Advanced
Technology in Marine Automation(CAMS'92), p.
173-183, Geneva, Apr., 1992

H. Yamato et al., "Automatic Berthing by the Neural

Controller", Proc. Of Ninth Ship Control Systems
Symposium, vol. 3, pp.3.183-201, Bethesda, U.S.A.,
Sep.,1990

K. Hasegawa and K. Kitera, "Mathematical Model of

Maneuverability at Low Advance Speed and its
Application to Berthing Control." Proc. Of The 2nd
Japan-Korea Joint Workshop of Ship and Marine
Hydrodynamics", pp.144-153, Osaka, June 1993.

K. Hasegawa and K. Kitera, "Automatic Berthing

Control System Using Network and
Knowledge-base", Journal of Kansai Society of
Naval Architects of Japan, Vol. 220, Sept. 1993
p.135-143 (in Japanese)

57
AN ELECTRONICAL AID FOR MANOEUVRING SHIPS IN RESTRICED WATERS

Reinhard Muller, Michaela Demuth

Maritime Department, University of Wismar, Germany

Abstract: Manoeuvre procedures which use rudder, propeller and thrusters are very
complex in which the crucial criteria for the feasibility of a manoeuvre is the navigational
space. Currently used automatic track systems configure successful way-point determined
routes although no technical support is given if a mix of rudder and engine manoeuvres are
necessary in harbour waters. The main project goal is to extend the function of the
automatic track systems. The first investigation involves developing a useful algorithm
which is carried out in the project. Examples of newly applied simulated automatic tracks in
harbour waters are given and discussed. The following paper describes the principles and
methods for developing an assistance system for the aid of navigation in harbour waters.
The project is supported by the Federal Ministry for Education, Science, Research,
Technology and Building. Copyright © 2001 IFAC

Keywords: path-prediction, helm order profile, restricted navigational area, decision

support.

1. INTRODUCTION In order to create a system for the manoeuvring and

Experiences are necessary for handling a ship in
harbor waters. Birthing maneuvers are influenced
by many different facts which characterize the
current situation. The captains knowledge and
experience access the success and efficiency of
maneuver procedures in restricted waters. The "art"
of navigation is to find out the appropriate
strategies for maneuvering in the current situation.
The bordering profile and the depth of the
navigational space in relation to the ships
dimensions influence the combination of the
various commands for the helm engine and
thrusters. To bundle the captain's experiences as
sets
... of different commands is one goal of the
mvestigatIO?.
The follow~ng p~per descnbes ~e ~evelopm~nt ?f
an e~ectromcal aid for the positiOnmg of ships ill
restncted waters.
2. PROBLEMS
The activities of manoeuvring procedures are the
analysis of navigational data, track planning and the
execution of helm and engine orders. For optimal
manoeuvres well defined alternations of course and
speed must be realised within a short amount of
time.
positioning of a ship it is necessary to find valid
paths and generate orders so that the ship can
follow the advanced predicted path. The
movements of the ship have to be controlled in
respect to the predicted data and the goal of the
manoeuvre. An experienced captain is a
prerequisite for maneuvering a ship successfully.
On modem ships bridges automatic course keeping
controller and path keeping systems are used. A set
of way points, given by the captain, defines the
continuous track which is then executed by the
automatic track system. If the next way point is not
reachable on a continuous track, the track system
cannot find a solution.
In the mentioned track sy stems the theoretical
approach doesn't contain astern ships movements.
But it is necessary to operate with negative speed
vectors in harbour waters or in the piers vicinity.
We have to introduce points of return for a
theoretical approach which include positive and
negative speed vectors. The kinetic energy of the
ship in points of return has to equal zero. For
finding out a point of return, a selection of simple
or combined manoeuvres and a scaling of them, has
to be carried out in advance.

59
 4. PRACTICAL TESTS

Further experiments will be prepared at the

Maritime Simulation Centre Wamemiinde
(MSCW) to improve the reliability of the algorithm
and to tune its heuristics.
The MSCW offers the possibility of a virtual and
dynamic model vessel to couple directly with the
test algorithm for analyzing.

5. CONCLUSION

The computed optimised solution contains the

inverse steering, path length, ship handlings which
are dynamically adapted to the environments. In
general, the dynamic manoeuvring characteristics
of the focused vessel will be respected. The
algorithm is also useful in open waters. The
published system is prepared for integration into
existing systems as additional support.

In the final version the described assistance system

could be used as technical support for navigating in
restricted waters. It would serve as "red line" for
the experienced captain or a basic plan for a young
Officer On Watch. The safety of decisions and the
access of the path findings would be optimized.
On the other hand the system would be very helpful
for preplanning new berthing procedures or
passages of narrow waters by the nautical staff.

63
 AUTOPILOT DESIGNED WITH FUZZY SET THEORY

A. Zirillil, A. Tianol, G.N. Roberts2 and R. Sutton3

I Departmentof Information and Systems, University of Pavia, Pavia, Italy Via Ferrata I, 1-27100 Pavia, Italy.
Also: Institute of Ship Automation C.N.R Via de Marini 6,1-16149 Genoa, Italy.
2Mechatronics Research Centre, University of Wales College, Newport Allt-yr-yn Campus, P.O. Box 180,
Newport, NP205XR, UK
3Department of Mechanical and Marine Engineering, University of Plymouth, Drake Circus, Plymouth, PIA
8AA, UK.

Abstract: In this paper, the design of course-keeping autopilot, is based on the cognitive modelling
approach and fuzzy set theory. Using this modelling approach, the aim is to design a control
system based on a model of the expert, who is able to specify the general properties of the system.
rather than on a model of the system to be controlled. After an heuristic description of the desired
manoeuvre characteristics, fuzzy set theory is used for the synthesis of the autopilot. Simulations
on a non-linear model of a containership are presented for course-keeping manoeuvres in the
presence of wave disturbances.

Keywords: Fuzzy system, fuzzy controller, direct adaptation.

1. INTRODUCTION developed by emulating certain characteristics of

The introduction of automatic autopilots for steering a
ship can be traced back to the 1922 with the
pioneering work of (Sperry 1922) and (Minorsky
1922). While Minorsky's contribution was of more
theoretical basis, leading to the celebrated
proportional-integral-derivative (PID) controller, the
major contribution of Sperry was of a practical
ground. In fact, as emphasised in (Bennet 1979),
Sperry observed that an experienced helmsman would
also, "meet" the helm, that is, back off the helm and
put it over the other way to prevent the angular
momentum of the ship carrying it past the desired
heading. Therefore he proposed the use of an
"anticipator" to build into his autopilot the "intuition"
of an experienced helmsman. The name "Metal-Mike"
give to the new automatic system by an officer of the
ship which had been used for the trials, emphasised
both the purpose (replace the helmsman) and the
origin of the proposed system.
The fast development of small and inexpensive
microcomputers and advances in computing
technology have fuelled the so-called "Intelligent
Control" theory, in which control algorithms are
intelligent biological systems (i.e. learning and
adaptation). Within the framework of intelligent
control systems, neural networks and fuzzy logic
systems have been widely used for the design of more
sophisticated and reliable control systems (White and
Sofge 1982). The majority of these applications define
the control design problem in the framework of
approximation theory, where a control mapping
relating the desired and measured performances with
the control inputs is defined. The universal
approximation capability of neural networks and fuzzy
logic systems is used to guarantee the existence of the
desired control mapping while adaptation and learning
algorithms are used to guarantee the optimality of the
approximation.
A different way for the synthesis of fuzzy based
controller, is to use a cognitive based approach. In this
instance, the aim is to design a control system based
on a model of the expert, who is able to specify the
general properties of the system, rather than on a
mathematical model of the system to be controlled.
The control strategy is then specified by a set of rules
deduced by a-priori knowledge of the system, that
constitutes the rule knowledge base of the controller.

71
Based on this stored knowledge, the actual situation is
evaluated in order to infer the appropriate control
action. The deduced control action, performed by the
so-called inference machine, is based on fuzzy logic
where uncertainties are easily handled.

By following the cognitive based approach, a heuristic

description of the manoeuvres is given and fuzzy set
theory is used for the design of a course-keeping
autopilot for ship. Although, with the above approach,
the interpretability of the overall system is still
preserved the tuning of the controller parameters is
somewhat heuristic and is mainly based on trials and
errors therefore the optimality of the overall system's
performances is not always guarantee. The heuristic
description of the manoeuvre is interpreted by means
of Fourier transform. The resultant autopilot is tested
for different sailing conditions in a simulation study
involving the non-linear model of a containership. It is
shown that the proposed fuzzy autopilot is able to
steer the ship acceptably well and represents a viable
control structure for further implementation of
adaptive and learning algorithms.

2. FUZZY AUTOPILOT

In order to design an optimal controller, performances

index have to be defined. In this respect. for a
particular sailing condition different factors may be
considered: 1) economy (fuel consumption), 2) safety
(related to accuracy and manoeuvrability), 3) user
predilections. In the design of steering control
systems, it is common practice to distinguish between
two main different mode of sailing. These modes are:
course-changing and course-keeping.

Course-changing: During course-changing

manoeuvres the ship heading angle is changed in such
a way that the ship can sail in the new direction
specified by a new (desired) heading angle. During
this manoeuvre it is possible to distinguish between
three main phases as: 1) the start of the manoeuvre,
where for safety reasons, the intention of the
manoeuvre must be clearly indicated to others ships,
2) a stationary turning, characterised by a constant rate
of turn, and 3) the end of the manoeuvre, where it is
important to control the overshoot of the manoeuvre in
order to avoid dangerous paths (Amerongen and
Naute Lemke 1980).

In each of the above phases, the helmsman will base

the control action on heuristic judgement of the
following information:

1. the new desired heading angle, given by the officer

of the watch,
2. the actual heading angle, given by the compass,
3. the ship turning rate, judged from the compass or
from observing fixed points in the shore or clouds.

Based on the above observations a reasonable choice

for a fuzzy course-changing controller is the PD-like
 FAULT-TOLERANT CONTROL BASED ON ADAPTIVE NEURAL NETWORK

T K Chang and D L Yu

Control Systems Research Group, Liverpool John Moores University, UK

Abstract: A fault-tolerant control (FTC) approach based on neural network (NN) inversion method
is proposed for multi-input, multi-output (MIMO) unknown non-linear dynamic system. An
adaptive multi-layer perceptron network (MLPN) on-line learns the system non-linear dynamic,
including behaviour of system with actuator or component faults. This MLPN is inverted based on
the Extended Kalman-filter (EKF) to estimate the appropriate control action to the non-linear
system. The stability of the NN inversion is proved with Lyapunov method. The results of FTC
application to a continuous stirred tank reactor (CSTR) process simulation show that the controlled
faulty system maintains the control performance and stability. Copyright © 2001 IFAC

Keywords: Fault-tolerant systems, adaptive control, non-linear systems, model inversion, Extended
Kalman filters.

1. INTRODUCTION FTC method is an accurately fault detection and

High-risk applications, such as nuclear and avionics
industry, require fail-safe operation, i.e. systems that
can withstand any single point failure without
effecting system operation. Also most of the modern
industrial plants are extremely complex and they
often include many kinds of systems which
compensate for the effects of faults, c;mponent
malfunction and misoperation. This requires
solutions that are very costly in both hardware and
development effort. The expense of such technology
is prohibitive for ordinary industrial automation
where implementation must be cheap. Nevertheless,
t~e ~nhancement in availabili~ and s~fety should be
sIgll1fi~ant. ~herefore, FTC IS very Important from
the vIewpOInt of safety as well as reduced
manufacturing costs. FTC offe~s avoiding inadverte~t
~rocess sh~t down s from sImple faults ' e.g. In
InstrumentatIOn an dcontro 1 1oops t hat cou ld d evelop
.. ~ .
1 laIlures.
Into productIOn stoppages or pant
The objective of FTC is (a) to improve the system
reliability or stability, and (b) retain acceptable
performance under system faults (Patton, 1993). All
present FTC methods can be classified into two
groups (Zhou and Frank, 1998):
First group is based on fault detection and isolation,
the essence of fault detection and isolation based
FTC methods is to detect and isolate system faults
on-line, and then to modify, redesign or reconfigure
the control laws to make the faulty systems stable
according to the isolated faults. The limitation of this
isolation is significant.
...
The s~con~ group IS Independent of fault detectIOn
and IsolatIO~. Some papers called fixed FTC
met~ods, whIch use fixed controllers and do not
consIder whether the fau~t ha~ occurred or. not.
Include methods such as IntegrIty control, relIable
stabilisation and si~ultaneous stabilisation, .. which
have b~en proposed In recent .years, see (FuJIta, M.
and ShImemura, E., 1988; JOShI, S.M., 1987). Clearly
it is a very difficult task to design such ~xed
~ontr?ller~, hence up to the. present only tIme-
Invanant lInear systems are consIdered.
In paper (Wang and Wang, 1999), a NN-based fault
compensator is proposed to modify the control
variables With the residual signal the fault
compens~tor able to modify the contr~l variables,
and drive the process track the desired trajectory,
~ ·
even lau It IS occurs.
In this paper, a new FTC approach is proposed. The
control variables are estimated based on an adaptive
MLPN, which on-lines learn the process dynamics.
Therefore, under actuator or component faults, the
proper control variables will be obtained based on the
adaptive MLPN, which has learns the fault behaviour
in the process. In order to achieve an accurate
adaptive MLPN, an EKF-based on-line learning
algorithm for MLPN is developed based on paper
(Iiguni, et al., 1992). The derivative of MLPN with
respect to the weight matrix is derived to improve the
modelling accuracy. A stable inversion algorithm of
MLPN is developed based on paper (Chang et al.,

77
 NEURAL NETWORKS FOR SYSTEM IDENTIFICATION OF COUPLED SHIP DYNAMICS

P. MartinI, M. R. KatebiI, I. Yamamoto2, K. Dai§o2,

E. Kobayashe, M. Matsuura2, M. Hashimoto2, H.Hirayama , N.Okamoto3

lIndustrial Control Centre, EEE Department, University OfStrathclyde, Glasgow, Scotland.

2Nagasaki R&D Centre, Mitsubishi Heavy Industries Ltd., Nagasaki, Japan.
3Technology Research Centre, Japan National Oil Corporation, Chiba, Japan.

Abstract: System identification of coupled ship dynamics is problematic with standard

least squares methods due to the non-linear, multivariable nature of the system. Neural
Networks have therefore been applied to this problem, as they are particularly suitable for
approximating non-linear, multivariable functions. In this paper, results of identification
with Neural Networks are given for a ship motion simulation based on a standard
mathematical model, and for real data collected from a l/50th scale model of the system.
The method is seen to be successful at various operating points, and ideas for extension
of the work are discussed. Copyright © 2001 IFAC

Keywords: Marine Systems, Neural network models, Identification.

1. INTRODUCTION position signals for three degrees of freedom of the

Floating Production, Storage, and Offioading (FPSO)
vessels are required to regularly transfer oil to a
Shuttle Tanker for transportation to the mainland.
During the offioading phase, the two vessels are often
coupled by a hawser, leading to complex dynamic
interactions. This paper details a recent investigation
into the use of neural networks for modelling such a
system.
The manner in which neural networks are used is
briefly explained. The mathematical model used to
both analyse ship motion and to produce a computer
simulation is given, and it is noted that this model
suggests the inputs and outputs to be used with the
neural network. The network is only used to
approximate the dynamics of the shuttle tanker in this
paper, but a combination of two networks would
model the entire system.
An initial attempt at training neural networks based
on data from the computer simulation is presented.
Results and figures are given to illustrate the
effectiveness of producing acceleration, velocity, and
ship.
Results are then given using data supplied by
Mitsubishi from their l/50th scale model tank tests.
These results are the most important as they give an
idea of the realistic potential of neural networks for
ship identification, which appears to be fairly high.
The results given in the main body of the report are
discussed, conclusions are drawn, and ideas
suggested for further work.
The system in question is depicted in Figure 1. The
circled numbers are (1) Hawser Angle with FPSO,
(2) Relative Position of FPSO and Shuttle Tanker,
and (3) Hawser Angle with Shuttle Tanker.
A turret towards the front of the vessel anchors the
FPSO, and control comes from a stem azimuth
thruster. The shuttle tanker is pulled away from the
FPSO by a tugboat, hence keeping the hawser in
tension, and control comes from a bow thruster. The
shuttle tanker also has a rear propeller and a rudder,
but these are not used when the hawser couples the
two ships.

83
The above Figures, 12, 13, and 14, demonstrate the
effective generalisation by the network for the final
third, after 3470 seconds. Similarly high quality
results have been obtained using another eight data
sets, but only individually. It was originally hoped
that a single network could be trained using each of
the data sets in turn, in order to generate a network
capable of providing an approximation to each of the
sets. However, this has not been the case, as re-
training a network with another set tends to degrade
the performance of the network acting on the
previous set.

The network was also trained with three similar data

sets where wave height was 0, 3, and 4m
respectively, before validation with a data set of 2m
wave height. This produced satisfactory performance
in the surge direction, but the results for sway and
yaw left room for improvement.

Hence, at the present time, when teaching data is

used from two or more experimental operating
points, attempts to predict motion not included in
these data sets have been unsuccessful. The selection
of suitable teaching data is thus considered a subject
for further inquiry.

6. DISCUSSIONS AND CONCLUSIONS

As the work stands at present, it would appear to be

possible to use neural networks to approximate the
motion dynamics of a shuttle tanker when coupled to
an FPSO via a hawser. The main problem
encountered in section 5 is that the network
experiences difficulty in learning to approximate
more than one data set simultaneously. This may be
due to the nine supplied data sets representing rather
disparate operating conditions, giving the network
too much to learn accurately. There is of course the
tacit assumption that the ship dynamics do not
change greatly over time and with environmental
differences. One idea to make progress with this
problem is to use several networks, one for each
likely operating point, then to train each network
extensively using data from each point. The ship's
motions could then be obtained from one network
corresponding to the present operating conditions, or
from a weighted sum of the various network outputs.

Further examination of the processes used in

producing the tank test data is required, as well as
examination of the use of integrators leading to
offsets, as in section 4. The use of feedback control
should mean that offsets are unimportant provided
that the remaining ship dynamics are an accurate
representation of the real ship.
 MODELLING AND CONTROL OF TOWED MARINE SEISMIC STREAMER CABLES

Egil Pedersen a, • and Asgeir J. ~rensen a

a Department of Marine Hydrodynamics,

Norwegian University of Science and Technology,
O. Nielsens vei 10, N-7491 Trondheim, Norway
• Kobe University of Mercantile Marine,
2-1-1, Fukae-minami, Higashi-nada, Kobe, 658-0022, Japan

Full-scale measurement data have revealed that slowly varying fluctuations in the
course of survey vessels performing marine seismic data acquisition with towed
recording (streamer) cables are capable of forcing the in-sea hardware in longitudinal
and transverse directions. This may give rise to undesired low-frequency oscillations in
the depth of streamer cables. The interaction between the transverse and longitudinal
motions takes place at the location of the depth-control devices that are mounted along
the streamers. A quasi-static model that describes the low-frequency dynamics of
streamer cables is proposed for the purpose of controller design and analysis. The
simulation study showed that the overall depth-control performance is sensitive to
appropriate tuning of controller gains, selection of sampling period and application of
dead bands. A successful depth control will rely on the number and lift capability of the
. mounted depth-control devices. Copyright ©2001 IFAC

Keywords: Seismic operations; model-based control; hydrodynamic:s;cables

1. INTRODUCTION received by the pressure-sensitive hydrophones in the

The main objective of marine geophysical
exploration is to detect offshore hydrocarbon
resources by determining the geological structures
below the seabed. The majority of marine seismic
surveying operations are performed as 3D data
acquisition by employing a multi-cable towing
configuration and accomplished by consecutively
shooting closely spaced parallel patterns of straight
lines over the area to be explored. The more cables
the lower the operating costs on a unit survey area
basis.
A towing configuration for 3D seismic data
acquisition purposes is principally composed of a
surface vessel that tows acoustic sources and a
multiple of long, neutrally buoyant, active depth-
controlled recording (streamer) cables in parallel a
few m below the sea surface, see Fig. 1. At specified
time intervals, the acoustic source emits a pulse into
the water that propagates away from the source. The
energies, which return from the earth strata, are
streamer cables. Post-processed signals provide
images of the stratigraphy and are interpreted by
geophysicists to determine whether additional, more
detailed surveys, or drilling are merited at the survey
prospect.
The streamer cables are mainly 3 - 6 kIn long and 60
- 70 mm diameter. The cable depth is actively
controlled by depth-controllers (often referred to as
'birds'), which are spaced along each cable, typically
300 m apart. Each depth-control device is equipped
with lifting surfaces (hydrofoils) and operates
independently to supply force to control the cable
depth. The lift capacity of the most common type is
150 - 120 N at the seismic survey speed range of 4 -
5 knots. The head end of individual streamer cables
is typically attached to a tow cable, and the cables are
diverted from the towing point of the vessel by lifting
devices. Floats are attached to the front and tail ends.
The separation between adjacent streamers is within
50 - 200 m on most seismic surveys.

Corresponding author on leave to Kobe University of Mercantile Marine, Japan

Tel.! Fax: +81 784352127, e-mail: Egil.Pedersen@marin.ntnu.no

89
 2. FULL-SCALE MEASUREMENTS

Pedersen (1996) presented a seismic multi-streamer

configuration that experienced serious depth-
oscillations in the higher, outer port streamer cable
with a frequency of motion corresponding to that of
the course deviations. The individual cable length
was 3000 m and 13 depth-controllers assigned a
target depth of 8 m.

The time series plots of the heading of the survey

vessel and the depth of mid-part of the affected cable
are given in Fig. 2. The largest deviation from the
target depth reached about 6 m. Figure 3 shows the
power spectra of the vessel course and cable depth
oscillations. The energies are concentrated within
periods of 900 and 1000 seconds.
Fig. 1. An illustration of a marine seismic towing
configuration utilising multiple recording
(streamer) cables for 3D data acquisition.

In offshore seismographic services, it is vital that the

towed streamer cables maintain the depth during data
acquisition, since deviations from the target depth
specification may give rise to uncertainties and errors
in the geophysical measurements that affect the
reliability and accuracy of the final post-processed
images of the geological structures. Experience
shows that sufficient depth-keeping performance
cannot be achieved under all surveying conditions.
Depending on how serious the variations in depth are
(Le. how many depth-controllers are outside their
specified depth range and for how long), the survey
line or part of it may have to be re-shot to meet the
depth requirements.

Pedersen (1996) showed that serious depth-keeping

problems of streamer cables could be associated with
slowly varying motions in the course of the survey
vessel, even in calm waters. It was argued that the
interaction between transverse (depth) and
longitudinal cable motions take place at the location
of the depth-controllers. In worst case, the amplitudes
of motion in depth may be as great as the towing
depth. Individual oscillations have been seen to last
several minutes and may be present from the start to
the end of a survey line. When this happens, a
significant loss in productivity may be experienced.

It is of crucial importance, from the viewpoint of

operational efficiency, to find practical solutions that
can negate slowly varying cable oscillations. This
paper presents a quasi-static model that describes the
low-frequency dynamics of streamers. The effect of
varying controller settings of the depth-controllers is
analysed. The simulation results show that the overall
depth-control performance is sensitive to appropriate
tuning of controller gains, selection of sampling
period and application of dead bands. Control
strategies and gain settings that may negate the
particular oscillations in depth of seismic cables are
proposed.
Pedersen (2001) carried out a comprehensive
measurement programme on board a 6-cable
configuration in calm sea conditions. Each cable
length was 3600 m with 16 depth-controllers. Figures
4 and 5 show the time series of the dynamic part of
vessel heading, and longitudinal motion, tension and
depth of the starboard higher outer cable. The
averaged tow (water) speed was 4.0 knots.
The mechanism of interaction was explained by
Pedersen (1996) and (2001) as follows: When a
seismic vessel deviates from a straight course, the
seismic cables are affected in the longitudinal
direction. This is because a course alteration to one
side will slacken the towing cables on that side, thus
decreasing the effective velocity through the water,
whereas the cables on the opposite side will become
more taut, thus increasing the effective towing
velocity.
If the streamer deviates from exact neutrality, which
it normally does, the depth-controllers provide a
static lift force to keep it at the assigned target depth.
This causes a local curvature at the position of each
depth-controller that runs with a non-zero static wing
angle. The variation in tension interacts with a
vertical force component at the location of depth-
controllers if, and only, if, the cable is not exactly
neutrally buoyant, which is capable of inducing cable
motions in the transverse (depth) direction. The
fundamental driving source for the particular
oscillations in depth of towed streamers is therefore
the axial velocity fluctuations of the cable. Figure 6
iIIustTat~s the mp.chanism of int~Taction.

The amplitude of longitudinal motion will depend

upon the streamer's diverted position relative to the
centre line of the vessel. Thus, the outer cables face
greater variations in the dynamic amplitudes of
longitudinal velocity than the inner ones. The tension
of the streamer will change in accordance with the
variations in the axial hydrodynamic force.
 6. DISCUSSION
The success of marine seismic data acquisition
utilising towed streamer cables is completely
dependent upon the ability of the streamers to keep
an assigned target depth during all surveying
conditions. However, problems with maintaining the
depth of streamers do occasionally arise. In such
cases, the acoustic reflections received may be
subject to errors and uncertainties that affect the
reliability and accuracy of the final post-processed
images of the geological structures. Thus, slowly
varying depth-oscillations may have a significant
negative impact on the operational efficiency of this
kind of nautical operation.
The incidents of undesired fluctuations in seismic
cable depths referred to (Pedersen, 1996 and 2001)
provides satisfactory documentation of an interaction
effect between slight disturbances in the horizontal
plane motion of the seismic towing vessel and
disturbances in the depth of streamers. The time
series plots of the heading and depth-controller
depths show a corresponding behaviour and the
power spectra show that the oscillation energies
roughly coincide.
Due to the very slowly variations in longitudinal and
transverse velocities, and therefore accelerations in
the cables, a quasi-static approach to the modelling
problem was considered. The model of the streamer
includes the driving source for the oscillation (Le.
. axial speed variation), buoyancy force and
hydrodynamic forces. Figure 10 shows that the
proposed model is able to provide a result where the
largest transverse (depth) motions occur at the mid
part of the cable, which is in accordance with
expeJ:lence from full-scale observations. Interaction
between transverse (depth) and longitudinal motions
takes place if, and only if, the net buoyancy force
deviates from neutrality.
The depth-control systems in practical use apply dead
bands on fin angle and depth error, typically -0.5°
and -0.5 m, respectively. Besides, the fin angle
update rate is normally about 20 sec. Simulation
scenarios with inclusion of dead bands proved
immediately reduced depth-control performance,
even for small values. The sampling rate is another
key parameter that needs to be kept low in order to
negate the oscillations in depths. Furthermore, the
simulation study showed that the performance is
sensitive to proper control gain settings.
The seismic industry should pay attention to the gain
settings, and applications of dead bands and sampling
rate of the depth-controller if slowly varying motions
in depth of towed streamers are being observed. In
addition a proper ship autopilot with appropriate
tuning of the corresponding controller gains will
reduce the ship-induced excitations of the streamer
cables.
94
7. CONCLUSIONS
The principal result presented in this paper is a quasi-
static model that describes low-frequency dynamics
of towed marine seismic streamer cables. The effect
of varying controller settings of the depth-controllers
that are spaced along a seismic cable is analysed.
The simulation results indicated that the proposed
quasi-static cable model is adequate for simulating
slowly varying oscillations of streamer cables. It is
shown that the depth-controllers may be able to
compensate the undesired cable motions if sufficient
lift (control) forces are provided in combination with
minimum dead bands and low sampling rate.
ACKNOWLEDGEMENTS
Kobe University of Mercantile Marine is appreciated
for being the first author's host institute during the
writing of this article.
REFERENCES
Fossen, T. I. (1994). Guidance and Control of Ocean
. Vehicles. John Wiley and Sons Ltd., UK.
Larsen, M. and BruAs, A. (2000). Modelling and
Control of Seismic Vessel during Towed Multi-
streamer Operation. MSc thesis. Faculty of
Marine Technology, Norwegian University of
Science and Technology, Trondheim, Norway.
Pedersen, E. (1996). A Nautical Study of Towed
Marine Seismic Streamer Cable Configurations.
Doctoral Thesis (Nautical Engineering). Faculty
of Marine Technology, Norwegian University of
Science and Technology, Trondheim, Norway,
MTA-report 1996:115, ISSN 0802-3271.
Pedersen, E. (2001). On the Effect of Slowly-
Varying Course Fluctuations of Seismic Vessels
during Towed Multi-Streamer Operations.
Journal of Japan Institute of Navigation, 104,
pp. 95 - 101, March, ISSN 0388-7405.
Triantafyllou, M. S. (1990). Cable Mechanics with
Marine Applications. Lecture notes, Department
of Ocean Engineering, MIT.
White, F. M. (1972). An Analysis of Axisymmetric
Turbulent Flow past a Long Cylinder. J. Basic
Eng., ASME, 94, pp. 200 - 206.
 NON-LINEAR SEMI-SUBMERSIBLE POSITIONING
SYSTEM DESIGN USING AN Hoo CONTROLLER

Decio Crisol Donha *.1 Eduardo Aoun Tannuri *.2

* Mechanical Department of University of Sao Paulo, Brazil

e-mail: decdonha@usp.br

Abstract: This paper describes the development of a Dynamic Positioning System for a
large production semisubmersibleplatform. The main purpose of the work is to present
the development of a full non-linear mathematical model for motion simulations and
the synthesis of a controller for a Dynamic Positioning System using the Hoo approach.
The non-linear model used for simulations and the linear model derived for control
synthesis include representations of all the important environmental disturbance
sources and reference dynamics. Robustness and performance are treated specifically.
Results are presented and analysed. Copyright ©200J IFAC

Keywords: non-linear systems; dynamic positioning; Hoo controller

1. INTRODUCTION dredging, etc. The diversity of these tasks asso-

A Dynamic Positioning (DP) system is an active
system required in marine vehicles to maintain a
desired heading and position or to track a desired
path, despite the environmental disturbances in-
duced by waves, winds and currents. The DP
system is basically composed of three subsystems:
a) the sensor system composed of a set of me-
chanical, electrical and acoustic devices designed
to measure the vesselposition and heading; b) the
actuator system, a thruster set strategically in-
stalled on the hull to produce the necessary thrust
to counteract the environmental action and c) a
logical unit, which uses the information provided
by the sensors to command the actuator system...
. . ··
Th e mam t as k 0f th e 1oglca1 urnt ISt 0 work as th e
syst em cont ro11er.
The DP of marine platforms and ships have been
used over the last three decades in a large number
of activities including: coring, prospection, min-
ing, drilling and production of hydrocarbons, fire-
fighting, cable and pipe-laying, diving support,
1 Partially supported by FAPESP
2 Supported by FAPESP
ciated with a similar diversity of vehicle config-
urations, ranging from conventional ship shape
vessels, tugs, barges, multi-hull semi-submersible
platforms to unclassified vehicles and the contin-
uous expansions of exploration frontiers to deeper
and less accessible waters make DP a challeng-
ing problem. At the present time DP problems
have well established solutions, and in industrial
applications the state-of-the-art controller is de-
signed using the LQG control theory. However,
DP still requires appropriate attention to opera-
tional costs, reliability and safety of the vessel as
well as to the performance and robustness of the
control system.
DP controllers stlll suffers a varIety of problems.
.
For example, poor controller performances ISob-
. . ·
troned when t h e DP contro1 system ISsubmitte d
to changes in the wave frequencies and wave en-
counter angles. Changes in the operational condi-
tions such as loading led to increased costs due to
controller commissioning.Non-linearities involved
in the process such as the saturation of thrusters,
for example, were also a major problem in the
performance deterioration. The main problem is
that the controller actions are normally based on

95
very simple linear time invariant models, which
only give good performance near nominal con-
ditions. Several adaptive schemes were tried but
never implemented probably due to the intrin-
sic complexity (Katebi et al., 1985), (Donha and
Brinati,1994) .
This paper presents the design of an fico con-
troller for DP applications where the objective
is to improve robustness of stability and perfor-
mance. To achieve these goals, a control system
which couples the three degree of freedom of the
horizontal plane (surge, sway and yaw) was devel-
oped. To assist in the performance and stability
robustness assessment of the system, simulations
are carried out using a non-linear model.
The work is organised as follows. In section 2 and
3, respectively, brief descriptions of the mathe-
matical models for simulation purposes and for
control synthesis are discussed. Section 4 sum-
marises the control strategy. Results are presented
and analysed in section 5 and conclusions are
drawn in section 6.
where M is the platform mass; Xu, Yv and Nt are
the added mass along the surge (x) and sway (y)
directions and the added yaw (around z) moment
of inertia, respectively; u, v and r are the surge,
sway and yaw speed in the body coordinates; Xn,
Yh and N h are the speed dependent surge and
sway hydrodynamic forces and yaw hydrodynamic
moment, respectively; Xw, Yw and Nw are the
wind induced surge and sway forces and yaw
moment, respectively; Xc, Yc and Nc are the
control forces and yaw moments produced by the
thrusters and lzz is the moment of inertia about
the vertical axis z; X, Y and IJ1 are, respectively,
the surge, sway and yaw motion in the inertial
system and 'IjJ is the yaw motion in the moving
reference system.
The hydrodynamic derivatives for the platform
RS-35 were estimated by tests in a conventional
towing tank (Donha and Brinati, 1990).
2.1 Hydrodynamic Forces
Assuming that this semi-submersible platform is
a hydrodynamic transparent structure, the hydro-
dynamic load is calculated through a modified
Morison equation, which includes a representatiQn
for excitation and reaction forces (Sarpkaya and
Isaacson, 1981).
Wave action is modelled as the sum of first and
second-order forces, the last one including slow
drift forces, which are crucial in dynamic posi-
tioning analysis. To calculate the hydrodynamic
load, the platform is splited up into small cylinder
elements, whose characteristic dimensions do not
exceed 20 percent of the incident wave lenght. The
first-order force in direction i of a cylinder element
moving in direction j is given by:
Table 1. Main Characteristics- RS-35
drift forces, reference signals were shaped by first
order transfer functions similar to that displayed
in equation (19).

4. CONTROL DESIGN

The 1-loo approach is particularly suitable for DP

systems where the models are generally poorly
defined (Grimble, 1994). A linear time invariant
control model is usually employed, although the
vessel is subjected to extremelly different condi-
tions, such as vessel loadings, changing environ-
mental conditions, etc. In consequence, stability
and performance robustness are very important
issues. With the 1-loo approach, the error between
the full realistic model and the low order control
model can be used already during the synthesis
stage. This procedure may also take into account
the nature of disturbances entering the system.
Moreover, there is the practical problem of low
modulation of thrusters and good disturbance re-
jection. A very important point to consider when
designing a DP controller is to garantee that the
control system does not respond to motions in-
duced by the high frequency disturbances.
Figure 2 presents the two port configuration used
in the control system synthesis. The weighting
functions Wi, Wd are used to reflect the available
knowledge about the input and output distur-
bances. Wd, for example, is used to reflect the
knpwledge of the prevailing frequencies of waves
that are acting on the platform. Wn weights the
noise input. The weighting We may be used to
reflect restrictions on the control signals, while Wt
may be used to shape the complementary transfer
function, to modify, for example, tracking features
of the system. The reference dynamic transfer
function R( s) weights the reference signal. The
weighting function Ws may be used to reflect
requirements on the shape of the 1-loo controller.
Weighting functions are also used for scaling pur-
poses.
of a DP controller. The controller proved to have
very good performance under a very wide range
of simulated environmental disturbances. This re-
duces the need for adaptation under changing
weather conditions or when the platform must
work under different operational conditions. The
control system still needs some improvement on
its filtering capabilities, which is very important
in DP applications. A second stage of this project
will include a careful tuning process, mainly for
the heading action. The excellent controller per-
formance may in part be attributed to the control
design procedure which uses information on the
environmental disturbance and sensors noise. The
sensitivity functions have a direct influence on the
parameter sensitivity robustness of the design and
certain guaranteed robustness properties can be
obtained, for given types of disturbances.
The approach presented here seems to provide a
step-forward in DP applications and can lead to
more reliable systems.
 METHODS OF OPTIMAL SHIP ROUTING FOR WEATHER
PERTURBATIONS

Krzysztof Stawicki, Roman Smierzchalski

Gdynia Maritime Academy, Morska St. 83, 81-225 Gdynia, Poland

Abstract: The article discusses sea route planning and optimisation methods which take
into account minimisation of the time needed to pass a so called time-minimum route at
dynamically changing weather conditions. The above task is a kind of dynamic
optimisation problem in which dynamic changes in weather conditions are defmed as
moving constraints. The effect of hydro-meteorological conditions on changes in the
parameters of ship motion and route was determined. The obtained solutions and
authors' experience gained during studying the presented methods have made the basis
for developing new evolutionary algorithms which were not used so far in planning
ship routes. Copyright ©200J IFAC

Keywords: ship routing, trajectory optimisation, and evolutionary computation

hydro-meteorological forecasts - with possible

1 INTRODUCTION
The aim of navigation, meant as the dynamic
process of determining actual positions of objects in
motion, is, most of all, to move the ship in a safe
and efficient way between two fixed points on the
surface of Earth. Of certain significance is also an
economical aspect which consists in searching for
the most favourable route taking into account such
elements as: technical performance of the ship, its
navigational equipment, as well as hydro- and
meteorological conditions specific for the route of
interest. One cannot neglect here a human factor,
which can manifest itself, for instance, in the need
to enforce temporarily sea watches.
Ocean sailing burden navigating officers with
special obligations which include: making use of
maps with in advance prepared safe routes (tactic,
strategic, and optimum routes), steering the ship
along these routes, monitoring the route with
respect to climatic data and great circle, studying
changes made in advance planned routes,
maintaining proper frequency of position checks,
and s.ending possible ~eports, if the navigation is
supplied by weather gUIdancecentres.
2 PLANNING OCEAN PASSAGES
The most widely applied method of planning the
ocean passages is planning the route by the captain
and officers on the basis of their own studies of
meteorological data (Wrobel F.: (1992». The
choice of route depends on the climatic zone of the
sailing area of interest, as well as on the technique
itself and the method of its execution. The
following types of routes can be distinguished with
respect to planning process: climatic routes,
strategic routes - planned on the basis of weather
forecasts, tactic routes - planned on the basis of
current weather conditions, optimum routes for
certain criteria assumed.

101
simplifications, one can defme the tenn, existing in
the literature, of a "typical" trajectory of the
tropical cyclone. This trajectory consists of three
sections:
• Equatorial section - along which the cyclone
moves west, with small north component
(south on the Southern Hemisphere). The
speed of the translator motion of the cyclone
along this section is, as a rule, low, within 5 to
12 knots (8-9 knots, as average). The cyclone
very often happens to alternately accelerate
and decelerate its speed, to move north only,
or to increase the north component.
• Meridional section - begins when the north
(south) component starts dominating over the
west component, and ends when the east
component starts dominating over the north
(south) component. This section includes a so
called point of recurvature. The cyclone speed
along this section increases to 12-20 knots.
• Polar section - the main direction of the
cyclone motion is north-east (south-east on the
Southern Hemisphere), the cyclone speeds up
to 20-30, or even 50 knots.
the area occupied by the tropical cyclone which
affect the ship, and may, for instance, result in
damage or incorrect operation of navigational
instruments, thus affecting the safety of sailing.
3.3 Modelling of the area of potential danger
Warnings about tropical cyclones most often
include the infonnation on predicted direction and
speed of the cyclone motion, frequently
complemented by predicted positions of the centre
of the cyclone system. It is difficult, however, to
evaluate both the exact position of the cyclone eye
at the time of analysis and the evaluation error
itself. Therefore, to make the manoeuvre of
avoiding cyclone effective, the area of potential
danger should be evaluated, i.e. the area which can
be reached by the eye of the cyclone moving in the
prescribed manner in the time of interest.

3.2 Cyclone as a navigational danger

Pieces of infonnation which are required by the

navigator the most are the radii of circles, inside
which winds have certain strength, including a so
called hurricane wind radius (strength ~12°B), and
a stonn wind radius (strength 8-11~). Outside
these circles there is a zone of winds lighter than
stonn winds (7°B and less) (Mars A., Styszynska A.
(1992)).
Observations show that the stonn wind radius may
vary from 25 to 270 nautical miles, while the
hurricane wind radius makes between 0.33 and 0.5
of the stonn wind radius. A general rule is to avoid
the stonn wind area, the more the hurricane wind
area. Although the entire area of the tropical
cyclone represents certain level of danger for
sailing, it can be divided into two halves along the
vector representing the translator motion of the
entire cyclone system. On the Northern Hemisphere
the right half (left - on the Southern Hemisphere) is
characterised by stronger waves and winds, which
makes sailing conditions worse than in the other
half. Therefore the right half of the tropical
cyclone on the Northern Hemisphere bears the
name of the dangerous half while the left one - the
navigable half (the opposite situation is on the
Southern Hemisphere). It is noteworthy that the
tenn "navigable" does not mean safe, but the
situation of the ship sailing across this half is
relatively more comfortable than when it sails
across the opposite, dangerous half. One should
keep in mind that along with wind and waves, there
are also other unfavourable weather phenomena in
 target point B is chosen as the origin and the above
construction is repeated to give a 24-hour
isochrone, and so on, until the target point B is
reached (see Fig. 3). At present, studies on the
application of dynamic programming to evaluating
optimum weather routes are in progress. The
authors used the abbreviation DPWR (Dynamic
Programming in Weather Routing) for this activity.
Using algorithms based on dynamic programming,
they made a series of simulation tests for directing
the ship between departure and arrival harbours, in
which the ship met areas, characterised by
unfavourable weather conditions. These studies
aimed at determining optimum trajectories with
respect both to the passage time and fuel
consumption.
F.Aligne et al. (1998) formulates and proves a
thesis that this goal could be only achieved because
the steering vector included not only the ship
course but also changing main engine settings. This
made it possible to steer both the ship course and
speed. It was proved that apart from situations
when the ship slows down to leave aside bad
weather areas, quite opposite situations may happen
in which the ship speeds up to enter those regions
to gain profits from the wind blowing from astern
sectors.

6 CONCLUSIONS

The presented discussion of the methods of

modelling the navigational environment with
weather perturbations and planning passage routes
has made it possible to formulate the following
conclusions and comments:
• For open areas, the acceptable approximation
of the optimum route planning solution is an
assumed steering which minimises the passage
time, or the diversion of ship's own trajectory
with respect to the assumed one.
• A mathematical model formulated in kinetic
terms is sufficiently adequate for route
planning process.
• It is possible to model weather perturbations,
equivalent to moving constraints of the process
state, as moving areas having shapes and
dimensions depend on assumed safe distances
at which these areas are to be passed.
• Practical solution of the ship trajectory
optimisation task which makes use of
maximum and optimum principles can be
obtained using in limited range.
Due to necessary adaptations of the ship's route to
rapidly changing weather perturbations it is
advisable to solve the optimisation problem using
evolutionary techniques, which provide
opportunities to study the problem in a wider range,
including dynamic changes of weather
perturbations modelled as moving areas. The route

105
planning problem can be reduced then to the multi- Banachowicz A. (1995) Environment and sea area
criterion dynamic optimisation task with static and parameters having an influence on
dynamic constraints. restriction of ship's moving. Gdynia
Further works on the subject will present an Maritime Academy, Navigational Faculty.
evolutionary method used for solving the problem The work under Andrzej Banachowicz
of ship route planning in the presence of changing leadership. Gdynia.
weather perturbations. Taking into account Jurdzinski M. (1989) Sea passage planning. Marine
continuously increased speed of computers it seems Publisher, Gdansk.
possible to use an interactive and multi-criterion Jurdzinski M. (1994) Influence of technical,
approach in the future to support navigator's constructional and operational parameters
decision, where the operator would choose a on ship's moving in optimalization
solution from a number of multi-criterion effective aspects. Gdynia Maritime Academy,
solutions prepared by the decision support system. Navigational Faculty. The work under
Miroslaw Jurdziilski leadership. Gdynia.
ACKNOWLEDGEMENTS M~droszkiewicz K. (1974) Sea passage
optimisation taking into account
The research reported in this paper was supported economical criterion. Publications of
by the Polish State Committee for Scientific Gdansk Polytechnic no. 220, Electrician
Research. XXXIV. Gdansk.
Marsz A., Styszynska A.: (1992) Exercises in
REFERENCES meteorology and oceanography. Part II,
Tropical cyclones. Gdynia Maritime
Aligne F., Papageorgion M., Walter E.: (1998) .. Academy, Gdynia ..
Incorporation power variations into NOWICkI A.: (19:8) Knowledge of maneuvermg
weather routing and why it may lead to ocean-gomg vessels. Problems of theory
better results. IFAC CAMS Fukuoka and practice. Marine Publisher, Gdansk.
Japan. ' Wisniewski .B.: (1986) Ship's route optimali:~tion
Chotkowski W., Madroszkiewicz K. (1974) takl~ account .weather. condltl~ns.
Analysis of methods used to control of .. M~tJtJme Academy m Szczecm, Szczecm ..
ship's route with minimum sailing time. Wismewski B., Holec M.: (1983) OceanographyIn
Publications of Gdansk Polytechnic no. bro~d outline. Part /J, Dy~amic of sea.
220, Electrician XXXIV. Gdansk. PolIsh Navy Academy, GdynIa.
Wrobel F.: (1992) Vademecum of navigator,
Marine Publisher, Gdansk.
 ASSISTED DYNAMIC POSITIONING SYSTEM FOR A FPSO BASED ON
MINIMIZATION OF A COST FUNCTION

Eduardo A. Tannuri, Celso P. Pesce, Decio C. Donha

Mechanical Engineering Department of University of saa Paulo

Abstract: This paper presents a two-layer controller for a turret-moored FPSO. The first layer
evaluates ship optimal heading, based on the minimization of a cost function relating important
operational parameters, namely first-order roll oscillation, dynamic traction on risers, mooring
lines offset and fuel consumption. Since this layer requires an estimate of the environmental
conditions, it is proposed an indirect evaluation of the wave spectrum by the measurement of ship
first-order oscillations. The second level receives the set point from the superior level and
commands the thruster system. This layer is composed by a robust heading sliding-mode
controller and the surge and sway motions are damped by a simple feedback linearization scheme.
The controller was tested through numerical simulations of a moored VLCC under common
environmental conditions found in Campos Basin, Brazil. Copyright © 2001 IFAC

Keywords: Control, Sliding Modes, FPSO, Turret, Assisted dynamic positioning.

position that surge and sway mean forces are

1. INTRODUCTION
Floating Production Storage and Offioading System
(FPSO) is a modern concept for floating offshore oil
exploration units. A tanker is moored in deep water
and a production plant is installed on its main deck.
Because of tanker large water line area, they are
exposed to severe environmental loads, which may
induce large displacements, possibly causing rupture
of the mooring lines and risers. Mooring systems are
designed to minimize such loads, allowing the ship to
be aligned with the resultant of the environmental
forces. The turret mooring system, for example, is
composed by a cylindrical structure (turret) supported
by an axial bearing system fixed to the ship and
moored to the seabed.
Under certain combinations of environmental agents,
waves may reach the ship almost perpendicularly,
inducing large first-order vertical motions. These
motions cause dangerous cyclic loads in risers and
mooring lines, difficulties in the production process
and discomfort for the crew (Leite et aI., 1999). Even
when the FPSO is equipped with an assisted dynamic
positioning system, it cannot directly compensate
first-order motions, due to high frequency
components that would require an enormous power to
be attenuated. To avoid these motions, it is proposed
to head the FPSO in a way to reduce first-order
motions. Furthermore, to minimize fuel consumption,
the controller must maintain the FPSO in such a
counteracted by the mooring system, and not by the
propellers forces.
The proposed FPSO control system is composed by
two layers (Tannuri and Donha, 2000). The first layer
calculates the optimal set-point position, minimizing
a cost function relating roll oscillation amplitude,
dynamic traction in risers, static offset of the system
and fuel consumption. To exemplify the importance
of these operational parameters, it may be considered
a relatively common non-extreme condition in
Campos Basin, composed by swell waves, which
come mainly from the south (or southeast), in the
presence of current coming from the north (or
northeast) direction. In such case, the angle between
waves and current is typically between 900 and 1800
and, depending on their relative intensities, the FPSO
may be subject to beam-sea waves with
approximately 4.0m significant height. This situation
may be unacceptable, causing high roll motions and
dynamic compression in the risers (pinto et aI.,
1999), being necessary sometimes to shutdown the
production plant. Of course, to minimize fIrst-order
motions it is required to align the ship with waves.
However, this is not the only parameter that must be
taken into account. Due to current and wind actions,
when the ship is aligned to waves, it may be observed
large displacements in mooring system, possibly
causing rupture of lines. Also, propellers forces
required to keep the ship in that situation may be
unfeasible or may require economically unviable fuel

107
 POSITIONAL GAME CONTROL OF SHIP IN COLLISION SITUATIONS

Jozer Lisowski

Gdynia Maritime Academy

83 Morska Str., 81-225 Gdynia, Poland
e-mail: jlis@wsm.gdynia.pl

Abstract: The paper concerns the application of elements of the game theory for the purpose
of optimal control of some dynamic continuous processes. Using as an example the process
of safe ship's control, the paper presents the problem of applying a positional non-
cooperative game of "j" objects for the description of the process considered as well as for
the synthesis of optimal strategies. A mathematical model of differential game is
formulated and its approximation in the form of linear programming problem is used for the
synthesis of safe ship's trajectory as a multistage process decision. The considerations have
been illustrated an example of a computer simulation the POSGAME programme to
determine the safe ship's trajectory in situation of passing a big number of the objects
encountered, recorded on the ship's radar screen in real navigational situation at sea.
Copyright ©2001 IFAC

Keywords: differential games, positional games, dual linear programming, safety analysis,
risk, ship control, computer simulation.

I. INTRODUCTION algorithms which are afterwards effected by the

The paper describes process of the safe ship control
in the collision situation using the dynamic game
model with ,j" participants. Depending on the degree
in which the ship's dynamics is taken into account
the following process approximate models have been
distinguished - basic model as the simulation one
and remaining models for the purpose of synthesis of
the algorithms describing the game ship control. The
process basic model comprises non-linear state
equations and non-linear time-depending constraints
of the state variables as well as the form of game
integral and final payment.
In practice the process of the handling of a ship as
multidimensional dynamic object depends both on
the accuracy of details concerning the current
navigational situation obtained from the ARPA
(Automatic Radar Plotting Aids) anti-collision
system and the form of the model of the process used
for the determination of the rule of handling
synthesis. While formulating the model of the
process it is essential to take into consideration both
the kinematics and the dynamics of the ship's
movement, the disturbances, the strategy of the
encountered objects and the formula assumed for the
goal of the ship's handling.
The diversity of selection of possible models directly
affects the synthesis of the ship's handling
ship's handling device directly linked to the ARPA
system and, consequently, determines the effects of
the safe and optimal control. In order to ensure the
safety of navigation ships are obliged to comply with
Rules of the Convention on the International
Regulations for Preventing Collisions at Sea
(COLREG). However, these Rules refer to two ships
under the conditions of good visibility. In case of a
limited visibility the Rules only specifY
recommendations of a general nature and are not able
to consider all necessary conditions which determine
the passing course. Consequently, the actual process
of ship's passing very often occurs under the
conditions of uncertainty and conflict together with
unprecise co-operation of the ships with reference to
the COLREG Rules. It is, therefore, reasonable to
investigate, develop and represent the methods of
ship's safe handling using the rules of a theory of
dynamic games and a theory of fuzzy sets for the
purpose of ship's operation (Lisowski, 1986; Merz
and Karmarkar, 1976; Vincent, 1977).
2. BASIC MODEL
The most general description of the own ship's
passing the ,j" number of other encountered ships is
the model of a differential game of a ,j" number of
objects (Fig. 1).

113
 RESEARCH ON
FLEXIBLE OSCILLATING FIN PROPULSION SYSTEM
AND ROBOTIC FISH

Ikuo Yamamoto and Yuuzi Terada

* MitsubishiHeavy Industries, Ltd., Japan

Corresponding address
Mitsubishi Heavy Industries, Ltd. Technical Headquarters,
Nagasaki R&D Center, Control System Lab.,
5-717-1 Fukahori, Nagasaki 851-0392, Japan
e-mail: yamamoto@csl.ngsrdc.mhi.cojp

Abstract: The purpose of this paper is to describe the feasibility research of marine vehicles
with oscillating fin propulsion control system. The oscillating fm propulsion system was
designed and constructed to be combined with a ship model. Tank tests using the ship
model have confmned the system's feasibility. As a result, several advantages of the
oscillating fin system have been found out. A neural network was successfully applied for
an identification of the ship model dynamics with the oscillating fin, and its effectiveness
was confmned. Robotic fish, intended as an amusement attraction for aquariums, using the
oscillating fm propulsion system have been developed. Its capability of untethered 3-
dimensional movement was confmned. Copyright ©20011FAC

Keywords: Flexible oscillating fin, Marine system, Propulsion device, Robotic Fish

1. INTRODUCTION
It is well known that marine creatures such as fish
swim using small power even at high speeds
(dolphin: 60 kmIh, swordfish: 80 kmIh), see Yoshida
(1976), and sweetfish, etc. are superior in their
position keeping characteristics. These characteristics
as creatures have been of interest as science from the
old times and much research has been conducted, see
Hertel, H. (1966), however; it is rare to study these
characteristics from the viewpoint of a engineering
subject, see Isiki et al. (1983).
....
One such research IS on a flexIble osclllatmg fm
. etc. has been deveI0ped as an applied product.
control system whIch could be used for the
propulsion of marine vehicles by positively making
the. most of the ~haracteristi~s of the flexibl~ p~
ThIs method obtams a propulsIon force by osclllatmg
fins equipped to vehicles on the analogy of the
..
moUon of manne creatures.
The linear theory analysis has been applied to the
rigid oscillating fin which regards the oscillating fm
as a rigid part, and past experiments and research, see
Isiki et al. (1983), have reported that the flexible fin is
superior to the rigid fin. Theoretical studies, see Isiki et
al. (1983), on flexible oscillating fins have been partly
conducted, while studies on the flexible fin including
the control system have not yet been performed. In this
research after the control system for a flexible
oscillating fm propulsion device and the oscillating fm
driving device were designed and manufactured, a
cruising test was performed first by a numerical
simulation and then with a model ship, and the
fundamental performance has been grasped and
prospects of putting them to practical use have been
obtained. And robotic fish for amusement in aquariums
'
2. OUTLINE OF THE SYSTEM
. '
In many cases the kineUc parameters 0f the oscillatmg
. ·
fin cannot be directly detected m controI 0f an oscillatlng
fm and there are problems choosing and identifying
parameters to be used for control, and a control system
able to cope with such problemsshould be architected.

119
 A NONLINEAR ST A TE-SP ACE MODEL OF
DIESEL PROPULSION PLANT OPERATION USING NEURAL NETS

Nikolaos I. Xiros and Nikolaos P. Kyrtatos

Laboratory of Marine Engineering

NATIONAL TECHNICAL UNIVERSITY OF ATHENS
Iraon Polytechniou 9, 15773 Zografos, GREECE

Abstract: Cycle-mean-value, quasi-steady, thermodynamic models of the slow-speed,

two-stroke turbocharged marine Diesel engine is a valuable simulation tool. Engine
operation models of this kind comprise of two differential equations and a nonlinear,
perplexed algebraic system for the engine/turbine/compressor torques. Neural nets are
proposed in order to establish the functional dependence of the torque variables upon
the state variables (engine/turbo rpm) and the control input (fuel index). In effect a
state-space model is obtained which is decomposed to a nonlinear nominal model and a
linear uncertain perturbation model, convenient for the application of robust control
methods. Copyright © 2001 IFAC

Keywords: Neural nets, marine propulsion, state-space models

1. INTRODUCTION State-of-art engine control is comprised by PI engine

Propulsion of the vast majority of contemporary
merchant ships (especially containerships and
VLCCs) utilises the marine Diesel engine as prime
mover. Typical marine propulsion plants include a
single, long-stroke, slow-speed turbocharged, two-
stroke Diesel engine directly coupled to the vessel's
single large-diameter, fixed-pitch screw propeller
(Faber, 1993). Most often a variable-speed AC
generator is also attached to the engine-propeller
shaft for efficient electric power generation. This
typical configuration has been adopted because it is
characterised by quite large propulsion power outputs
(typically more than 30-40 MW from a single unit)
and yet by operational robustness and increased
efficiency due to its conceptual simplicity.
The Diesel engines used in such marine propulsion
powerplants are designed for steady-state operation
(Faber, 1993; Faber 1994; MAN B&W, 1996) and
they are selected, during the ship design stage, so that
they can cover the vessel's propulsion power
requirements at an rpm value that does not exceed
100-120rpm(Faber,1994).
speed (rpm) regulation, enriched with a number of
nonlinear functions, such as control action limiters,
and heuristic PI gain scheduling in order to improve
closed-loop system transient performance at the
various engine operating points and loading profiles
(Lan, et a/., 1996). It should be noted out that the fact
that all marine propulsion engines are turbocharged
imposes additional concerns to engine control
developments (Roskilly and Mesbahi; 1996). As in
many cases, however, the control problem originates
from a closely related problem of appropriate
modelling of plant operation. In this respect, the
objective of this work has been the development of a
nonlinear state-space model that can depict the
marine engine-turbocharger interaction and
operation, on one hand, and can be directly used for
the application of advanced controller synthesis and
robustness analysis, such as nonlinear and robust
control methods.
Opposite to the case of automotive engines, which in
many cases are not turbocharged, the experimental
determination of the state equations is not possible
for marine propulsion engines. This is due to a

131
Current generation unmanned underwater vehicles, equipped with robotic manipulators, are teleoperated and consequently
place a large workload burden on the human operator. A greater degree of automation could improve the efficiency and
accuracy with which underwater tasks are carried out. These tasks can involve manipulator motion that is both
unconstrained and/or constrained. For unconstrained motion, where a trajectory requires following, a prerequisite is good
joint angle control. An adaptive self-tuning pole-placement controller is used for joint angle control. Practical results show
the benefits compared to conventional fixed-gain control. For constrained motion, often simultaneous control of position
and force is required. An adaptive hybrid position/force controller is proposed and compared to a fixed-gain version.
Simulation and practical results illustrate the merits and drawbacks of each scheme. Copyright © 2001 IF AC

1. Introduction

Unmanned Underwater Vehicles (UUVs) are commonly
used in offshore, marine science, defence and salvage
activities for inspection, maintenance and repair.
Workclass vehicles are equipped with one or two
manipulators, and are controlled using tele-operation by
an operator on the surface. The manipulators are generally
hydraulic, with no tactile or contact sensing for
monitoring or controlling contact forces during task
execution. Consequently, damage to manipulator and
environment are common during typical operations. The
use of sensed force and torque data would greatly assist in
execution of many tasks [1]. Such an ability is central to
activities such as weld inspection and grinding. In
addition for unconstrained manipulator motion, possibly
involving complex reference trajectories, accurate joint
angle control is essential. In this paper, we consider the
problems of joint angle control and simultaneous control
of manipulator position and applied force using an
underwater hydraulic arm with a fixed base.
There has been an abundance of joint angle control
schemes proposed over the last two decades e.g.
fuzzy/neural, model reference, computed torque, variable
structure, optimal and self-tuning. Any successful control
scheme must be able to cope with the following problems.
The highly nonlinear dynamics of both the manipulator
and actuator, including inertia, gravitational, Coriolis and
centrifugal effects, friction, mechanical flexibility,
backlash and actuator geometry. Accurate control is
required over a wide range of operating conditions. There
is cross-coupling between neighbouring inputs and
outputs of the system and the system dynamic parameters
are time-varying. An adaptive control scheme was
deemed most suitable for this application, as it can
automatically accommodate the wide variations in
manipulator dynamics. A self-tuning approach, using
pole-placement control, was chosen.
There have also been many proposed control strategies to
achieve simultaneous position and force control. Two of
the most popular are hybrid position/force control [2] and
impedance control [3]. A hybrid position/force control
scheme decomposes the task space into two orthogonal
sub-spaces, one controlling the manipulator
position/orientation in certain Cartesian axes, the other
controlling forces/torques in the remaining directions.
This enables the various tasks encountered during subsea
intervention, such as weld inspection and mating of
connectors, to be readily accomplished. Often, underwater
tasks require specific forces to be applied to unknown
environments and this is difficult to guarantee with
impedance control schemes. We have therefore chosen a
hybrid position/force control approach.
During subsea intervention the manipulator operates in an
unstructured and time varying environment, with wide
changes in operating conditions. These variations must be
accommodated by the control scheme used. It has been
shown that a fixed gain hybrid position/force control
scheme which performs well for one particular set of
operating conditions, degrades considerably away from
this operating point [4]. In this paper we consider a self-
tuning adaptive hybrid scheme to cope with any
unmodelled disturbances and variations in the robot and
environment dynamics.
The paper is organised as follows. Section 2 describes the
self-tuning adaptive joint angle control scheme. The
following section shows the practical results of the
adaptive scheme and compares it to a fixed-gain

149

...
4.3. Imtwl Shallow Water Dock Trwls
In order to validate the SRPS system and the
Swimmer vehicle in a controlled manner, shallow
water tests were performed. The docking station was
installed on the seabed (12m below sea level) in
Cybernetix's dock facilities in Marseille. Several
tests were performed in order to validate system
operations and docking/detaching operations.
....
Imtlally ~he SWImmer vehIcle was. suspended !Tom a
heavy 11ft crane on the docksIde (see FIg. 8).
Although not driven, the Swimmer's onboard system
were monitored using "test" cables carrying Ethernet
/ TCP-IP data to monitoring a PC on the dockside. To
ensure the successful mating of the vehicle to the
docking station, the vehicle must be positioned
within the following specifications:
• Within a +/- 20cm misalignment tolerance
to the docking station.
• Less than +/- 5° pitch/roll in order to
perform a successful coupling.
159
During the tests, some buoyancy and power problems
were exposed. After some modifications, the tests
were repeated and the vehicle successfully docking
onto the submerged station.
4.4. Manual Operation
To test the operation and control of the thrusters,
positional sensors and the variable buoyancy system,
a series of manual-docking tests were performed. By
operating the vehicle via a joystick interface (without
the crane), the next phase in the test programme
utilised the vehicle sensors' including the SRPS to
pass positional information to the pilot to aid in the
remote docking of the vehicle. Once docked the
vehicle's docking latches were operated to ensure the
vehicle was adequately coupled to the docking
station (Note the wet connectors were not installed
for this test). The final phases of the manual test were
to disengage the docking latches and return to the
surface. In addition, tests were undertaken to test the
Variable Buoyancy System, and fail-safe recovery of
the vehicle. The vehicle was piloted to the seabed
and all systems were powered off, (by powering off
the vehicle a 150kg electro-magnetically coupled
clump weight was released and the vehicle now
being positively buoyant returned to the surface).
4.5. Automatic Operation
The final set of shallow water trials (to be performed
sometime in May and June 2001) will utilise all the
vehicles systems. The initial position information
will be supplied by the vehicle's onboard sensors
(together with a external GPS / Gill reference
because the LRPS cannot be used in shallow water,
[Bechaz 2000]). This will be fed into the ASap
algorithm and be used to initialise the docking
process. The system will automatically navigate from
the starting position to the final position where the
vehicle will be mated with the docking station. Once
power and communications have been established to
the Swimmer vehicle, and the Phenix ROV (via the
wet mateable connector), the ROV will be detached
from the Swimmer vehicle and piloted around the
docking station. Upon satisfactory completion of the
ROV tests, the Phenix vehicle will dock with the
Swimmer vehicle, the umbilical will be automatically
re-coiled by the Swimmer's Tether Management
System (TMS). The Swimmer vehicle will then
detach from the docking station
(power/communications connectors will be
disengaged.) and the vehicle will then return to the
surface.
4.6. Deep Sea trials
Final deep-water (>300m) trials are scheduled to take
place in Summer 2001. off the Mediterranean coast.
The work-class ROV will be mounted on the
Swi~r AUV and launched .from D~ (Dy?amic
PosItlomng) vessel. The docking statlon wIll be
installed in -300m of water, and the vehicle deployed
 Abstract: Accurate identification of nonlinear time variant MIMO systems, especially in
case of AUVs is essential for implementation of control algorithms and navigation
purposes. Control problems of AUVs have also difficulties due to the nonlinear
dynamics behaviors of vehicles and also unpredictable effects come from the
surrounded water mass. These nonlinear effects are so complicated that bring
difficulties for dynamics modeling and position control descriptions while using
conventional methods. The proposed method here uses neural networks as a general
idea for dynamics modeling and position control of any six-degree of freedom rigid
body and are applied to an AUV, named Twin Burger 2, as an example. Supervised
Learning and Unsupervised Learning are used for adjusting the neural networks'
synaptic weights and the results are illustrated. Path planning of AUVs using neural
network is also addressed here as of a complicated control scheme and Reinforcement
Learning is used for adjusting the neural network parameters of the path planning
module via some obstacle avoidance examples. Copyright ©2001IFAC

Keywords: Autonomous Underwater Vehicle (AUV), dynamics modeling, position

control, path planning, neural network, SISO, MIMO, obstacle avoidance.

1. INTRODUCTION of MIMO systems such as AUVs with the above

This paper discusses about three different issues
regarding to AUV systems, which are 1) dynamics
modeling, 2) position control, and 3) path planning,
all are investigated using neural networks.
The most popular method for deriving dynamics
equation of motion of a moveable rigid body is using
conventional Newton LaGrange mechanics. In order
to implement this method and derive equation of
motion for navigation and control purpose, in
addition to describing kinematics of the rigid body
precisely, it is necessary to identify all of the external
force and moment terms as precisely as possible.
Dynamics modeling of AUVs using Newton
LaGrange methods cannot be done easily due to the
difficulties in identification of environmental effects
caused by surrounding water mass and nonlinearities
of actuators' performances. Position control problem
mentioned nonlinearities cannot be easily
accomplished using conventional control theories
too. In this paper it is going to find proper substitutes
for conventional dynamics modeling and
conventional position control to overcome to these
difficulties using neural networks.
Recently, different kinds of neural network structures
are applied to system identification and controls
purposes of various AUV systems (Fujii and Ura,
1991). Neural network structure proposed by Fujii
consists of two parts, which are Forward Model, and
Controller Network. Forward Model network is
responsible to identify the dynamics behavior of the
vehicle and is trained off line using Supervised
Learning. Position control network is used for
keeping the vehicle in a desired position and is
trained on line using Unsupervised Learning. The
only disadvantage of the control system proposed by

167
Fujii is its consumption time while on line
adaptation. Imaginary Training algorithm proposed
by Ishii et ai. (1995) modifies the structure proposed
by Fujii by dividing it into two different parts called
Imaginary World. and Real World. Real World part
is used for on line implementation of the control task
and Imaginary World part is used for off line
adaptation of the controller. The problem regarding
to the network proposed by Ishii is that the Forward
Model and Controller Network are just applied to
SISO system and have not enough generality to be
applied to MIMO ones. In this paper it is going to
overcome to this problem by extension of his idea to
dynamics modeling. and position control problems of
six degree of freedom rigid bodies. Twin Burger 2 is
considered as an example and some results are
illustrated (Sayyaadi et aI.. 2000).

The other different feature addressed in this paper is

path planning of AUVs using neural networks.
Obstacle avoidance for an AUV system is a very
important task that has to be considered during a
mission in order to prevent damages to t?e vehicle
due to collision with the hazardous environmental
objects. Motion planning of an AUV can be
decomposed into path planning. and. trajec.tory
planning. Path planning for obstacle aVOidanceISto
generate a collision-free path in an environment
while optimizing it with respect to some criterion.
Trajectory planning is to schedule the ~ovement ?f
an AUV along the planned path. ThIS paper wIll
focus on the first issue. path planning.
. designed for control adaptation. Real World Part is
The most recent works on path.planmng for obstac~e
avoidance of AUV systems IS based on GenetIc
Algorithm (Sugihara and Yuh, 1997). General
drawbacks of Genetic Algorithm methods are their
low speed toward converged results, because the
algorithm has to search most of the possible solutions
to the problem using random probability for new
generation through crossover and mutation
mechanisms. Fujii et al. (1998) suggested a new
method for obstacle avoidance problems based on the
Reinforcement Learning method. This algorithm was
implemented for multi-robot syste~ in sim~lation
and actual environment. Although hISmethod ISvery
practical. it has one drawback and that is his
controller cannot produce a continuos control action.
because the control action space and the velocity
state space have to be partitioned into some discrete
values in advance.
The proposed module for path planning while
obstacle avoidance of AUVs in this paper is an
example of advanced control by neural network and
uses normal distribution function for continuos
output generation. Learning method is based on
Reinforcement Learning. Proposed method be just
applied to the model extracted for AUV. Twin
Burger 2. for its planar motion in a horizontal plane
considering its Surge. Sway, and Yaw motion and
having constant Heave.
Imaginary Training algorithm proposed by Ishii et al.
(1995) as an on-line adaptation mechanism for neural
network controller is depicted in Fig. 1. This
algorithm is realized by introducing the structure.
which consists of two separate processing parts; Real
World Part. and Imaginary World Part. In this
structure the neural network controller can be
adjusted independently of the actual operation of the
controlled object. Real World Part is designed for
actual control operation. and Imaginary World Part is
an ordinary feedback system that consists of a
controller (C_R) and a controlled object. Inputs to
the C R are the differences between the reference
signal~ r(t) and the state variables Se(t) given by
...
the controlled object. l~aglOary World Part consIsts
of a controller (C_I). WhICh~as ~e same s~cture as
C_R. and a n~ural network IdentIfier. The lOp~tS to
C_I are the dIfference between the reference sIgnals
r(k) and the state variables Sf (k ). come from the
neural network identifier. Inputs to the neural
network identifier are control signals u(k) given by
. fl . th I . W ld
the C_I. where the bme ow 10 e maglOary or
Part is denoted by k . which is independent of that
of the Real World Part. Ishii applied his idea to an
AUV test bed. named Twin Burger 2 as an example
for its Heading motion mode. The difficulty
regarding to Ishii's model is that his model is
applicable to control problem of SISO systems and
has not enough generality to be applied to MIMO
systems. In this research work it is going to adopt his
architecture and modify it for MIMO control
purposes.
For MIMO systems. any kind of neural network
architecture regardless of the number of layers and
the number of neurons in each layer are very

168
complex and large that cannot be trained easily in a
single step of training. As a result it has to be trained
gradually in some sequences of training procedures.
Dynamics modeling of MIMO systems is called here
as Coupled Model Neural Network Identifier
(CMNNI), and position control is called as Couple
Model Neural Network Controller (CMNNC). Each
one is decomposed to a numbers of SISO systems,
called Single Degree of Freedom Neural Network
Identifiers (SDFNNI), and Single Degree of Freedom
Neural Network Controllers (SDFNNC). After
training of all of these SISO identifiers and
controllers, additional interconnecting synapses
compose MIMO identifier and controller and then
these new synaptic weights are adjusted through
additional training processes.
Identifier and controller networks have three layers,
which are input layer, hidden layer, and output layer.
To simulate the motion of the vehicle by giving input
force and only initial condition of velocity, the output
of the first integration layer has to be fed back to the
input layer through RC-l loop. The recurrent
connection from the hidden layer to the input layer,
called RC-2, allows the network to express the
dynamics behaviors with reduced number of input
state variables. The RC-3 loop makes the controller
able to provide control signal (force/moment) by
inputting only its initial value at the beginning of the
control task. Inputs to the controller network are past
time step of control signaVsignals (force/moment),
position error/errors, and velocity error/errors.
Output of the controller network is control
signaVsignals of the next time step. Inputs to the
identifier network are control signaVsignals, and past
time step velocity/velocities.
Every neural network needs some kind of data sets or
teachers to be used as of teaching data. Using
Supervis~ Le~ing carries out training of the neural
network Identifiers (dynamics modeling). It means
that input/output behaviors of the system are sampled
through some experiments and then these sampled
data sets are used for adjusting the parameters of the
neura~ net~o~ks. The proposed identification
techmque IS Impl~mented to a four degrees of
freedom AUV, Twm Burger 2, which is shown in
Fig. 3. This AUV has four motion modes called:
Surg~ (longitudinal), Sway (lateral), Heave
(Vertical), and Yaw (rotational). Doppler Sonar
Sensor detects state variables of the vehicle. Number
of neurons in hidden layer of every SDFNNI and
S?FNNC are 6 neurons and as a result this number
will be 24 neurons for CMNNI and CMNNC.
...
~rammg algonthm of neural network identifiers in
smgle degree of freedom modes and coupled mode is
carried out in 4 consequent steps. Error Back
Propagation rule is used while training processes. In
the first step of training, inputs to the identifier
networks are control signaVsignals and
velocity/velocities come from the sampled data sets.
Evaluation error is based on the discrepancies
between simulated acceleration/accelerations and
actual one/ones. In the second step, inputs are
control signaVsignals come from data pack and
simulated velocity/velocities. Evaluation function is
the same as step one. In the third and fourth steps,
inputs are the same as step two. Evaluation functions
are made from discrepancies between actual
velocity/velocities and simulated one/ones in step
three, and actual position/positions and simulate
one/ones in step four. All of the synaptic weights of
the SDFNNIs are set to some small random values
when training starts and then these parameters are
adjusted gradually. After getting proper convergence
in the evaluation function of each SDFNNI, the
CMNNI is made using additional interconnecting
synaptic weights. These weights are also set to some
small random values initially and then will be
modified while training algorithm.
In case of Twin Burger 2, there are totally eight
series of simulation results in Surge, Sway, Heave,
and Yaw motion modes for SDFNNIs, and CMNNI.
Simulation results in Sway mode of CMNNI, which
includes the most severe coupling effects, are
illustrated in Fig. 4 as an example. For more
additional examples refer to the other publications by
the authors (Sayyaadi et aI., 2000). It can be
concluded that the proposed identifiers are very good
ones to model dynamics behavior of any kind of six
degrees of freedom rigid body. In the velocity chart
of Fig. 4, the simulated velocity using SDFNNI is
shown for comparison. By looking at this chart, it
can be concluded that SDFNNI is not a suitable
identifier for the problems of coupled type.
Training algorithm for neural network controller
(position control) is carried out using Supervised
Learning, and Unsupervised Learning together. In
case of SDFNNC, a rough controller, rudimentary
controller, is used to generate the primary data sets
for Supervised Learning. This controller cannot keep
the controlled object in the desired reference level,
because it is not tuned up well. But it is very useful
at the beginning of the training procedure where
there is no any kind of knowledge about the control
action. This step of training is just for initialization
of the controller network. For each different mode,
synaptic weights of SDFNNC are initially set to
some small random values, and then are modified to
imitate rudimentary controller.
170
Supervised Learning (initialization algorithm) for
SDFNNC is carried out in two steps. In the first step
inputs to the network are position error, velocity
error, and control signal of the past time step, all
come from the data pack generated by the
rudimentary controller. Evaluation function is made
from the discrepancies between control signal of the
next time step simulated by the network and that of
the rudimentary controller in the data pack. In the
second step of training, inputs to the network are
position error, velocity error, and initial control
signal come from the data set, and control signal of
the next time step is fed back while learning
algorithm progresses. The rudimentary controller is
a simple positive/negative switching mechanism
around target point.
The next stage of training for SDFNNC is fine-tuning
up of the controller network, because the initialized
network cannot make the system converge to the
desired point due to its oscillatory behavior around
target point. This stage is adaptation algorithm and is
carried out using Unsupervised Learning. In this
stage identifier network and controller network are
considered as of a unique network for adaptation.
Inputs to the network are position error, velocity
error, and initial value of control signal. State
variables simulated by the identifier network are fed
back to the input layer as well as control signal while
training algorithm progresses. Evaluation function is
weighted summation of the position error, and
velocity error. Error Back Propagation is used for
adjusting the synaptic weights and only the weights
regarding to the controller network are modified.
Initialization algorithm and adaptation algorithm are
applied to four independent motion modes of Twin
Burger 2, and the results for its Sway mode are
illustrated in Fig. 5.
In case of MIMO system, it is not possible to make a
rudimentary controller due to the coupling effects
between different modes. Thus the training
algorithm for CMNNC is only based on
Unsupervised Learning (adaptation algorithm). The
additional synaptic weights, used for interconnecting
of the independently trained SDFNNCs in a unique
CMNNC structure, are set to some small random
values at the beginning of the training procedure and
then will be modified during training. CMNNC and
CMNNI are considered as a unique network and
Error Back Propagation rule is applied to this
network in such a way that the only weights relating
to the controller part are modified. Adaptation
algorithm is carried out in two steps. In the first step,
all of the reference values are set to zero and training
is carried out. Inputs to the network and evaluation
function are similar to those of SDFNNC while
adaptation algorithm. In the second step all of the
reference values are set to the desired values and the
same procedure as of the first step is repeated. Fig. 6
illustrates the performances of the CMNNC for Twin
Burger 2.
3. PATH PLANNING OBSTACLE AVOIDANCE
MODULE FOR AUVS

The last problem addressed in this paper is path

planning for obstacle avoidance module. This
module has neural network architecture also and is
trained using Reinforcement Learning. Neural
network structure here is composed of two different
networks, which are Predictor Unit, and Learning
Unit. Predictor Unit is responsible to predict the
reinforcement signal for the next time step and its
output is treated as of standard deviation corresponds
to a normal distribution function. Training algorithm
for this unit is based on Supervised Learning and
Error Back Propagation rule is used to minimize the
evaluation function, which is made from the
discrepancies between actual reinforcement signal
and predicted one. Learning Unit is said as of the
main unit for the desired task and its out put is treated
as of mean value corresponds to the normal
distribution function, which its standard deviation
was identified as of output of the Predictor Unit.
Normal random variable corresponds to these two
parameters is the output of the module.
 Abstract: This paper deals with controlling of large ships in harbour manoeuvres by
applying the Decoupling Control Method. Ship motions are expressed by a multi-term
non-linear model. The Decoupling Control Method applied to the ship non-linear model
is described and issues concerning designing the controller are discussed. The method
and its excellent effectiveness are illustrated by simulation of a Very Large Crude
Carriage (VLCC) in a typical pattern of approaching and berthing manoeuvres.
Robustness of the controller is also studied by examining its ability in dealing with
influences of parameters' uncertainty and effects of environmental disturbances (such as
strong wind and current). Copyright ©200J IFAC
Keywords: non-linear control, decoupling control, ship control, tracking, robustness

1. INTRODUCTION To control such a non-linear system, a robust control

In harbour manoeuvres, ship's masters or pilots often
face the most complicated and sophisticated types of
ship handling. Therefore, from safety and economy
points of view, it is an essential task to develop a
controller to automatically guide the ships in harbour
manoeuvres.
To develop such a controller, several problems must
be solved. Among them the most difficult is the how
to lead the ships follow a desired trajectory precisely.
A suitable mathematical model of ship manoeuvring
motions in harbours and a proper control method are
then necessary for that purpose. The dynamics of
ships in harbour manoeuvres are fundamentally
non-linear in nature; therefore, to be able to describe
a wide range of ship manoeuvring motions in
harbours, a multi-term mathematical model of ship
motions is adopted in this study. The model, which is
suitable for describing ship harbour manoeuvres, was
originally presented by K. Kose et al. (1989) and has
further been developed by Le and Kose (2000), Le
and Nguyen (2000).
method must be employed. Over the last three
decades, the problems of achieving decoupling, or
non-interaction, in MIMO control systems has been
widely studied and it is not surprising that
Decoupling Control has been motivated by the needs
of a wide range of applications. The highly coupled
nature of ship dynamics in harbour manoeuvres and
the high performance requirements, together with the
lack of a good MIMO design procedure for the field
made this study essential. Moreover, since ships'
manoeuvring motions in harbours can be divided into
elemental motions (Kose, 1987), the Decoupling
Control would be the best choice for the control
method.
Although several studies concerning automatic
control systems for ships' harbour manoeuvres have
been carried out recently (Iwamoto, 1999; Ogawara
and Iwamoto, 1998; Berge, Ohtsu, and Fossen, 1998;
Ohtsu, Takai, and Yoshihisa, 1991), in most of those
studies, bow and stem thrusters were used as the
means to provide controlling forces and moment.
However, control of large ships in harbour areas

179
 Abstract: Due to more largely becoming ship units its manoeuvring freedom in
enveloped sea areas is limited. Therefore a more exact knowledge of the sea area
conditions and variable disturbances in the zone of manoeuvre than with smaller ship
units is necessary. The on-board sensor technology can determine only the body-related
variable disturbances. By using of an external acoustic doppler current profiler (ADCP)
in connection with a telemetry mechanism and an up-to-date identified ship's model the
calculation results of the manoeuvre predictor should be improved. The harbour
approach Puttgarden will be the example area for the investigations. Copyright © 2001
IFAC

Keywords: Naval automatic control, Manoeuvre prediction, ship manoeuvre simulation,

Shipping, Navigation

1. INTRODUCTION the on-board sensor technology fulfils these special

Manoeuvre prediction systems are well-known since
introduction of the computing technique also in
shipping. Their application results in the increase of
the ship's safety when manoeuvring in narrow
channels and docklands in particular with
inexperienced ship management personal. In addition
the ships become more largely due to economic
conditions. This results in a restriction of ships
manoeuvring area because the dimensions of the
narrow channels are remained more or less constant.
A certain compensation of these restrictions will be
achieved by modem manoeuvring mechanisms, as
Schottel propulsion devices and thrusters. However,
these aids due to a strong dependency from velocity
are usable only conditionally.
By using a manoeuvre prediction device as an
objective manoeuvre consultation the navigational
safety of vessels has to be increased. An objective
manoeuvre consultation represents the basis for safe
decision making. The objectivity can be achieved on
the one hand by the identification of the actual
manoeuvring capabilities of everyone of the special
actuators, which depends on the load condition of the
ship and on the other hand by the determination of
the variable disturbances wind and current which are
effective for the manoeuvring behaviour. The
generation of appropriate measured values only from
demands on objectivity of the consultation only
concerning the self-behaviour of the ship and the
behaviour of disturbances to the point of start (time
and position) of the prediction. Therefore objective
manoeuvre consultation is only possible with such
devices if either no disturbances for the manoeuvre
zone are available or the ship is moving in a
homogeneous field of wind or current.
For the reasons mentioned above the following
specified requirements to an objective prediction
system for ship manoeuvres can be derived:
- Adaptation ability of the used mathematical ship
model for the consideration of the manoeuvring
characteristics which are depentent of the load
conditions
- Inclusion of sea-area-specific disturbance
variables into the prediction model by external
sources of information
- high acceptance through the ship management
personnel which can be achieved by simple
operation and analysis in integrated navigational
systems as well as transparent representation
forms
The network project MAPSYS (MAnoeuvre
Prediction SYstem for Ships), promoted by the
German Federal Ministry for Education and
Research, has the objective among other things to
improve the result of the manoeuvre prediction at the

185
which form these data are available in other
navigation systems. The data of the enumerating I to
4 were measured directly as analogue voltage signal
of ±10 V. These are supplied to an analogue input
modules with high impedance of a field bus node.
The field bus node converts these voltage signals into
appropriate NMEA 183 strings and put it out byaRS
232 interface module. Here it is mentioned that the
NMEA 183 standard is no longer sufficient with
dynamically relevant data, because for applications
of real times a data inquiry of 0.1 seconds is
necessary.
The main item of the manoeuvre prediction device
represents an industrial suited Pc. All signals arrive
into this so-called MAPSYS computer without direct
electrical connection. They are sampled here in real
time and stored as data files, in order to use these
from other PC's or consoles e.g. ECDIS. An
appropriate identification software tool calculates in
the so-called Sleep mode the actual parameters of the
ship motion model depending on the load condition
from the measured ship data and. disturbance
variables. In the Active-mode the MAPSYS-PC calls
automatically the necessary data of the sea area-fixed
disturbance model, here the model of flow, from sea
area stationed component. Afterwards the actual
prediction is executed with the aim to calculate the
control signals for the actuators for the ship motion
along a given track line.
The electronic sea chart (ECDIS) serves as user
interface. This one is an electronic data base with
graphic surface for the representation of the sea area
topography, navigation marks and other navigational
information. This becomes generally accepted in on-
board operation due to rationalisations and increasing
system integration. Here the appropriate control
instructions are given and the prediction results are
displayed by its own software module. Additionally
the sea-area-specific disturbance variables are to be
represented.
For the correlation of the on-board measuring data
the data of an additionally installed inertial sensor
(GUI 3023 from Perform Tech company) are
recorded. This one provides the earth-fixed direction
angles, the body-fixed velocities, rates of direction
angles as well as the body-fixed accelerations in a
binary format with a data refreshing rate of 60 Hz.
2.2 Structure of the sea area-fixed system component
Enveloped sea area specific information are
necessary to increase the prediction quality, because
all area information is not available always from on
board sensors and data bases, e.g. ECDIS. The
approach of the ferry harbour Puttgarden was
determined as an example area for the project
MAPSYS in order to examine the influence of the
187
locally strongly changing currents on the ship
motion. An ADCP-sensor from RDI company was
procured to get some real time information of
currents in the sea area. Several wind sensors or
Internet services, e.g. charts of currents, come in
consideration as further data sources depending upon
sea area. These are not considered for the time being
for an on-line processing. However, they serve for
the generation of a local stream model within the area
of harbour approach. The determination of the
surface flow by means of high frequency radar
technology, how it is used in the port Rotterdam,
separates from cost reasons.
The head of the ADCP sensor was installed in a
distance of approximately 200 m from the east mole
head and outside of the harbour approach. Fig. 2
shows the flow sensor head in its sea ground
anchorage before sinking. A GSM modem for the
transfer of the data to a ship, which operates in the
D2-radio-net, at the mole head was installed. The
converted stream data will transfer as ASCII or
binary-coded data. These data can be received by a
further ship based GSM modem if necessary. They
are made available the MAPSYS computer, which
consults them for the calculation of prediction and at
the same time provides for a display in the ECDIS.
By the wireless transfer it is further possible to make
these data accessible also for different users e.g. the
port captain or the vessel traffic service management.
By this means a substantial advantage lies in relation
to the situation momentary.
3. INFLUENCE OF THE CURRENT ON THE
MANOEUVRING BEHAVIOUR OF SHIPS
The known theories of the description of the motion
of ships suppose the motion on the undisturbed
surface of the sea. In consequence influences of
current in modem simulators are realised by a
constant drift in the inertial reference system'
because a constant impulse is transmitted to the ship
(Hilten, 2000). This assumption is always ingenious
in such cases, if the ship moves in the free sea area in
which the change of gradient of current is very slight
 Abstract: In this paper the problem of ship trajectory tracking with £-accuracy is
considered. Results presented are based on a three-degrees-of-freedom model with full
dynamic interaction between motions in roll, sway and yaw. The aim of control is to
minimise the influence of wave disturbances on roll motion and to track with £-accuracy
a ship trajectory. For this purpose a linearised and a non-linear ship model are used. The
Matlab-Simulink simulations confirm the viability of the proposed method for the high
precision control of ship track-keeping. Copyright © 2001 IFAC

Keywords: tracking systems, MIMO, marine systems, disturbance rejection

1. INTRODUCTION to efficiently and safely carry out demanding

The remarkable growth of transport of passenger and
cargo at sea as well as the initial exploitation of ocean
resources has determined the construction of an
increasing number of new surface ships and
underwater vehicles. For an efficient operation at sea,
it is essential that these craft should be equipped with
advanced control systems. Control systems to be
installed on board ships are generally designed in such
a way to reduce fuel consumption, to minimise
disturbing wave induced motion and at the same time
to improve navigation accuracy. It has been verified in
the recent years that traditional control methods are
generally inadequate for designing efficient control
systems.
Design solutions for marine craft motion control are
often rather difficult to find within the context of
classical control theory, owing to the intrinsic non-
linear dynamic behaviour of the plant itself and to the
disturbances, which act upon it. Many researches and
simulation studies have been conducted in the recent
years in order to design and put into operation a new
generation of ship control systems, such as autopilots,
stabilisers and dynamic positioning systems, capable
navigation tasks in a widely varying range of
environmental conditions. In principle, the design of
such control systems should be based on a
multivariable approach, which takes properly into
account the couplings between the different motion
and determines control systems within the framework
of optimal control theory. A number of successful
simulation studies and sea trials have been carried out
which apparently supported such approach.
It is worth and surprising enough noting, however,
that the adoption of these optimal designs has been
until now quite irregular on board new ships, were old
fashioned PID autopilots are often preferred. This
apparent discrepancy is partly due to the fact that
most of the proposed designs are critically dependent
on the availability of accurate mathematical models of
ship and environment, which are generally quite
complex and difficult to be determined and properly
on-line tuned. It may be therefore attractive to explore
the applicability of robust control methods, which
potentially can reduce the negative effects on control
system performance of the uncertain factors affecting
the ship dynamic behaviour. Most of these methods
can be derived by an H~ optimization approach

191
 ON THE DESIGN OF ALERTS FOR
FUTURE PLATFORM MANAGEMENT SYSTEMS
Michael A. Tainsh, F.Erg.S., Eur.Erg.

Centre for Human Sciences, Defence Evaluation and Research Agency,

Farnborough, Hants GU14 OLX, UK

There is a requirement for improved user performance with warnings

and alarms within the Platform Management Systems (PMSs) on
current ships. A programme has been conducted for MoD/Ship
Support Agency to investigate the optimal design characteristics of all
alerts for future PMS. The past approach to alarms and warnings was
considered 'equipment centred' while the approach taken here is 'user
centred'. The requirements for future major Royal Navy platforms
were considered: the likely functionality of their PMS workstations,
the means of on-line control, and concepts for the design of the
human-computer interaction. Hypotheses were formulated on
potentially desirable categories of information to be associated with
alerts, These categories are identified as reference (including function,
location and organisation), significance (covering redundancy,
timeliness and hazard) and statements of quantified information
including encyclopaedic material. Results from two trials with Subject
Matter Experts (SMEs) supported the hypotheses.

Keywords: alarm systems, human factors; human centred design

1. INTRODUCTION

The control of a ship and its services is carried out
within the Platform Management System (PMS).
There is sound evidence that there are problems
within current Royal Navy (RN) PMSs. The status
information, warnings and alarms are displayed,
acknowledged and controlled in ways that do not
match user characteristics. In fact, the general design
concept that appears to have been applied in many
cases seems to be more associated with the design of
electromechanical dials. In this case, the physical
measure from a sensor (status) is available plus
additional information on whether it is slightly
different from a desired range (a warning) or
substantially different from that range (an alarm).
This is an 'equipment centred' approach to displaying
information and takes little account of the users'
characteristics. In the past, this may have been the
only information that could have been provided in
non-computer based systems. However, in current
and future systems this can be judged a very limited
approach.
The Defence Evaluation and Research Agency
(DERA) Centre for Human Sciences (CHS) was
approached by the MoD/Ship Support Agency (SSA)
to study the handling of alerts in future RN ships to
support procurement programmes in line with current
practice (References 1, 2 and 3). It was agreed that
the study should employ a number of techniques,
including:
(a) Top-down analysis of known assumptions,
high-level issues, or system characteristics in
order to derive likely human factors issues;
(b) Observations of lessons learned in the past
on similar projects and general concepts of best
practice;
(c) Exercises in early-prototyping, trials and the
use of Subject Matter Experts (SMEs).
In this case, all three were used to investigate design
options and mitigate risks. The work described here
addresses the issues associated with future ship PMSs
and aimed for entirely novel design solutions.

203
 2. DEFINITIONS
The initial high level issue stems ftom the established
set of categories for describing alerting information:
status, warning and alann. The established set was
seen as 'equipment centred' based on what electro-
mechanical technology could offer. These are known
to be associated with displays that are poorly used
and disliked on some RN ships. In this study, a
provisional definition of an alert was adopted: the
mechanism by which a user is infonned of a change
of information. This general definition provided a
starting point for an empirical study. This definition
of alerts is both arbitrary and general. However, it
was believed that this defmition would ensure that
there would be a broad opportunity for a wide range
of possible design solutions to arise ftom the
investigation. It was hoped that this would be
achieved by adopting the general definition that
makes no a priori distinctions between the categories
of alerting information that might be provided to a
user.
3. AIM OF THE INVESTIGATION
The aim was to investigate the categories of
infonnation (a taxonomy) that users would prefer to
be associated with alerts, in contrast to the set that
appears to be assumed in all cases: status, warning
and alarm. The aim was to be achieved through a set
of trials which involved SME's , and the analysis of
users' beliefs, knowledge and perfonnance.
4. THE STUDY
4.1 The Approach
The study proceeded in five stages:
(a) Consideration was given to the set of possible RN
platfonns currently under investigation for future
procurement, and their operations. The main concepts
to be taken into account here include the
consideration of 'Whole Life Costs' which has
substantial implications for the introduction of
automation and the need for greater integration of all
ship systems.
(b) A description was obtained of the possible
functionality of a future PMS workstation including
the equipment and plant that was likely to be
associated with it, and the user organisation. It is
envisaged that jobs/roles and tasks may be radically
changed as new possibilities for automation are
introduced. These will have profound consequences
for work organisation, complementing and the
layouts of the ships. Improved communications mean
that the internal arrangement of spaces is likely to be
204
quite different in the future ftom any ships in current
operation.
(c) An understanding was gained of the alternative
means of on-line control of the plant, equipment and
services likely to be available to the PMS, including
automation and support. The control of the PMS is
likely to depend on more sensors and more processed
infonnation. In particular, the use of automation to
reduce the need for the user to execute manual
control functions was seen as central to the
consideration of the users' roles.
(d) A concept of the Human-Computer Interaction
(HCI) for a future PMS was develo~ed. The
implications for the HCI are profound. There may be
a central workstation but equally the user may act
ftom a remote position.
(e) The generation of implications of the HCI concept
for the provision of alerts. It was inferred that the
scope of alert infonnation should cover three main
categories:
(i) Reference infonnation. This covers the
functions carried out by the equipment and plant,
the organisation of the ship's company and in
particular the team associated with the PMS, and
the spatial layout of the ship and the plant and
equipment within it.
(ii) Significance. Priorities must be indicated
associated with hazards, availability of
equipment and plant (redundancy) and the
temporal factors associated with a failure or other
undesirable state.
(iii) Statements of quantified infonnation. This
was considered to include infonnation associated
with the state of the equipment and plant, history
including trends and allied encyclopaedic
information.
4.2 Implications For Trials
The findings from the study led to the hypothesis that
future users may wish to have alert information that
covered the above set of categories. The questions
then are:
(i) Precisely what infonnation might the users
want associated with each category;
(ii) how may they wish to have this infonnation
presented to them.
The emphasis in the first trial was intended to focus
on the first of these questions.
 5. TRIAL ONE The tasks were:

5.1 The Approach · Platfonn Engineering Officer of the Watch comes

on duty
A scenario for equipment and plant design and · Automatic plant reconfiguration
configuration and operation was constructed based on · Galley fire
possible options for a new aircraft carrier. This was · Platfonn Engineering Officer of the Watch watch
chosen because work is still in its very early phases handover
and none of the Subject Matter Experts (SMEs) · Planning of maintenance routine
involved in the trials would have any well-fonnulated · Go to State One
views about existing or possible design solutions. · Go to flying stations
· PMS automatically reconfigures to match load
Three sets of trial conditions were constructed with requirements
ten SMEs (two were used to pilot the techniques with · Failure of weapons lift
the remaining eight used to generate the results). The · Failure of gas turbine
starting point was a set of thirteen tasks that could be · Go to action stations/readiness state
taken as representative of a day in the life of the · Fire in gas turbine compartments
Marine Engineering Officers of the Watch and · Failure in heating, ventilation and air conditioning
Platfonn Engineering Officers. The fonner group is Examples of the trials material are given in Figures 1
typically highly experienced Chief Petty Officers and 2.
while the latter are typically senior officers.
Tool) Laboratory at DERAICHS, Portsdown. Each
SME was interviewed separately.
5.2 Results
The conversations with the SMEs were analysed for
content. They provided evidence of wide range of
preferences. In particular, the SMEs requested the
supplementary information that could be provided
through the addition of reference information on
plant and equipment availability. The SMEs were
particularly keen to have information on the
significance of an alert. This included summary
information on the availability of plant and
equipment and the provision of information on the
priority of the alert in terms of the time required
before it had to be actioned, and hazard
considerations.
These results imply a set of display and control
designs that could enhance user performance and it is
believed it could improve the safety of the system.
The user centred approach based on a combination of
top-down analysis and a user trial resulted in a
concept for alerts that is both novel and in line with
SMEs' preferences for future ships. This supports and
elaborates the general definition of an alert that was
stated earlier.
207
 components of object motion (vx" Vy) analytically
detennined on the basis of consecutive
measurements. Each network smooths out one
component of motion vector V. The network is
dynamically expanded as new information about
the target is acquired. The measurement sequence
thus constructed is the teaching sequence of the
network, which has exactly as many radial neurons
as there are elements of the measurement sequence.
The maximum length of the measurement sequence
should be detennined experimentally depending on
the kind of measurements and the characteristic of
their errors. The suggested length of this sequence
is contained within 40-60 measurements. The
remaining measurements not included in the time
window are subject to oblivion. In this way, the
measurement sequence, on attaining an established
maximum value, is successively renewed, which
means that the appearance of a new measurement
brings about the oblivion of the oldest one. A single
GRNN network pair produces a fairly well
estimated response, but the oscillations are still too
strong. This was the reason for applying the second
smoothing stage. The response of the first output
network pair is given to the output of the successive
pair. At the output of the second-stage smoothing
there are already obtained suitably smoothed
components of the object motion vector, which later
serve the purpose of detennining its course and
speed. The two-stage neural state estimator thus
constructed, as proved by numerical experiments,
fulfilled the requirements of functioning in
conditions with large error occurrence and unifonn
. rectilinear motion.
values of speed and course do not exceed the
established tolerance limits. On ascertaining that
the ship's movement parameters have been
stabilized, the system goes over into the stable work
variant. In this variant the filter has a double-stage
structure and lengthened teaching sequence, and at
the same time this sequence is gathered anew from
the moment the manoeuvre of the tracked target
was ended.
4. RESULTS OF NUMERICAL
EXPERIMENTS.
In order to verify the method worked out an
application using NeuroWindows (Ward Systems
Group, 1993) has been constructed. The unit's
movement was simulated, which in the 2nd
observation minute made a right turn of 40°. This
tum took one minute to take place. The unit's
position was detennined within one minute. The
above was accepted in accordance with IMO
requirements. The condition for manoeuvre
detection was the detection of course alteration by
5° or speed up to 1 knot allowance. The condition
for returning to stable functioning was course
oscillation up to 5° tolerance or speed with
tolerance up to 1 knot. Experiment results have
been presented on particular diagrams. Error values
of the estimated speed and course of the tracked
target have been presented on the vertical axis in
degrees and knots respectively.

In order to adapt the system for functioning in non-

linearity conditions a manoeuvring detection
system was applied consisting in controlling the
course and speed of the target calculated on the
basis of the network's response. The manoeuvring
detection system is activated after the first 20
measurements. In a situation where it has been
stated that three consecutive calculated values
exceed the established deviation values in relation
to the average value of ten preceding measurements
(of course or speed) in the same direction of
changes, the filter changes over to the manoeuvring
variant. At the same time it should be noted that the
value exceeding the allowance is not taken into
account when calculating the average value. In the
manoeuvring variant the second network pair is
excluded and the maximum length of the filter's
teaching sequence is reduced. The length of this
sequence is detennined experimentally and should
amount to about 20-30 measurements. In the
working period of the manoeuvring variant the
filter works out momentary values of the object's
course and speed on the basis of a reduced
measuring sequence, until it is found that three
consecutive differences between the calculated

213
 ENHANCED NEURAL NETWORK MODELLING FOR PROCESS FAULT DIAGNOSIS

T K Chang, D L Yu and D Williams

Control Systems Research Group, Liverpool John Moores University, UK

Abstract: A neural network (NN) based fault detection and isolation (FDI) approach for unknown
non-linear system is proposed. This FDI scheme is able to detect both actuator and sensor faults. An
enhanced parallel (independent) NN model is trained to represent the process, and used to generate
residual. A mean-weight strategy is introduced to overcome the unmodelled noise and disturbance
problem. An information pre-processor is implemented to convert the quantitative residual to
qualitative form and applied to a NN fault classifier to isolate different faults. The developed
techniques are demonstrated with a multi-variable non-linear tank process. Copyright © 20011F AC

Keywords: Fault detection, fault isolation, neural networks, non-linear systems, process
identification.

1. INTRODUCTION the mathematical model by a NN that has been

Over the past two decades, much attention has been
paid to the problem of FDI in dynamic systems. All
present FDI strategies can be classified into model-
free or model-based (Liu, 1996). The model-free
methods directly examine the plant measurements,
these methods include limit checking, frequency
analysis, and expert systems. The model-based
approaches for linear systems can be classified
(Patton, 1994) as follows:
1. Observer-based methods or observer-based
residual generation (Frank, and Wunnenberg,
1989).
2. Parity space methods or parity equation residual
generation (Frank, 1990).
3. Parameter estimation method (Isermann, 1993).
The above approaches work well for linear systems.
However such methods fail in the presence of
significant non-linearities in the system dynamics.
Hence some authors (Seliger, and Frank, 1991;
Hengy, and Frank, 1986) have proposed residual
generation schemes that use the theory of non-linear
observers. Recently, a few non-linear, observer-based
methods have been developed for certain classes of
non-linear systems, for example, Seliger and Frank
(1991), but these appear too complex to be of
practical use.
The model-based approaches to FDI in automated
processes have received considerable attention for
the last two decades. However, an accurate
mathematical model of a complex non-linear
dynamic system is difficult and costly to develop
under realistic situations. One solution is to replace
trained to adapt to the process input-output
behaviour. It has been proved that certain NN
architectures, such as the multi-layer perceptron
network (MLPN) and radial basis function network
(RBFN), can represent any non-linear mapping
between input and output with arbitrary accuracy by
given suitable weighting factors and architecture
(Narendra, and Parthasarthy, 1990). For FDI with
NNs, the NNs are normally used for classification,
output prediction and state estimation (Kolvo,
1994).
In most known publications (Kolvo, 1994), the NN is
applied as a fault classifier based on the process
measurements. In such diagnosis systems, only
system output is used. For some diagnosis problems,
this can be valid. However, for FDI in non-linear,
dynamic systems this is not the case as system input
could have an effect on features of system output. An
effective approach to solve this problem is presented
in (Frank, 1995; Patton, et ai., 1994), where the
residual generation concept known from the model-
based methods was applied (Patton, et ai., 1989).
Most of the NN-based FDI approaches for unknown,
non-linear systems contain detection and isolation
parts (Yu, et ai., 1999; Patton, et ai., 1994). In their
approaches, a series-parallel NN is implemented to
perform one-step-ahead prediction of the system
output, and another NN, known as fault classifier is
trained to learn the fault patterns. Because series-
parallel NN model uses the past process output with
fault effect as input, it will be influenced by the
actuator or sensor faults. Therefore it will lead that
series-parallel NN follows the faulty system output.
The residual generated by this series-parallel NN is

215
small and sensitive to modelling error. The fault
classifier will suffer from this small residual. Ideally,
a parallel NN should be used to generate residual.
However, an efficient algorithm for a parallel NN is
not available, due to its recurrent operation.

The FDI techniques (Zhou, and Frank, 1998; Wang,

1995) are mostly developed for either actuator or
sensor fault only, as they assume that only actuator or
sensor fault occurs in the system. Those actuator or
sensor FDI methods should be used for industrial
systems.

In this paper, the proposed NN-based FDI scheme is

able to detect and isolate actuator and sensor faults in
an unknown, non-linear system. An enhanced
parallel NN model is developed to improve the
parallel modelling accuracy. This enhanced parallel
NN model contain two NNs, the first NN is trained
by the series-parallel model method, then
implemented as a parallel model. Because this
parallel model is trained with series-parallel method,
its parallel modelling error is greater than the series-
parallel modelling error. Therefore, the second NN,
called a parallel modelling error reducer, is
introduced to reduce the parallel modelling error.
Realistically, unmodelled noise and disturbances are
unavoidable in any network model. Thus, a mean
weight strategy is introduced to overcome them. The
strategy uses three enhanced NN models to average
the estimations from these three network models.
Thus, unmodelled noise and disturbance are
compensated by network models each other.

The difference between the parallel NN output and

process output is used as residual to detect faults. For
fault isolation, an information pre-processor is used
to convert the quantitative residual to qualitative
form. This qualitative information is then used to a
NN, known as fault classifier, which is trained to
learn the fault patterns, which could be obtained from
the past fault records. The developed NN-based FDI
techniques are demonstrated in a multi-variable, non-
linear tank process. The MLPN is used in this NN-
based FDI simulation.
 Abstract

In both naval and commercial marine sectors, ship owners and operators are faced with the need to
operate sophisticated vessels with lean-manning levels. This is graphically illustrated by the goals of
the DD21 program in the US, where the target complement of95 is one of the main drivers behind a
70% reduction in operating costs. Such lean-manning policies are one motivator towards increasingly
sophisticated data-driven systems being fitted on-board. The design of such systems is no longer the
remit of purely control engineers, but relies heavily on both on human science and simulation.

This paper outlines the changing customer requirements in advanced marine applications and the
impact of these trends on a major supplier of marine systems. The development of advanced Platform
Management Systems in support of cost reduction and lean manning is then reviewed in more depth,
with particular focus on the role of human factors and simulation.

Keywords: Automation, Dynamic modelling, Human factors, Operability, Simulation.

1. INTRODUCTION Management is a function, rather than a

The drive towards reduced construction and
ownership cost is giving rise to corresponding
increases in the adoption of ship automation
and advisory systems. Depending upon how
well such systems are developed, configured
and tested, they can either detract from, or
enhance, the operational effectiveness of the
ship.
Operating a highly automated platform with a
reduced complement gives rise to a new set of
Human Factors challenges and risks. Operator
workload needs to be carefully managed: too
high a workload increases human error rate,
especially in high stress situations; too Iowa
workload may reduce situational awareness.
Either extreme can have a negative effect on
damage readiness, performance, morale and
staff retention. The successful implementation
of ship automation systems can only be
achieved by understanding the complex
interactions between the human operator and
the machine. In this sense, Platform
system. A Platform Management System,
together with the operators, together form a
complex, socio-technical "meta-system",
involving the successful interaction of ship's
crew and machines. It is also clear that the
successful implementation of such systems is
no longer the exclusive domain of Controls
Engineers.
This paper describes the importance of Human
Factors in this area, and describes how Rolls-
Royce is integrating such techniques into the
system engineering process. It goes on to
review the role of simulation and modelling in
the development of the successful Platform
Management Systems of the future.
2. KEY MARKET DRIVERS
2.1 Whole Life Cost
Analysis of through life costs shows the clear
linkage between manning levels and total cost
of vessel ownership. Defining a suitable

221
complement is a complex issue; it is a function
of the vessels operational role, battle damage
policy, platform design, career structure and
other issues. Pressures on operational budgets
are driving down manning levels. Crew
reductions may be achieved by a number of
means, including (but not limited to) further
automation. This requires trade-offs to be
carried out between Non-Recurring Costs
(NRC), Unit Production Cost (UPC) and
Through-Life Cost (TLC), leading to
automation as a "spend to save" opportunity.
2.2 Risk Management
Availability of suitable technology is not
generally seen as a major barrier in the
development of marine automation schemes.
Advances in shore-based industrial automation
systems have provided a plethora of
commercial off the shelf (COTS) equipment,
which in general can be configured for use in a
maritime environment. However, it is the
deployment of this technology in a fighting
vessel which brings with it technical and social
issues which are unique, and require special
considerations. The Royal Navy, the Defence
Procurement Agency (DPA) and the Prime
Contractors wish to reap the benefit of this
technology, whilst managing the attendant
problems. This is highlighted within the UK
MoD, who have recently placed a five-year
contract aimed at de-risking the application of
PMS to future Naval Platforms. An industrial
team, comprising BAE Systems, Vosper
Thornycroft Controls and Rolls-Royce, will
execute a program of studies and possibly
trials, focussed on mitigating and avoiding the
risk associated with the application of
advanced technology in Naval controls &
surveillance systems. These studies will
consider the functional and non-functional
requirements including safety, human factors
and supportability. The work will support the
Future Surface Combatant, Future Carrier
(CVF), Future Attack Submarine (FASM) and
other ship programmes.
2.3. Controls Integration
Control schemes can be classified within a
simple five-layer model, which is given in
Figure 1. The simplest controller (perhaps
locally controlling a winch) would correspond
to Levell. A controller for a collection of
local equipment comprising a sub-system,
would fall into level 2 (examples of this
would be a gas turbine controller or an
automatic stabilising system). Level 3
integrates entire ship systems such as
222
command, electrical power, propulsion,
damage control. Level 4 examples would be an
integrated bridge system or PMS which
integrates together Level 3 functions. Level 5
covers Aftermarket and Fleet Management
applications, which are increasingly web-
enabled.
Shipbuilders, owners and end-users have
recognised that there are tangible benefits if
the different control systems throughout the
vessel can be specified and procured in a co-
ordinated manner. This gives opportunity to
reduce the number of different control systems
(and suppliers) involved, hence reducing
interfaces and risks. A reduction in the
number of different hardware platforms offers
savings in carried-on-board and carried-at-base
spares inventory. Where security and
availability considerations allow it, there is
scope for integrating control functionality from
various levels, on a common network and
operating platform. This offers potential for
reducing passive inter-connections, cabling
and bulkhead penetrations, and allows operator
interfaces to be presented in a consistent,
integrated manner.
Figure 1 Control System Framework
This is of course the territory of Integrated
Platform Management Systems, which offer
the potential to provide integration in two
orthogonal directions. Horizontal integration
can be achieved using a resilient ship-wide
network to allow control and information to be
distributed around the vessel. Vertical
integration between operational controls,
health monitoring, communications, CCTV
and information management applications are
also possible, with data sharing and fusion at
all levels. This is analogous to the industrial
approach of sharing data from sensor to
boardroom.
With the operator effectively out of the loop
for much of the time, and working in a
supervisory role, situational awareness may be
impaired. For example, the designers of
autopilots and Dynamic Position controllers
are conscious to avoid the potential for
auto/manual mode confusion errors.

To overcome this, previous work on meta-

systems has proposed a grey-scale of task
allocation. However, this is also not without
its drawbacks either: assigning "confirmatory"
tasks to human operators may in theory help
maintain situational awareness, this is at the
expense of workload margin. What is worse,
Figure 2. Use and Re-Use of Simulation
confirmatory tasks may be perceived by the
throughout the Project Life-Cycle
user as an unnecessary and irritating
"overhead". Other examples in the category of
Models are typically generated in the pre-
overheads are navigating tasks such as
contract phase to demonstrate key system
windowing and zooming operations. An area
performance parameters and to evaluate
of interesting development is dynamic function
alternative system concept options. Once a
allocation, where the responsibility for control
preferred option is identified, more detailed
actions varies dynamically in accordance with
system models are generated to confirm
some parameter or rule-set (eg user preference,
system performance and optimisation.
workload, damage control state).
Within Rolls-Royce, high fidelity models are
_3._5Workload Analvsis
used to implement full-scope manoeuvring
room simulators. These are used as a formal
Techniques and processes for workload
media by the RN for operator training and
analysis can also be aided by the application of
qualification.
computer-based tools. Whilst this is
sometimes thought of as an imprecise science,
The same models used for trainers are also
numerical methods have a useful role to play,
being used to stimulate the prototype control
and such tools allow workload budgets to be
system, which is useful not only as a realistic
shared between operators, and workload peaks
dynamic test facility, but also for operability
to be identified and designed out.
assessment. This technique is proving useful
in developing the PMS for the Astute Class
:!. MULTI-ROLE SIMULATION
propulsion system, covered in Section 5.
METHODOLOGY
Once the control scheme is finalised, dynamic
Of course, modelling and simulation has been
testing work may proceed followed by final
used for many years in the development and
implementation and acceptance testing.
evaluation of marine equipment and systems.
Whilst not removing the need for prototyping
Rolls-Royce are pioneering the technique of
or physical trials, they are increasingly useful
embedding a model of the system under
techniques for de-risking new designs. As
control and its control system within the
considerable resources are expended in the
controller itself. The real system performance
development of such models, they should be
can be automatically compared with the
re-used as far as possible. It is important that
reference data, and differences/trends used to
modelling is seen as an investment, which is
locate faults and inform maintenance
protected and exploited.
management decisions.
Simulation can be used in a wide variety of
Such models are of increasing importance in
ways throughout the lifecycle, and this is
the aftermarket - especially when products are
shown simplistically in Figure 2.
sold on a Power by the Hour™ basis.

224
 ~ SMCTD
A number of years ago, Rolls-Royce
developed the Submarine Machinery Controls
Technology Demonstrator (SMCTD). This
facility comprises a dynamic simulation of a
full-scale control console, connected to a
model of a generic nuclear submarine
propulsion system and electrical system. The
facility (shown in Figure 3) has been used
primarily to evaluate advanced control
technologies, examples of which are SUR VEX
(an expert surveillance system offering
advanced operator diagnostics) and FEED (a
fuzzy-logic based boiler feed water control
system).
Figure 3 Submarine Controls Technology
Demonstrator
In recent years the SMCTD facilities had
become difficult to support, and like many of
the other Manoeuvring Room Simulators used
by the RN, there came a need to replace the
computer systems. In order to retain the
considerable investment made in generating
the model library, and to bring the facility up
to date, the models were re-hosted on a PC-
platform, using automatic code conversion
tools, developed originally for Simulator
lifetime extension. Hence SMCTD II was
created, and is currently being used to support
development of the propulsion and power
management systems for the Astute Class of
SSN's (under contract from CAE Electronics
Ltd).
Two key goals of this project are lean manning
and operability, with the consoles and PMS
being designed for a reduced Manoeuvring
Room team.
225
Following task and workload analysis, various
automation schemes have been identified
which allow the designers and future operators
to evaluate the design prior to full
implementation. Moreover, as the Platform
Management System is fully developed, the
SMCTD models can be used as a dynamic test
facility, allowing the system to be exercised in
a comprehensive range of operational
scenarios, including emergencies. This also
represents a useful training experience for
current and future ship's staff.
6. CONCLUSION
Marine controls engineers and hardware/
software "box" suppliers are enabling the
introduction ofPMS into Naval service. There
are significant technical challenges facing the
specifier, integrator and user of the advanced
ship automation systems of the future, for
example: supportability, obsolescence,
wireless interconnectivity, survivability,
software safety justification to name a few.
However, the successful introduction of
Platform Management Systems cannot be
carried out effectively by controls engineers
alone. This is an immensely wide-ranging and
broad subject area, and in the author's opinion,
no one person should pretend to be an
"expert".
What is clear is that the design of successful
Platform Management Systems in the future
will be significantly affected by human factors
considerations. The contractual acceptance of
operability will be a key element in the
handover of such systems. Human trust in the
technology, acceptance of reduced manning
levels and appropriate career structures and
training arrangements are greater risk
considerations than technological risk alone.
Obviously some of these factors are outside
the influence of marine suppliers. Industry is
turning to simulation as a route to the leaner
manning challenge: a challenge, which is
being faced universally as a method of
reducing operational costs.
Copyright © IF AC Control Applications in Marine Systems,
Glasgow, Scotland, UK, 2001

GENETIC ALGORITHM OPTIMISA TION OF OIL TANKER CONTROL SYSTEMS

Eva Alfaro-Cid and Euan W. McGookin
Centre of Systems and Control & Department of Electronics and Electrical
Engineering, University of Glasgow, Glasgow, Gl2 8LT, UK
phone: +44 141 330 6031, fax: +44 141 330 6004, email: alfaro@elec.gla.ac.uk
Abstract: The optimisation of the controller's gains for an oil tanker control system
using genetic algorithms is considered. Manual tuning of a PID controller's gains
can be a very slow process so an optimisation technique based on genetic algorithms
has been used to optimise the controller's gains. These gains have been optimised
for two different configurations: course changing and course keeping. Copvright ©
2001IFAC -
Key words: Oil tanker control, PID controller, Genetic Algorithm optimisation

1. INTRODUCTION are highly cross-coupled, the resulting model of the

The extremely rapid increase in tanker size that has
occurred in recent years, due to the necessity of
long distance transportation of big quantities of oil,
has made the vessels very difficult to manoeuvre.
Presently the length of a super tanker can be in
excess of 300 m. It follows that the size and limited
speed of these vessels makes them difficult to
manoeuvre quickly. However, their large bulk does
help with the stability of these craft.
...
Usmg automatIc controllers. for the. 011 tanker
cGntrol systems can solve thIS. In thIS work the
design of a controller based on a PID structure is
studied. This is a very simple controller whose
performance depends on the values of the
controller's gains. During the initial design process
Ziegler-Nichols method is used to obtain a first
approach and afterwards the controller's gains are
manually tuned to improve the performance of the
controller. Howevet:, this tuning approach is a
tedious and slow process.
A solution to this problem is the use of optimisation
techniques based on Genetic Algorithms (GAs) to
optimise the controller's gains. This paper presents
the results obtained after the optimisation of the
controller for two different configurations: course
changing and course keeping.
This investigation has been carried out through
simulation studies in Matlab based on the
mathematical model for an ESSO 190000 dwt oil
tanker (Fossen, 1994; McGookin, 1997). Computer
generated simulations based on this model have
proven to be sufficiently representative of the full-
scale manoeuvring dynamics of this large tanker.
~s ~onfidence in the accuracy of these
sImulatIons allows them ~ be used to a~curately
evaluate the manoeuvrabIlIty of these ShIpS (van
Berlekom and Goddard, 1972).
2. OIL TANKER MODEL
The mathematical representation used in this study
models the heading and propulsion dynamics of a
190000 dwt super tanker for transporting oil
(Fossen, 1994; McGookin, 1997; van Berlekom and
Goddard, 1972). Since these modelled dynamics
227
tanker is extremely nonlinear. Therefore, the
heading and propulsion dynamics can be
represented by the following nonlinear state space
equation (Fossen, 1994), where x is the state vector
and u is the input vector:
x = f(x,u) (I)
..
Althou~h m the. model ther~ are two ma~n
dynamIcs: propulsIOn and headmg, the automatIc
control system governs only the heading motion.
The propulsion dynamics are usually controlled by
simple steps commands and the dynamics of the
system self regulate the response.
The control of the heading comes trom the
regulation of the heading ang!e, If/, by manipulation
of the rudder actuator deflectIon. Consequently, the
system decouples to a third order system with states
the yaw rate, r, the yaw angle, I{I and the actual
rudder deflection, Or, and with the commanded
rudder deflection Ore as the input (McGookin, 1997)
This results in a decoupled, linear, SISO system of
the following state space form:
As it can be seen trom ~e equatio~ abov~, the
commanded rudder deflectIon, after bemg subjected
to rate and amplitude limits is a state of the SISO.
The~efore ~e is .a comma~ded .va!ue that may be
~utsIde t~e phYSIC~l operatmg lIilllts of the rudder
(I.e. ± 30 and 2.33 /sec).
3. GUIDANCE CONFIGURATIONS
The particular control systems are designed around
two PID controllers which provide course changing
and course keeping abilities for the tanker. The
difference between these two systems is that course
changing controller controls the heading of the
vessel irrespective of the vessel's positional course.
It reacts to heading commands trom a pilot, while
course keeping provides a positional feedback loop
that ensures that the vessel keeps to a
predetermined course quite accurately. In the
course keeping control system, the helmsman
 Abstract: The research considers a fast ferry with active appendages (actuators): aT-foil
near the bow and transom flaps. The objective is to attenuate the ship's vertical motions.
There is a problem of control design, to move the actuators in the most effective way.
For an easy and safe control study, a simulation tool has been developed using
MATLAB and SIMULINK: a modular architecture results where it is easy to integrate
control strategies to be analysed. Simulated experiments can be run with the tool, to see
the behaviour of the important variables: vertical motions of the ship, motions of the
actuators, sea-sickness incidence, etc. The paper describes the simulation tool from a
functional point of view, and explains some details of the main parts of its internal
structure. Copyright ©20011FAC

Keywords: simulation, ship control, marine systems, transportation control, MIMO and
mathematical models.

1. INTRODUCTION experimental and simulated studies have been

Our research deals with the attenuation of vertical,
heave and pitch, accelerations of fast ferries. For
such purpose, the use of aT-foil near the bow and
transom flaps is considered. By means of a control
system, these appendages must move to counteract
the effect of each wave. In consequence, there is a
control design problem to be solved.
Direct experimental tests on ships, or scaled down
replicas are time consuming and expensive (and
maybe dangerous on real ships). To be able to
conduct control studies on computer, reducing
experiments to a minimum, a simulation environment
is welcomed. Therefore a main objective of our
research has been to develop a good simulation
environment, focusing on the vertical motions of a
ship. The behaviour of the ship with actuators, under
the control strategy under study, must be predicted
with high fidelity.
Using the facilities of CEHIPAR (Canal de
Experiencias Hidrodinamicas de El Pardo, Madrid; in
English: EI Pardo Model Basin) a series of
accomplished, to establish the models for the
simulation environment (De la Cruz, et aI., 1998; De
Andres, et aI., 2000; Aranda, et aI., 2000). As a first
step, the research is restricted to head seas.
The simulation environment is designed for an easy
study of control design alternatives. Hence a modular
structure has been achieved, on the basis of
MATLAB and SIMULINK "modus operandi". The
results is that ,in effect, it is easy to incorporate any
control strategy into the environment, to see what
happens to the ship's response to waves when this
control strategy is applied.
Along the development of the simulation tool, a first
version was achieved. Some aspects of this version,
in particular those related to actuators modelling,
were presented in a previous paper (Esteban (a), et aI,
2000). The recent experiments made by CEHIPAR,
to test first control solutions, showed several new
aspects that should be considered for a more realistic
simulation. For instance there is a mixture of noise
and vibrations, so filters must be applied. Also, the
interactions of ship and actuators and the non-linear

267
 When an experiment finishes, all data obtained along
time are there, as computer files, together with other
information pertaining to a qualification of the
results. This information is in the form of numerical
indexes, to measure the general effects of the control
along the complete experiment. For instance, the MSI
(motion sickness incidence) index
(O'Hanlon, MacCawley, 1974), and other related to
the reduction of vertical accelerations. Therefore we
have the means to compare different control
solutions.

Figure 3 shows a row of three windows at the top of

screen. This is an example of the windows we can
open, resize, zoom, close, etc., to observe any
variable of interest (for instance, in the figure 3,
motions of the T -foil and flaps, WV A acceleration).

The simulation tool is used to run experiments. By

clicking on the SIMULINK block labelled
"Experiments", a dialog window opens. Figure 4
shows this window. The combo box "Experiment" let
us to select the kind of experiment, without
appendages, with appendages or controlled. The user
specifies the total time of the experiment,
"Simulation time". There are also, in the "Speed"
line, three possible ship's speeds: 20, 30 and 40
knots. Finally, there is a set of regular and irregular
waves (the same studied in the CEHIPAR
experiments) that can be used for the experiments.
The environment compares the simulated results with
the experimental results for the experiment selected.
In this new version the direct feedback from "Ship"
block to "Actuators" block (Esteban et al. 2000a)
has been eliminated using new models of the ship
with appendages. This is very important for the
simulation because this behaviour is very complex
and the feedback in the old version means an
algebraic loop, difficult to handle for the integration.
With algebraic loops it is necessary to use a very fine
integration step: this implies a slower simulation.

The "Control Process" block include different sub

blocks, figure 6. There is a sampling block to force
the user to employ sampled signals like in the
experiments. A filter has been added to eliminate
noise in the WV A signal. The "Controller" is the
block to be defined by the user, to study on the
simulation the performance of a candidate for the
control. Since the P.I.D. controller is the
conventional solution, our first studies focused on
this alternative. The objective is to get results for
reference purposes (to confirm whether we can find
better solutions).
 1. INTRODUCTION
The main problem for the development of high speed
craft is concerned with the passenger's comfort and
the safety of the vehicles. The vertical acceleration
associated with roll, pitch and heave is the main
cause of motion sickness. The roll control is the most
attractive candidate for control since increasing
damping can be obtained more easily. However,
shipbuilders are also interested in increasing pitch
and heave damping. In order to solve the problem
antipiching devices and pitch control methods must
be considered. Previously, models for the vertical
ship dynamics must be developed for the design,
evaluation and verification of the results.
Once the modelling stage of vertical dynamics of a
high speed ferry (De la Cruz et aI., 1998; Aranda et
aI., 2000) and actuators (Esteban et aI., 2000) is
finished then the following stage is to design a pitch
273
and heave control to command the positions of the
actuators. The final goal is decreasing the vertical
accelerations that produce motion sickness.
In this paper a comparative study of different
multivariable classic controllers: PD, first order filter
and second order filter is made. Using these designs
a decreasing of the motion sickness incidence is
obtained.
2. USED MODELS
2.1 Linear models of vertical dynamic of a high
speed ferry.
Vertical dynamic of a high speed ferry can be
described (De la Cruz et aI., 1998; Aranda et aI.,
2000) by four transfer functions:
 1. INTRODUCTION
This work is part of the investigation about
alleviation of fast ferry vertical accelerations
using appendages. Before it, the following
stages of the project have been completed: the
experiments with the scale model without
actuators had been done in CEHIP AR (Canal de
Experiencias Hidrodinamicas de El Pardo,
Madrid; in English: El Pardo Model Basin)
[CEHIP AR, 2000], vertical dynamic models of
the ship had been obtained from them,
SIMULINK simulation environment had been
tested and our work group had started to design
the control algorithms.
The next logical step was to apply the control
on the scaled down replica with appendages.
279
We decided to employ an industrial PC to test
the control algorithms with the scaled down
replica. It was important to check the different
control strategies, and to be able to do changes
quickly in the CEHIP AR installations. This was
motivated because the trials had to be
performed in a short time without interrupts.
At present, our group is working in the
development of a software tool for control code
automatic generation. This kind of tools are
suitable to solve the design of control software
for reactive systems like the control experiments
with the replica. The tool runs on MS-Windows
and generates C++ code that is supported by a
real time operating system called RTKemel.
The results, during our experiments, have
satisfied the expectations. We have proved the
efficiency of the tool to redesign quickly the
control software "in situ". Moreover we always
could work with a stable control program that
met the specifications.
This tool can be used in other experiments
related with ship control. Because of it we
consider interesting its presentation. This paper
begins with a introduction to ROOM
methodology, after that we explain the tool
developed and its utilization to solve the control
experiments with the replica. Finally we present
the conclusions of this work
2. ROOM METHODOLOGY
ROOM (Real-Time Object Oriented Modelling)
was introduced by Selic in 1994 [Selic, et aI.,
1994 ]. It is a formalism for modelling real time
systems using the object oriented paradigm.
This paradigm was introduced in the beginnings
of SIMULA and SMALTALK languages and it
is characterized to confer behaviour to the
objects of computation. In a reactive system, the
object of the computation will be actors whose
behaviours cooperate to solve the system.
3. THE ENVIRONMENT FOR FAST
CREATION OF REAL TIME CONTROL
SYSTEMS
As a result of the investigation performed by
our group, a CASE tool called EdROOM has
been development. This tool is capable to
generate automatically real time control code
starting from a ROOM model of the system. It
includes a graphic editor to define the actor
structure of the model and the behaviour of its
actors. The figure 1 shows part of the structure
of a model, made with EdROOM, that includes
5 actors and their interconnections. The figure 2
shows the behaviour of one of this actors edited
with EdROOM. The tool also lets define the
communication protocols, and the petitions to
the communication, timing and scheduling
servIces.

ROOM lets us describe the structure and

behaviour of the real time system using
.diagrams. The main components are actors that
communicate between them by message
passing. The behaviour of each actor is defined
using a kind of state chart, called ROOMCharts,
based on the Statecharts introduced by Harel
[Harel, 1987]. The received messages lead the
trigger of the transitions between the states.

Working with ROOM, the designer of software

control system has to define actors, give them a
behaviour, and trace connections between them
to establish their communications. ROOM also
defines the set of scheduling, communication
and timing services to satisfy the requirements
of the real time software systems. The
communication include the definition of
communication protocols between actors, the
assignment of priorities to the messages and the
management of message queues.

Similarly to the functional blocks hierarchy

employed in SIMULlNK, in ROOM one actor
can be construct from other actors. This fact lets
define several levels in the structure of the
system. The ROOM formalism can be
implemented in different ways and its adoption
is very interesting to develop a clear and
structured real time control code.
EdROOM runs on Ms-Windows, but it
generates C++ code supported by RTKernel [3].
RTKernel is a multitasking real time kernel
which uses pre-emptitive scheduling. It uses
Ms-DOS to provide some services as file
management but works under it and manages its
own scheduler to assure the necessary

280
detenninism for working in a real time
conditions. Ms-DOS is not a real time system
and it does not offer multitasking therefore we
avoid it.
After we have constructed the graphic model of
the system and we have defined a few functions
integrated with it, we can generate C++ control
code using a specific utility of EdROOM. This
source code is compiled and linked with a
library called mv_rtk.lib which implements the
communication, timing and scheduling services
in RTKemel. The result is a compact executable
file which can run on a any PC with Ms-DOS.
The implementation of mv_rtk.lib in other real
time operating systems lets us construct portable
and highly reusable code using EdROOM.
4. REPLICA CONTROL HARDWARE
The Replica constructed by CEHIPAR is 4.5
meters long. It has incorporated a T-foil near the
bow and two transom flaps. There is a step
motor to move the T-foil wings and other to
move the Flaps. We have used an industrial PC
to control them. Figure 3 shows a picture of the
replica with its appendages. Figure 4 shows a
picture of one of the step motors.
maximum rotation speed of the motors is 67.5 0/
seg. This value corresponds with a 13.5 0/ seg
speed in the real ship.
The sensors located in the replica measure the
following variables: heave, pitch, the height of
the arriving wave, the drag forces ( starboard
and port) and the accelerations in several points
of the replica.
The replica is moved by a carriage which has a
complete installation of data acquisition devices
and video cameras. The control is perfonned by
the industrial PC fixed to the carriage and
located near the replica. The PC includes the
Advantech PCL812PG data acquisition card
connected to its bus. Six of its sixteen analog
inputs channels are used by the control program
to perfonn the periodic sampling and four of its
sixteen digital outputs manage the control of the
motors. Depends on the control algorithm it is
necessary to sample a different set of signals.
For example the first control studies have been
done using the WVA acceleration.
The industrial PC has also a console in which
we can observe the values of the main variables.
This has been very useful in the calibration
tasks and to complete the basic control tests
(moving by hand the replica and checking if the
value of the motor rotation signs truly
counteract the wave effects).
5. EDROOM APPLIED TO THE CONTROL
OF THE REPLICA
The model of the replica control includes 5
actors connected between them. The figure 5
shows this aspect of the model structure.
another actor "fileAnalogInputs". This enabled
us to do trials of the program in our laboratory,
using two motors similar to those mounted in
the replica, before to go to the CEHIP AR.

6. "IN SITU" REDESIGN EXPERIENCES

One of the advantages that we wanted to obtain

with the use of EdROOM was to make easy the
redesign of the control program. The experience
shows that the design of this kind of system
consists of an iterative process of test and error
that it is convenient to be clear and agile. The
experiments with the replica have been a good
occasion to check the efficiency and suitable
characteristics of this tool.

For example, after being ready the automatic 7. CONCLUSIONS

control of the replica based in the algorithm, it
was also considered opportune adding a manual
control to perform calibration tasks and to do
initial tests before the experiments. This
redesign modified some of the actor's behaviour
but it was ready in less than a day. The new
model continued to be clear and the program
generated needed little debugging.
Besides, during the experiments with waves, we
realized that the motors did a considerable noise
which affected to the sensors measurements.
Then, it was considered suitable to use a
software filter to decrease this noise. This
decision modified only the actor
"periodic_Sampling" dedicated to perform the
samples with a period of 2 ms and, after
filtering, send them to the "planner" each 10 ms
as happened before. Again, redesign could be
ready in less than a day and the improving
obtained, with respect to the control capabilities,
has been notable.
.
A vIsual CASE tool for real time automatic
control code generation has been created. The
tool, called EdROOM, runs under Ms-
Windows. It follows an object oriented
methodology of software engineering, called
ROOM, based in the use of actors. The use of
ROOM guarantees a structured and clear design
of control programs. The executable code is
generated quickly and can be used in Intel Pc.
This code uses a low cost real time kernel called
R TKernel which works in a stable way together
with Ms-DOS. The library mv rtk.lib
implements the communication, timing and
scheduling services using RTKernel primitives
necessary to execute the ROOM model. This
library is linked with the C++ code generated to
create the executable. The figure 10 shows the
relationships between C++ code generated,
mv_rtk.lib library, Ms-DOS and RTKernel.

That fact is also the key point of the reusability

of the ROOM actors and it let us say that actors
work like "plug and play" software components.
As the result of all these aspects the
development of real-time systems can be
affronted like the evolution of a basic prototype
that runs properly in each stage.

The code size of the control program is less than

400 kbytes ( including the graphic display of the
variables) and the PC used has been a Pentium
at 200 MHz. As a whole we have obtained a
low cost control system which meet the
specifications. The figure 10 shows the console
of the control program generated with
EdROOM. The graph on top represents the
values of the accelerations and the graph below
represents the positions of T-foil and flaps.
EdROOM has been employed to perfonn the RTKernel 4.0 and RTKernel 4.5 Real-Time
control of a fast ferry replica within the Multitasking kernel for C/C++. User's
experiments in the CEHIPAR. The control Manual.
moves aT-foil near the bow and two transom
flaps to alleviate, in the best possible way, the Selic, Bran, Gulleckson, Garth, and Ward, Paul
vertical accelerations of the ship. Thanks to T. (1994), Real-Time Object Oriented
EdROOM the experiments have been done Modelling. NewYork, John Wiley and
quickly and the debug time has been reduced Sons.
considerably. Besides, it has been able to test
most of the control program in laboratory
conditions, avoiding unnecessary experiments
in CEHIPAR.

One of the advantages of using EdROOM have

been its capability to make easy the redesign of
the control program. During the experiments
there have been several situations in that this
redesign have been necessary. It has been
proved EdROOM can solve efficiently this sort
of problems.

In comparison with other alternatives, like the

RealTime Workshop (MATLAB), our approach
is more in the spirit of software engineering
methods for a modular, clear development of
potentially big applications. Besides EdROOM
can be used for several RTOS platfonns, with
no changes in the design.

The generality of the EdROOM approach let us

tackle easily other experimental problems like
the test with other replicas and several control
strategies. For example, it is foreseen one set of
experiments in multivariable control based on
heave and pitch values.

It is also possible to develop other versions of

the mv_rtk.lib library, which supports the code
generated by EdROOM, to allow for the
generation of code for other hardware targets
with other operating systems. The only
requirement is the existence of a C++ compiler.

8. ACKNOWLEDGEMENTS

The authors want to thank the support of the

CICYT Spanish Committee (project TAP97-
0607-C03-01), and the collaboration of the
BAZAN and CEHIPAR staff.

9. REFERENCES

CEHIPAR (2000): www.cehipar.es

Harel, David. (July 1987), Statecharts: A Visual

Fonnalism for Complex Systems. Science
of Computer Programming 8 :231-274.

284
 Abstract: This paper is about a research on the use of active appendages to smooth the
motions of fast ferries. According to BAZAN specifications, the research focused on a
fast ferry with a T-foil and transom flaps. A collaboration of three research groups, with
the experimental support of CEHIP AR was established to accomplish the objectives.
The research was scheduled as two main steps. A first step of control-oriented
modelling has successfully been achieved. The second step is dedicated to control
design and experimental evaluation, seeking for the best solution. First experimental
results confirm good expectations with the use of the active controlled appendages. This
paper describes the main aspects of the research: the problem to be solved, the
methodology and fulfilment of the research project, and the most relevant results
obtained. Copyright © 2001 IFA C

Keywords: marine systems, ship control, transportation control, MIMO

1. INTRODUCTION investigation groups, from three universities, was

Fast ships are acquiring great importance. New
technologies are applied for more speed and better
performances. Also, new control systems and
appendages are added for several uses. One of the
problems with high speed is that considerable vertical
accelerations, due to waves, can appear. These
accelerations can originate sea-sickness (O'Hanlon,
MacCawley, 1974), degrading the comfort of
passengers.
Our research deals with the use of active appendages
to smooth the motions of fast ferries. There are
important reasons for smoothing the motions: to
improve the comfort of passengers, to avoid negative
effects on the ship, to increase the operational
capabilities of the ship. The origin of the research is
the proposal from the company BAZAN to start
working on this topic. A collaboration of three
established to embark on a research project. To have
a solid experimental basis, the services of CEHIP AR
(Canal de Experiencias Hidrodinamicas de El Pardo,
Madrid), a prestigious towing tank institution, have
been of particular importance.
The research project was defined and submitted to
CICYT (Comit, Interministerial de Ciencia y
Tecnologja), the most important government entity
for research promotion and support. The project
obtained the approval and financial support from
CICYT, and started in 1997.
It is interesting, to give an idea of the kind of
problems to be treated, to look at a graphical
representation of the impact of vertical accelerations
on passenger's sea-sickness. There is an index,
computed with a formula, denoted "MSI" (motion
sickness incidence) that can be used as a statistical

285

It was clearly noticed, from the beginning, that the
research involved several novel aspects. Fortunately,
the difficulties found along the research, related to
the experimental and the modelling tasks, were
successfully solved.
The paper describes the main traits of the research. It
begins dealing with the methodology and realization
of the project. Next, the main results are presented.
These results are satisfactory and promising.
2. STATEMENT OF THE RESEARCH
In order to have a specific ship, to centre the
research, Bazan selected a particular example
belonging to a series of fast ferries recently built by
this company. The research was restricted to head
seas. Bazan determined also the use of aT-foil near
the bow and transom flaps, to smooth vertical
motions. Figure 2 shows a photograph of the fast
ferry. Figure 3 shows a lateral view of the ship, with
a zoom on the actuators.
The T-foil and the flaps can move, to counteract the
effect of each incident wave. By means of a control
system, these appendages should move in the most
effective way: that means a problem of control
design. The objective of the research is to solve this
problem, following some important steps:
- To obtain a dynamic, control-oriented model of the
ship
- To define the criteria and specifications the control
designs should obey.
- To develop an evaluation system for the control
designs.
.
In View of the main objective and the necessary
aspects, a time scheduling and an operational
methodology was established for the complete
research team. The experience with control
investigations says that the first step, devoted to
modelling, usually means the longest effort, since it
~equires an experimental basis. This modelling step is
important because:
- When resorting to first principles analysis, it
promotes a study in depth of the system to be
controlled.
- On the basis of good models, reliable simulation
environments can be created. The study of control on
simulation environments is more flexible and fast,
and less dangerous, than the direct experimental
study.
- A simulation tool allows for a quick and easy
evaluation of different control alternatives.
Furt~e~ore, it favours the development of more
sophisticated and powerful control designs.
The main support for the physics analysis is given by
(Fossen, 1994; Lloyd, 1998; Lewis, 1994).
To obtain the experimental data for modelling,
CEHIPAR built a scaled-down (1/25) replica of the
fast ferry. With the replica, CEHIPAR, using its
facilities, performed a set of experiments at 20, 30
and 40 knots. In addition, CEHIPAR, using the
program PRECAL, obtained simulation data about
heaving and pitching motions, heaving force and
pitching moment, for a set of regular waves and
ship's speeds of 20, 30 and 40 knots. The program
PRECAL computes motions, forces and moments,
using a CAD description of the hull.
The data and experiences obtained were useful for
research feedback: to improve the accuracy of the
simulation environment, to refine the control.

At present, several different control design

alternatives are under development. They will be
evaluated in the simulation environment. If there are
better solutions than the P.LD., they will become
candidates for experimental testing.

4. RESULTS

Many times the investigators involved in control

problems pointed up that most of the time and efforts
during a research, are invested in modelling the
system or process to be controlled. This is also the
case with our research. Hence, the most interesting
and important results are reached in the last steps of
the project.
The most important and recent result of our research
is the confirmation that is possible, by means of
actuators and adequate control, to obtain a significant
attenuation of vertical accelerations. This is highly
positive for passenger's comfort and to increase the
operational capabilities of the ship. Figure 9 shows,
as an example, the experimental results with irregular
waves SSN 5 and a speed of 40 knots. Near the bow,
passengers experiment the worse vertical acceleration
("WV A"). The figure shows measures of this
acceleration without appendages, and with controlled
moving appendages. There is a reduction of about
75%.
The simulation environment enables us to do on
computer the same experiments as in CEHIP AR with
the replica. During an experiment, the simulation
shows animated pictures and data displays about the
motions of the ship and actuators, and about the
pertinent variables. When an experiment is finished,
the simulation tool perform a data processing to
obtain qualification indexes for the evaluation of the
control (for instance, the MSI).
Other important result is a set of models of the ship
and the actuators. They are control-oriented models
in MA TLAB-SIMULINK. Among other uses, they
allow for the estimation of the ideal contribution of
the actuators to decreasing the vertical acceleration.
Figure 11 shows the expected ideal reduction of
pitching acceleration for irregular waves. Since there
are three ship's speeds (20, 30 and 40 knots) and
three sea states (SSN 4, 5 and 6), nine cases are
studied.
With regard to methodologies, the main contributions
belong to two important aspects. First, a modelling
methodology has been achieved, based on the
hybridising of genetic algorithms and non linear least
squares. Second, a fast method for real-time testing
of control algorithms on the experimental system has
been implemented. This has been accomplished using
EdROOM: a software engineering tool developed by
one of the research groups.

From another perspective, the research has obtained

as important result the creation of an infrastructure of
people, means, methods and experiences, that can
embark on new topics of marine systems control.

5. CONCLUSIONS

The paper describes the definition and realization of a

research project. The target is to attenuate as much as
possible the vertical accelerations of fast ships, by
using controlled actuators.

To accomplish the project, a research team has been

organized with participation of Bazan, CEHIPAR
and three university research groups. Funding has
been obtained from CICYT. A first stage of the
research, along three years, has been concluded.

The research focus on a specific fast ferry. CEHIPAR

built a replica, and numerous experiments have been
performed for the several needs of the research.

It has been confirmed that the controlled actuators are

effective to attenuate the vertical accelerations. In
addition, a simulation environment has been
developed for the control design and evaluation. New
models of the ship and the actuators have been
established, and used as the basis of the simulation. It
is a set of results really satisfactory.

In achieving the research, an infrastructure and a

methodology has been established, enabling for new
research topics on marine systems control. The
collaboration of different entities, from industry and
academia, has been a fruitful and pleasant
expenence.

In the next future, the context of the study will be

enlarged, considering other ship's motions.

This paper has the objective of giving a general view

of our research. Other papers of our team, in this
conference, will present more details on each of the
main aspects of the research.
In order to overcome the above mentioned difficulties,
a multi-sensor fusion technique is proposed as shown
in Fig. 1. In this technique, information other than the
optical information is used in decision making for the
navigation of the underwater vehicle. Position of
different points on the cable is measured and this
information is used as a rough position model of the
cable. The rough position data model of the cable is
used to predict the most likely region ofthe cable in the
image, which reduces the amount of image processing
data decreasing the processing time. Due to the
narrowing of the region of interest in the image, the
chances of misinterpretation of similar features
appearing in the image can be avoided. In the proposed
technique, dynamics of the AUV are also used to predict
the position of the cable in the image (Balasuriya et al.,
1997, 1998). This compensates the delays introduced
by the image processing algorithm (Fujii et aI., 1993).

A rough model of the layout of the cable is generated

by taking the position (xj'Y) of a few points along the
cable as shown in Fig. 2. The line connecting xjYj to
\+IYj+1 acts as the model.of c~ble an~ is used to ~redict
the most probable regIOn m the image for image
processing and is used for navigation command
generation when the vision processor cannot recognize
the cable in the environment. The use of the model for
cable pose prediction is explained in Fig. 3. The region
of interest corresponds to the region of uncertainty of
the rough model. Using the camera model, the
corresponding position ofthe model line is determined
and the interested region for image processing is
selected according to this position. The interested
region is selected according to the uncertainty of the
cable model. Uncertainty is considered as the maximum
possible deviation of the actual cable with respect to
the model line. Narrowing of the region of interest in
the image reduces the chances of misinterpretations of
other similar features in the image and also increases
the speed of processing due to the reduction in the
input data for image processing. If the cable is not
detected in this region then, the navigation is carried
out by following the model line. The model line features
(p,S) are used in the Hough plane too as shown in Fig. 4.
The cable image introduces a high concentration of
pixels in a particular direction forming a line in the
imageplane. This line feature introduces a peak in the
Hough space and its existing region can be predicted
using the uncertainty ofthe model line as shown in Fig.
4. This technique avoids the extraction of other peaks
in the Hough plane. In other words, it is possible to
distinguish the cable of interest even when there are
similar cables appearing in the image. If a peak
representing the cable cannot be found in the region
predicted, the features of the model line is used for the
navigation of the AUV.
The integration of different components for cable
tracking can be schematically represented as shown in
Fig. 5. The low-level controller provides the dynamic
state of the AUV and the optical image, captured by a
CCD camera. The optical image is pre-filtered for optical
noises using the SoLoG filter explained elsewhere
(Balasuriya et al., 1998). The position of the AUV
determined by deadreckoning is used to predict the
region of interest in the image

292
plane and that region is transfonned into the Hough
plane. Also using the position data of the AUV, the
interested region in the Hough plane is predicted and
the features of the cable image are extracted from the
Hough plane. The size of the region varies from frame
to frame and as a result, the image processing time will
not be a constant. Therefore, in order to compensate
the delays caused by image processing, a predictor
based on the dynamics ofthe AUV which is explained
in detail in (Balasuriya et at., 1998) is used as the delay
compensator. Once the instantaneous features of the
cable is detennined, these features are verified to check
whether they represent the cable. The properties of the
cable features are; 1) there are no abrupt changes, and
2) number of pixels on the line feature should be greater
than a defined threshold. If the features extracted by
image processing represents the properties of the cable
then those 2D features are transfonned into the vehicle
coordinate system for detennining the navigational
references. Else the model line features are used for
detennining the navigational references. The generation
of navigational references are discussed in (Balasuriya
etal., 1997,1998).
Two modes of operation are carried out to mltlal1ze the
position at the cross point of the mark shown in Fig. 7.
The initial dive of the AUV is done by keeping the
heading direction, determined prior to the mission, along
the Y axis. Once the defined height of the AUV with
respect to the bottom is reached the AUV switch into
the INITIAL SWAYING mode. Here, using visual
feedback the AUV is navigated in the direction parallel
to X ax~s only with sway actuator. At this instant heading
and height references are kept constant. This mode is
can:ied out until the image of the cable appears at the
honzontal centre of the image. In this task the camera is
kept in a forward-down looking position. The camera
axis is parallel to the centre axis of the AUV. Once the
cable is at the centre ofthe image the X-axis position of
the AUV is made zero.

The high-level controller explained in (Balasuriya et al.,

1997) is modified in this algorithm as shown in Fig. 6. It
is important to use the same coordinate system for both
2D model of the cable and AUV positioning. This task
is achieved by positioning the AUV initially with respect
to a defined target (initial mark) using visual data. The
initial mark is selected in such a way that it has two line
features falling in the x and y axis of the coordinate
system used for AUV positioning. The direction of the
cable at the initial location is selected as the y axis and
Next, by keeping the same sway position, heading and
height the AUV is moved forward by surging until the
mark appears at the vertical centre of the image. It is
called the INITIAL SURGING mode. Therefore, the mark-
cable intersection point is taken as the (0,0) point ofthe
coordinate system used for the 20 model of the cable,
and for the positioning of the AUV. If the initial
positioning is failed, the AUV ascends to the surface
indicating that it could not recognize the initial mark.
Else it starts tracking the cable and during tracking if
the vision system cannot recognize the cable, the mode
of operation is switched to the OEADRECKmode which
uses the 20 model line for navigation until the cable is
visible to the CCO camera mounted in the front of the
AUY.

The proposed algorithm therefore, can handle the

situations when the cable is invisible to the CCO camera
and when there are many similar cables appearing in the
image. The introduction of the 20 model of the cable
improves the performance of the autonomous
underwater cable tracking system.

angle of 55° with respect to the horizontal. The surge,

The experiment is carried out in a testing tank, whose
depth is 305m and the setting ofthe cable is as shown in
Fig. 9. The aim of the experiment is to test the
performance of the proposed system to handle the two
main practical points mentioned in section 2. The
problem in vision processing to recognize the cable to
be tracked when there are similar features appearing in
the image is tested by laying a similar cable close to the
cable of interest. The problem in vision based cable
tracking to track a cable when it is invisible in the image
is tested by discontinuing the cable. Also a similar cable,
whose layout is different, is introduced to the image
when the tracking cable is invisible in the image.
sway, yaw and heave motions are controlled for the
autonomous underwater cable tracking mission. Path
planning according to the visual features are carried
out as explained in (Balasuriya et aI., 1998). The cable is
laid for a distance of about 27m and is invisible for a
distance of about 8m. Similar hose pipe is laid close to
the interested hose in order to check the performance
of the tracking algorithm. The surge speed control
reference is kept at 0.1 m/s for vision based tracking
and 0.2m/s for deadreckoning. Sway speed control
reference is dependent on the change of position of the
cable in the image in vision based tracking and is
dependent on target position for deadreckoning. Yaw
rate control reference is dependent on the position
change of the cable in the image and is dependent on
the position model line in the case of deadreckoning.
Heave reference is generated to maintain a constant
height with reSDect to the cable.

In this experiment a yellow colour hose pipe is used as

the underwater cable to be tracked. The relative vertical
distance between the cable and the AUV is controlled
to be at 1m. An aluminium beam is used as the starting
point mark. The CCO camera, mounted in the front of
the AUV, is positioned with pan zero and with a tilt
 1 INTRODUCTION
An autonomous vehicle may be defined as "a
vehicle with a sensorial system and an actuator
system, managed by a control architecture able to
undertake a user specified mission". So, an
architecture is a framework in which the following
processes are implemented: sensing, control laws,
error detection and recovery, path planning, task
planning and monitoring of the events along the
execution of a particular mission.
There are three main approaches for organising the
architecture. Deliberative architectures, which are
strongly based on planning as well as a world
model, which allows reasoning about and making
predictions about the environment. Behavioural
architectures, also known as reactive architectures
or heterarchies, are broken down according to the
desired behaviours required of the robot. Normally,
the missions are described as a sequence of phases
with a set of active behaviours. The behaviours
react continuously to the situation sensed by the
perception system. The robot's global behaviour
emerges from the combination of the basic active
behaviours. The real world acts as a model to which
the robot reacts, based on the active behaviours. As
active behaviours are based on the sense-react
principle, they are very robust and very suitable for
dynamic environments.
311
Finally, hybrid architectures, which take advantage
of the two previous ones minimising their
limitations. Usually, these are structured in three
layers: (1) the deliberative, based on planning, (2)
the control execution layer (set onlset off
behaviours) and (3) a functional reactive layer.
Comparative studies of these different approaches
in the field of underwater robotics have been
published recently in the literature (Ridao, et ai.,
1999) and (VaIanavis, et ai., 1997). Other
interesting surveys were (Blidberg, et al. 1995),
(Coste-Maniere, et ai., 1995), (caccia, et aI., 1995),
(Hall, et. aI., 1992) and (Arkin 1998). The control
architecture of the GARB I follows the hybrid
approach. A description of GARBI (see Fig.ll-f)
can be found in (Amat, et aI., 1996). In this paper
we present the overall description of the 02CA2
architecture and a detailed description of its
Reactive-Layer.
This paper is organised as follows. Section 2
presents an overall description of the three layers.
The Reactive-Layer is explained in section 3.
Section 4 describes the model used for the
simulation and the experiments used to test the
concepts proposed by the architecture. Finally,
conclusions are presented in section 5.
compared with the actual sensor readings. If there is a discrepancy, the diagnostic engine is invoked to determine the
most likely cause of this discrepancy.

For diagnostics, the actual sensor values are propagated through the causal-network model to determine the faulty
component and its operational mode that would produce the abnormal sensor values. Conversely, for system
reconfiguration, the desired machine or process output values are propagated through the causal network to
determine the equipment settings necessary to achieve the desired output (subject to the constraint that certain
components are unavailable or are in failure mode). In this way, the same causal-network model can be used for
both diagnostics and for assisting control reconfiguration. When multiple diagnostic solutions or reconfiguration
solutions are found in the model, the best solution can be determined by evaluating probabilities or costs associated
with the various component modes or settings. Details on the model-based methodology for diagnosis and control
reconfiguration can be found in Provan and Chen (1999).

We can create causal-network models from a library of component models by interconnecting the components. For
example, the pump loop shown in Fig. 5 is composed of 6 component types: motor/pump, pipe, valve, tank, pressure
sensor and flowmeter. Rockwell Science Center has developed software tools for defming component models,
interconnecting the components to form causal-network models, and generating embeddable C code from the causal-
network models.