Page 374 - Practical Design Ships and Floating Structures

Practical Design of Ships and Other Floating Structures 349
You-Sheng Wu, Wei-Cheng Cui and Guo-Jun Zhou (Eds)
8 2001 Elsevier Science Ltd. All rights reserved

BAYESIAN AND NEURAL NETWORKS
FOR PRELIMINARY SHIP DESIGN

H. B. Clausen, M.Liitzen, A. Friis-Hansen, N. Bjmneboe
Department of Mechanical Engineering,
Maritime Engineering Technical University of Denmark
DK-2800 Kongens Lyngby, Denmark

ABSTRACT
To ease the determination of the main particulars of a ship at the initial design stage it is convenient to
have tools which, given the type of ship and a few other parameters, output estimates of the remaining
dimensions. To establish such a tool, a database of the characteristics of about 87,000 ships is acquired
and various methods for derivation of empirical relations are employed. A regression analysis is
carried out to fit functions to the data. Further, the data is used to learn Bayesian and neural networks
to encode the relations between the characteristics. On the basis of examples, the three methods are
evaluated in terms of accuracy and limitations of use. For a chosen type of ship, here container vessels,
the methods provide information on the relations between length, breadth, height, draught, speed,
displacement, block coefficient and TEU capacity. Thus, useful tools are available for the designer
when he is to choose the preliminary main characteristics of a ship.

KEYWORDS
Neural Network, Bayesian Network, Main Particulars, Preliminary Ship Design, Data Fitting.

1 INTRODUCTION

The main particulars of a ship are determined at a preliminary stage, based on more or less detailed
customer requirements. The naval architect has to find a design with the ‘optimal’ main dimensions
and often the approach is to consider ships with similar characteristics and use an iteration loop to
modify the dimensions so that all specifications are met. hother approach is to use empirical relations
for finding the initial design parameters. Many authors have refined this approach (e.g. Bertram and
Wobig (1999), Watson and Gilfillan (1977)) by fitting regression lines to statistical data, yielding
explicit empirical expressions for the relations between various parameters. However, the empirical
expressions only give the relation between two parameters at a time irrespective of the rest. A method
to find the simultaneous relations between more parameters is addressed with neural and Bayesian
networks.