Page 315 - Practical Design Ships and Floating Structures

290

specification, a direct search on a design database may yield results that are suitable without further
modification.

0 2670000
2523000
H 237600 0
1 2229000
208200 0
193500 0
178800 0
1641000
149400 0
134700 0
120000 0
12.00
Transportation cost (€/tomes)
Figure 1 : Annual Cargo Vs Transportation Cost.
However, the direct database search approach may not yield any design that exactly meets the design
specification but may only identify some designs that are in reasonably close harmony with the given
design specification. Hence, it is necessary to “interpolate” between these designs to obtain a desirable
design that will meet the design specification. Such “interpolation” needs only involve the variables
identified above as involved in sub-problem one. In this example, the design data is viewed as being
active so that the designer need not know how the data is derived. Furthermore the data can be used
without involving relatively tedious and often iterative mathematical procedures. As in object oriented
programming approach, the “interpolation method” can be regarded as being attached with the data
and can be used in a transparent manner. In this case an objective-directed search employing a genetic
algorithm based multiobjective optimisation method was applied (Sen and Yang 1998) as the
“interpolation” method. The necessary ship design knowledge (e.g. stability requirement, powering
estimation, etc) is embedded within this method. The result of the interpolation is shown in Figure 1
which shows Pareto optimal solutions with respect to two economic objectives. For illustrative
purpose, the range of the search is wide so that a clear range of efficient solutions can be shown. From
Figure 1, a designer can then select efficient design solutions that most closely meet the specification.
As discussed before, reuse of design data can go beyond direct search and interpolation applications.
It is possible to extend the design data to satellite applications. For example, suppose the designer
would like to incorporate consideration of seakeeping characteristics (in terms of natural periods of roll,
heave and pitch) into the main design database without carrying out full-scale analysis. His current
database only has a relatively limited set of ships with known sea-keeping characteristics. If the
database is reasonably large and populated with reliable data then an Artificial Neural Net (ANN) can
be used to fit a response surface to the existing data. A three-layer feed-fomard ANN with seven
nodes in the input layer (nodes il - i7, length, breadth, depth, draught, block coefficient, waterplane
coefficient and metacentric height), seven (nodes 1 - 7) in the hidden layer and three in the output
layer (nodes 08 - 010, roll, heave and pitch period), was set up. This ANN was trained with a set of 110
training data from the designer’s current database. A separate set of 16 test data was used to test the
trained ANN. The approximate error given by the trained ANN for the test data was found to range
from 0.00% to 2.00% with a majority of results being within 1 .O%. The trained ANN is then able to
give approximate roll, heave and pitch periods of all vessels within the database given the required